6.4. Adding a custom core
You may want to integrate a custom RISC-V core into the Chipyard framework. This documentation page provides step-by-step instructions on how to achieve this.
Note
RoCC is currently not supported by cores other than Rocket and BOOM. Please use Rocket or BOOM as the RoCC base core if you need to use RoCC.
Note
This page contains links to the files that contains important definitions in the Rocket chip repository, which is maintained separately from Chipyard. If you find any discrepancy between the code on this page and the code in the source file, please report it through GitHub issues!
6.4.1. Wrap Verilog Module with Blackbox (Optional)
Since Chipyard uses Scala and Chisel, if the top module of your core is not in Chisel, you will first need to create a Verilog blackbox for it so that it can be processed by Chipyard. See Incorporating Verilog Blocks for instructions.
6.4.2. Create Parameter Case Classes
Chipyard will generate a core for every InstantiableTileParams
object it discovered in the TilesLocated(InSubsystem)
key.
This object is derived from``TileParams``, a trait containing the information needed to create a tile. All cores must have
their own implementation of InstantiableTileParams
, as well as CoreParams
which is passed as a field in TileParams
.
TileParams
holds the parameters for the tile, which include parameters for all components in the tile (e.g.
core, cache, MMU, etc.), while CoreParams
contains parameters specific to the core on the tile.
They must be implemented as case classes with fields that can be overridden by
other config fragments as the constructor parameters. See the appendix at the bottom of the page for a list of
variable to be implemented. You can also add custom fields to them, but standard fields should always be preferred.
InstantiableTileParams[TileType]
holds the constructor of TileType
on top of the fields of TileParams
,
where TileType
is the tile class (see the next section).
All custom cores will also need to implement instantiate()
in their tile parameter class to return a new instance
of the tile class TileType
.
TileParams
(in the file BaseTile.scala) ,
InstantiableTileParams
(in the file BaseTile.scala),
CoreParams
(in the file Core.scala),
and FPUParams
(in the file FPU.scala)
contains the following fields:
trait TileParams {
val core: CoreParams // Core parameters (see below)
val icache: Option[ICacheParams] // Rocket specific: I1 cache option
val dcache: Option[DCacheParams] // Rocket specific: D1 cache option
val btb: Option[BTBParams] // Rocket specific: BTB / branch predictor option
val hartId: Int // Hart ID: Must be unique within a design config (This MUST be a case class parameter)
val beuAddr: Option[BigInt] // Rocket specific: Bus Error Unit for Rocket Core
val blockerCtrlAddr: Option[BigInt] // Rocket specific: Bus Blocker for Rocket Core
val name: Option[String] // Name of the core
}
abstract class InstantiableTileParams[TileType <: BaseTile] extends TileParams {
def instantiate(crossing: TileCrossingParamsLike, lookup: LookupByHartIdImpl)
(implicit p: Parameters): TileType
}
trait CoreParams {
val bootFreqHz: BigInt // Frequency
val useVM: Boolean // Support virtual memory
val useUser: Boolean // Support user mode
val useSupervisor: Boolean // Support supervisor mode
val useDebug: Boolean // Support RISC-V debug specs
val useAtomics: Boolean // Support A extension
val useAtomicsOnlyForIO: Boolean // Support A extension for memory-mapped IO (may be true even if useAtomics is false)
val useCompressed: Boolean // Support C extension
val useVector: Boolean = false // Support V extension
val useSCIE: Boolean // Support custom instructions (in custom-0 and custom-1)
val useRVE: Boolean // Use E base ISA
val mulDiv: Option[MulDivParams] // *Rocket specific: M extension related setting (Use Some(MulDivParams()) to indicate M extension supported)
val fpu: Option[FPUParams] // F and D extensions and related setting (see below)
val fetchWidth: Int // Max # of insts fetched every cycle
val decodeWidth: Int // Max # of insts decoded every cycle
val retireWidth: Int // Max # of insts retired every cycle
val instBits: Int // Instruction bits (if 32 bit and 64 bit are both supported, use 64)
