Ecosyste.ms: Awesome

An open API service indexing awesome lists of open source software.

Awesome Lists | Featured Topics | Projects

https://github.com/bluespec/Flute

RISC-V CPU, simple 5-stage in-order pipeline, for low-end applications needing MMUs and some performance
https://github.com/bluespec/Flute

Last synced: 13 days ago
JSON representation

RISC-V CPU, simple 5-stage in-order pipeline, for low-end applications needing MMUs and some performance

Awesome Lists containing this project

README 

        
= Flute, a free and open-source RISC-V CPU

Rishiyur S. Nikhil, Bluespec, Inc. (c) 2021

:revnumber: v1.0

:revdate: 2021-05-19

:sectnums:

:toc:

:toclevels: 3

:toc-title: Contents

:description: Highly-parameterized RISC-V CPU, from small RV32 embedded to Linux-capable RV64

:keywords: RISC-V, Bluespec, BSV, Flute



'''



// ================================================================

// SECTION



== CPU microarchitecture and options



Flute is one of a family of free and open-source RISC-V CPUs from

Bluespec, Inc.  (https://bluespec.com[]) For a description of the

family, see Section <>.



Flute's microarchitecture is a 5-stage, in-order pipeline.  The source

code is highly parameterized to produce hundreds of variants, ranging

from a small, bare-metal RV32 CPU to a medium-size, Linux-capable RV64

CPUs with multiple privilege levels.



There is a sketch of the module hierarchy in `Doc/Microarchitecture/Microarchitecture.pdf`.



=== RISC-V ISA Spec options



* RV32I or RV64I

* Optional ISA A (Atomic instructions)

* Optional ISA C (Compressed instructions)

* Optional ISA M (Integer multiply and divide)

* Optional ISA F (Single-precision IEEE floating point) +

  Optional ISA D (Double-precision IEEE floating point)

    ** For F and D, optional hardare divider (or trap on divide instructions)

* Optional Privileged Architecture MU (Machine and User levels) or MSU

  (Machine, Supervisor and User Levels)

    ** For RV32, Privilege S implements the Sv32 Virtual Memory scheme

    ** For RV64, Privilege S implements the Sv39 Virtual Memory scheme



=== CPU microarchitecture options



* Barrel-shifter (faster, more expensive hardware) or serial shifter

  (cheaper, slower hardware) for Shift instructions



* Multipliers inferred by RTL Syntheis tool (faster, more expensive

    hardware), or serial multiplier (cheaper, slower hardware)

    multiplication for ISA M option



=== "`Near-Memory options`" (Caches and MMUs)



Flute has separate I- and D-caches.



* WT_L1: L1 only for I-Cache and D-Cache; write-through policy

* WB_L1: L1 only for I-Cache and D-Cache; write-back policy

* WB_L1_L2: L1 for I-Cache and D-Cache, shared coherent L2 cache; write-back policy



For all of them, an MMU is included if Privilege S is selected.



=== Development options



* Optional standard RISC-V Debug Module that can connect to GDB, to

  control and observe the CPU.



* Optional Tandem Verification trace generator that outputs an

  instruction-by-instruction trace that can be compared with a golden

  model RISC-V CPU (e.g., RISC-V Formal Specication, simulator like

  Spike or Bluespec Cissr)



// ================================================================

// SECTION



== The provided system to run out-of-the-box, including CLINT and PLIC



This repository contains a simple system (a small SoC) that

instantiates the CPU. The system can be compiled into a Bluesim or

Verilog simulation, and can be synthesized for FPGA.



The immediate layer (the "`core`") that surrounds the CPU includes a

CLINT (we call it "`Near Mem IO`") which contains the standard RISC-V

MTIME (Real-time timer) and MTIMECMP (Timer Compare) memory-mapped

registers and an MSIP "`Software Interrupt`" memory-mapped register.



The immediate layer (the "`core`") that surrounds the CPU also

includes a standard memory-mapped PLIC (Platform-Level Interrupt

Controller).



The SoC surrounding the core is based on an AMBA AXI4 interconnect, to

which is connected a boot ROM model, a DRAM memory model and a UART

model for serial communication.  The interconnect is parameterized for

the number of M (Manager) and S (Subordinate) ports, so one can attach

more memories, peripherals, or accelerators (needs recompilation).



The system is run in simulation by loading the DRAM memory model with

a data from a standard Verilog "`Mem Hex`" file (hexadecimal memory

contents), which can be produced from a RISC-V ELF binary file.