val nLocalInterrupts: Int // # of local interrupts (see SiFive interrupt cookbook)
val nPMPs: Int // # of Physical Memory Protection units
val pmpGranularity: Int // Size of the smallest unit of region for PMP unit (must be power of 2)
val nBreakpoints: Int // # of hardware breakpoints supported (in RISC-V debug specs)
val useBPWatch: Boolean // Support hardware breakpoints
val nPerfCounters: Int // # of supported performance counters
val haveBasicCounters: Boolean // Support basic counters defined in the RISC-V counter extension
val haveFSDirty: Boolean // If true, the core will set FS field in mstatus CSR to dirty when appropriate
val misaWritable: Boolean // Support writable misa CSR (like variable instruction bits)
val haveCFlush: Boolean // Rocket specific: enables Rocket's custom instruction extension to flush the cache
val nL2TLBEntries: Int // # of L2 TLB entries
val mtvecInit: Option[BigInt] // mtvec CSR (of V extension) initial value
val mtvecWritable: Boolean // If mtvec CSR is writable
// Normally, you don't need to change these values (except lrscCycles)
def customCSRs(implicit p: Parameters): CustomCSRs = new CustomCSRs
def hasSupervisorMode: Boolean = useSupervisor || useVM
def instBytes: Int = instBits / 8
def fetchBytes: Int = fetchWidth * instBytes
// Rocket specific: Longest possible latency of Rocket core D1 cache. Simply set it to the default value 80 if you don't use it.
def lrscCycles: Int
def dcacheReqTagBits: Int = 6
def minFLen: Int = 32
def vLen: Int = 0
def sLen: Int = 0
def eLen(xLen: Int, fLen: Int): Int = xLen max fLen
def vMemDataBits: Int = 0
}
case class FPUParams(
minFLen: Int = 32, // Minimum floating point length (no need to change)
fLen: Int = 64, // Maximum floating point length, use 32 if only single precision is supported
divSqrt: Boolean = true, // Div/Sqrt operation supported
sfmaLatency: Int = 3, // Rocket specific: Fused multiply-add pipeline latency (single precision)
dfmaLatency: Int = 4 // Rocket specific: Fused multiply-add pipeline latency (double precision)
)
Most of the fields here (marked “Rocket spcific”) are originally designed for the Rocket core and thus contain some implementation-specific details, but many of them are general enough to be useful for other cores. You may ignore any fields marked “Rocket specific” and use their default values; however, if you need to store additional information with meaning or usage similar to these “Rocket specific” fields, it is recommended to use these fields instead of creating your own custom fields.
You will also need a CanAttachTile
class to add the tile config into the config system, with the following format:
case class MyTileAttachParams(
tileParams: MyTileParams,
crossingParams: RocketCrossingParams
) extends CanAttachTile {
type TileType = MyTile
val lookup = PriorityMuxHartIdFromSeq(Seq(tileParams))
}
During elaboration, Chipyard will look for subclasses of CanAttachTile
in the config system and instantiate a tile
from the parameters in this class for every such class it found.
Note
Implementations may choose to ignore some fields here or use them in a non-standard way, but using an inaccurate value may break Chipyard components that rely on them (e.g. an inaccurate indication of supported ISA extension will result in an incorrect test suite being generated) as well as any custom modules that use them. ALWAYS document any fields you ignore or with altered usage in your core implementation, and if you are implementing other devices that would look up these config values, also document them. “Rocket specific” values are generally safe to ignore, but you should document them if you use them.
6.4.3. Create Tile Class
In Chipyard, all Tiles are diplomatically instantiated. In the first phase, diplomatic nodes which specify Tile-to-System interconnects are evaluated, while in the second “Module Implementation” phase, hardware is elaborated. See TileLink and Diplomacy Reference for more details. In this step, you will need to implement a tile class for your core, which specifies the constraints on the core’s parameters and the connections with other diplomatic nodes. This class usually contains Diplomacy/TileLink code only, and Chisel RTL code should not go here.
All tile classes implement BaseTile
and will normally implement SinksExternalInterrupts
and SourcesExternalNotifications
,
which allow the tile to accept external interrupt. A typical tile has the following form:
class MyTile(
val myParams: MyTileParams,
crossing: ClockCrossingType,
lookup: LookupByHartIdImpl,
q: Parameters)
extends BaseTile(myParams, crossing, lookup, q)
with SinksExternalInterrupts
with SourcesExternalNotifications
{
// Private constructor ensures altered LazyModule.p is used implicitly
def this(params: MyTileParams, crossing: HierarchicalElementCrossingParamsLike, lookup: LookupByHartIdImpl)(implicit p: Parameters) =
this(params, crossing.crossingType, lookup, p)
// Require TileLink nodes
val intOutwardNode = None
val masterNode = visibilityNode
val slaveNode = TLIdentityNode()
// Implementation class (See below)
override lazy val module = new MyTileModuleImp(this)
// Required entry of CPU device in the device tree for interrupt purpose
val cpuDevice: SimpleDevice = new SimpleDevice("cpu", Seq("my-organization,my-cpu", "riscv")) {
override def parent = Some(ResourceAnchors.cpus)
override def describe(resources: ResourceBindings): Description = {
val Description(name, mapping) = super.describe(resources)
Description(name, mapping ++
cpuProperties ++
nextLevelCacheProperty ++
tileProperties)
}
}
ResourceBinding {
Resource(cpuDevice, "reg").bind(ResourceAddress(tileId))
}
// TODO: Create TileLink nodes and connections here.
6.4.4. Connect TileLink Buses
Chipyard uses TileLink as its onboard bus protocol. If your core doesn’t use TileLink, you will need to insert converters between the core’s memory protocol and TileLink within the Tile module. in the tile class. Below is an example of how to connect a core using AXI4 to the TileLink bus with converters provided by Rocket chip:
(tlMasterXbar.node // tlMasterXbar is the bus crossbar to be used when this core / tile is acting as a master; otherwise, use tlSlaveXBar
:= memoryTap
:= TLBuffer()
:= TLFIFOFixer(TLFIFOFixer.all) // fix FIFO ordering
:= TLWidthWidget(masterPortBeatBytes) // reduce size of TL
:= AXI4ToTL() // convert to TL
:= AXI4UserYanker(Some(2)) // remove user field on AXI interface. need but in reality user intf. not needed
:= AXI4Fragmenter() // deal with multi-beat xacts
:= memAXI4Node) // The custom node, see below
Remember, you may not need all of these intermediate widgets. See Diplomatic Widgets for the meaning of each intermediate
widget. If you are using TileLink, then you only need the tap node and the TileLink node used by your components. Chipyard also
provides converters for AHB, APB and AXIS, and most of the AXI4 widgets has equivalent widget for these bus protocol; see the
source files in generators/rocket-chip/src/main/scala/amba
for more info.
If you are using some other bus protocol, you may implement your own converters, using the files in generators/rocket-chip/src/main/scala/amba
as the template, but it is not recommended unless you are familiar with TileLink.
memAXI4Node
is an AXI4 master node and is defined as following in our example:
// # of bits used in TileLink ID for master node. 4 bits can support 16 master nodes, but you can have a longer ID if you need more.
val idBits = 4
val memAXI4Node = AXI4MasterNode(
Seq(AXI4MasterPortParameters(
masters = Seq(AXI4MasterParameters(
name = "myPortName",
id = IdRange(0, 1 << idBits))))))
val memoryTap = TLIdentityNode() // Every bus connection should have their own tap node
where portName
and idBits
(number of bits to represent a port ID) are the parameter provides by the tile.
Make sure to read TileLink Node Types to check out what type of nodes Chipyard supports and their parameters!