// ================================================================

// SECTION



== Directory structure



All the hardware designs in this repository are written in BSV, an

HLHDL (High-Level Hardware Description Language).  BSV sources files

have extension `.bsv`.  These are compiled to standard synthesizable

Verilog RTL using the free and open source _bsc_ compiler

(https://github.com/B-Lang-org/bsc[]).



// ----------------------------------------------------------------

// SUBSECTION



=== CPU and Core, in `src_Core/`



The `src_Core/` directory contains the sources for the CPU and immediate surrounding core(s).



The CPU itself is in sub-directories `CPU`, `ISA`, and `RegFiles`.



The CLINT (unit with memory-mapped MTIME, MTIMECMP and MSIP registers) are in `Near_Mem_IO`.



The PLIC (Platform-level Interrupt Controller) is in `PLIC`.



The alternative "`near-memory`" subsystems (caches, MMUs) are in

`Near_Mem_VM_WT_L1`,

`Near_Mem_VM_WB_L1`, and

`Near_Mem_VM_WB_L1_L2`.



The optional RISC-V Debug Module is in `Debug_Module`.



Two alterative "`cores`" are in `Core` (used with WT_L1 and WB_L1) and

`Core_v2` (used with WB_L1_L2). These directories also contain the

optional Tandem Verification generators.  These cores instantiate the

CPU, chosen near-memory subsystem, CLINT, PLIC, optional Debug Module

and optional Tandem Verification generator.



// ----------------------------------------------------------------

// SUBSECTION



=== SoC and simulation top-level, in `src_Testbench/`



The SoC is in `src_Testbench/SoC`.  This includes a boot ROM model, an

AXI4 interconnect fabric, a memory controller for DRAM, and a UART

model for serial communications.



The file `SoC/SoC_Map.bsv` specifies the system's address map

(addresses for memory, boot ROM, CLINT, PLIC, UART, etc.)



The subdirectory `src_Testbench/Fabrics/` contains the code for AXI4 interfaces,

transactors and fabrics.



Everything in the SoC and below is synthesizable, and can be

synthesized for FPGA.



The `src_Testbench/Top/` subdirectory is the only part that is meant

for simulaton only (not synthesizable). It contains a thin layer

around the SoC to provide a simulation clock and reset, a memory model

for the DRAM, and connections from the UART to the terminal.



// ----------------------------------------------------------------

// SUBSECTION



=== ISA tests (in `Tests/`)



The directory `Tests/isa` is a copy of the "`official`" RISC-V ISA

tests (original is at https://github.com/riscv/riscv-tests[]).  It

also contains compiled versions of all the tests (each has a RISC-V

ELF file and an "`objdump`" file that shows its disassemby).



The directory `Tests/elf_to_hex` contains a small C program to convert

an ELF file to a "`Mem Hex`" memory hexadecimal contents file.



// ----------------------------------------------------------------

// SUBSECTION



=== Provided example builds (including generated Verilogs)



As mentioned earlier, one can generate hundreds of variants of Flute

depending on the choice of configuration parameters.  This repository

contains "`build`" directories for a few particular configurations,

both to provide an out-of-the-box experience and to serve as example

templates which you can modify to create your own variant:



The following are for RV32I + C (Compressed instructions), bare-metal

(M and U privilege levels). One builds for the Bluesim simulator, the

other for Icarus Verilog (iverilog):



[source]

----

builds/Flute_RV32CI_MU_WT_L1_bluesim_tohost/

builds/Flute_RV32CI_MU_WT_L1_iverilog_tohost/

----



The following are for RV64GC (RV64IMAFDC), privilege levels M, S and

U, virtual memory Sv39. One builds for the Bluesim simulator, the

other for a Verilator simulator.  These have booted FreeRTOS, Linux

and FreeBSD, in simulation and on FPGA.



[source]

----

builds/Flute_RV64GC_MSU_WB_L1_L2_bluesim_tohost/

builds/Flute_RV64GC_MSU_WB_L1_L2_verilator_tohost/

----



// ================================================================

// SECTION



== Building and running simulations (Bluesim or Verilog simulation)



For some of the actions below, you need to have installed the free and

open-source _bsc_ compiler, which you can find at

https://github.com/B-Lang-org/bsc[].



// ----------------------------------------------------------------

// SUBSECTION



=== Building for Bluesim simulation (_bsc_ needed)



In one of the Bluesim build directories,

e.g,. `builds/Flute_RV32CI_MU_WT_L1_bluesim_tohost/` the command



[source]

----

$ make compile simulator

----



will compile and build a Bluesim simulator.



See section below for how to run the simulation.



// ----------------------------------------------------------------

// SUBSECTION



=== Building for Verilator simulation (_bsc_ not needed)



Each of the Verilator build directories,

e.g,. `builds/Flute_RV64GC_MSU_WB_L1_L2_verilator_tohost/`

contains a `Verilog_RTL` directory where we have already generated the

Verilog RTL sources for you from the BSV sources.



_You do not need the free and open-source bsc compiler to just build

the Verilog simulator from the Verilog sources._



The following command will build a Verilator simulation executable.



[source]

----

$ make simulator

----



See section below for how to run the simulation.



// ----------------

// SUBSUBSECTION



==== Regenerating the Verilog RTL (_bsc_ needed)



[source]

----

$ make compile

----



will regenerate the Verilog files in the `Verilog_RTL` directory.



// ----------------------------------------------------------------

// SUBSECTION



=== Running a simulation (Bluesim or Verilog simulation)



Once you have built a Bluesim, IVerilog or Verilator simulator as

described in the previous sections, you can run it as follows (_bsc_

is not needed for this).



To run a single ISA test:



[source]

----

$ make test

----



This runs the default ISA test (`Tests/isa/rv32ui-p-add` for RV32,

`Tests/isa/rv64ui-p-add` for RV64), and prints an instruction trace

during execution.  (First, it uses an `elf_to_hex` program, provided

in the `Tests/` directory, to convert the relevant ELF file into a

"`Mem Hex`" memory-contents file, which is loaded into the memory

model at the start of simulation).



You can choose a different ISA test from the `Tests/isa/` directory by

specifying it on the command line, like this:



[source]

----

$ make test TEST=rv64ui-v-ld

----



Note: if you specify a test that contains an instruction outside the

set of instructions for your build (e.g., an ELF that uses C

(compressed) instructions for a build that does not support C) this

will result in an illegal instruction trap, as expected.



You can run all relevant ISA tests (i.e., all those tests that are

relevant for the build's chosen ISA options) with:



[source]

----

$ make isa_tests

----



This will spawn multiple parallel processes to run run through all the

relevant tests.  The `Logs` subdirectory contains a log for each ISA

test that was run.



`make isa_tests` actually invokes the Python program

`Tests/Run_regression.py`, which you can run directly if you wish.

Running it with `--help` will describe its command-line arguments,

including the ISA architecture string, using which it selects the

"`relevant`" ISA tests.



The following will "`clean`" your build directory.  The first command

just deletes intermediate files and directories created during

creation of the simulator.  The latter will also deleted the simulator

itself and restore the directory to its pristine state.



[source]

----

$ make clean

$ make full_clean

----



// ----------------



==== Running your own ELF file on Flute



If you look at the actions taken by the Makefile in the above

examples, you can see how you can substitute your own ELF file as a

program to run.



// ================================================================

// SECTION



== Creating a new architecture configuration



In the `builds/` directory, you can create a new sub-directory to

build a new configuration of interest.  For example:



[source]

----

$ cd  builds

$ Resources/mkBuild_Dir.py  ..  RV32IMAC  MU  WT_L1  bluesim  tohost

----



will create a new directory: `Flute_RV32ACIM_MU_WT_L1_bluesim_tohost/`

populated with a `Makefile` to compile and link a bluesim simulation

for an RV32I CPU with M,A, and C ISA options, M and U privilege

levels, L1 I-Cache and D-Cache with write-through policy (no L2

cache), for building a Bluesim simulator, and which observes the

`tohost` memory location for test completion (which is the standard

method in ISA tests to signal completion).



You can build and run that simulator as usual:



[source]

----

$ make compile simulator test isa_tests

----



// ================================================================

// SECTION



'''

[#CPU_Family]

== Other open-source RISC-V CPUs from Bluespec, Inc.



This is one of a family of free, open-source RISC-V CPUs created by Bluespec, Inc.



* Piccolo (https://github.com/bluespec/Piccolo[]): 3-stage, in-order pipeline +

  For low-end applications (Embedded Systems, IoT, microcontrollers, etc.).



* Flute (https://github.com/bluespec/Flute[]): 5-stage, in-order pipeline. +

  For low-end to medium applications that require 32-bit or

  64-bit operation, an MMU (Virtual Memory) and more performance.



* Toooba (https://github.com/bluespec/Toooba[]): superscalar, deep, out-of-order RV64

  pipeline, using MIT's RISCY-OOO core.



All of them are written in entirely in BSV, an HLHDL (High-Level

Hardware Description Language).



The three repo structures are nearly identical, and the ways to build

and run are nearly identical.