Also, by default, there are boundary buffers for both master and slave connections to the bus when they are leaving the tile, and you
can override the following two functions to control how to buffer the bus requests/responses:
(You can find the definition of these two functions in the class BaseTile
in the file
BaseTile.scala)
// By default, their value is "TLBuffer(BufferParams.none)".
protected def makeMasterBoundaryBuffers(implicit p: Parameters): TLBuffer
protected def makeSlaveBoundaryBuffers(implicit p: Parameters): TLBuffer
You can find more information on TLBuffer
in Diplomatic Widgets.
6.4.5. Create Implementation Class
The implementation class contains the parameterized, actual hardware that depends on the values resolved by the Diplomacy framework according to the info provided in the Tile class. This class will normally contains Chisel RTL code. If your core is in Verilog, you will need to instantiate the black box class that wraps your Verilog implementation and connect it with the buses and other components. No Diplomacy/TileLink code should be in this class; you should only connect the IO signals in TileLink interfaces or other diplomatically defined components, which are located in the tile class.
The implementation class for your core is of the following form:
class MyTileModuleImp(outer: MyTile) extends BaseTileModuleImp(outer){
// annotate the parameters
Annotated.params(this, outer.myParams)
// TODO: Create the top module of the core and connect it with the ports in "outer"
// If your core is in Verilog (assume your blackbox is called "MyCoreBlackbox"), instantiate it here like
// val core = Module(new MyCoreBlackbox(params...))
// (as described in the blackbox tutorial) and connect appropriate signals. See the blackbox tutorial
// (link on the top of the page) for more info.
// You can look at https://github.com/ucb-bar/cva6-wrapper/blob/master/src/main/scala/CVA6Tile.scala
// for a Verilog example.
// If your core is in Chisel, you can simply instantiate the top module here like other Chisel module
// and connect appropriate signal. You can even implement this class as your top module.
// See https://github.com/riscv-boom/riscv-boom/blob/master/src/main/scala/common/tile.scala and
// https://github.com/chipsalliance/rocket-chip/blob/master/src/main/scala/tile/RocketTile.scala for
// Chisel example.
If you create an AXI4 node (or equivalents), you will need to connect them to your core. You can connect a port like this:
outer.memAXI4Node.out foreach { case (out, edgeOut) =>
// Connect your module IO port to "out"
// The type of "out" here is AXI4Bundle, which is defined in generators/rocket-chip/src/main/scala/amba/axi4/Bundles.scala
// Please refer to this file for the definition of the ports.
// If you are using APB, check APBBundle in generators/rocket-chip/src/main/scala/amba/apb/Bundles.scala
// If you are using AHB, check AHBSlaveBundle or AHBMasterBundle in generators/rocket-chip/src/main/scala/amba/ahb/Bundles.scala
// (choose one depends on the type of AHB node you create)
// If you are using AXIS, check AXISBundle and AXISBundleBits in generators/rocket-chip/src/main/scala/amba/axis/Bundles.scala
}
6.4.6. Connect Interrupt
Chipyard allows a tile to either receive interrupts from other devices or initiate interrupts to notify other cores/devices.
In the tile that inherited SinksExternalInterrupts
, one can create a TileInterrupts
object (a Chisel bundle) and
call decodeCoreInterrupts()
with the object as the argument. Note that you should call this function in the implementation
class since it returns a Chisel bundle used by RTL code. You can then read the interrupt bits from the TileInterrupts
bundle
we create above. The definition of TileInterrupts
(in the file Interrupts.scala) is
class TileInterrupts(implicit p: Parameters) extends CoreBundle()(p) {
val debug = Bool() // debug interrupt
val mtip = Bool() // Machine level timer interrupt
val msip = Bool() // Machine level software interrupt
val meip = Bool() // Machine level external interrupt
val seip = usingSupervisor.option(Bool()) // Valid only if supervisor mode is supported
val lip = Vec(coreParams.nLocalInterrupts, Bool()) // Local interrupts
}
Here is an example on how to connect these signals in the implementation class:
// For example, our core support debug interrupt and machine-level interrupt, and suppose the following two signals
// are the interrupt inputs to the core. (DO NOT COPY this code - if your core treat each type of interrupt differently,
// you need to connect them to different interrupt ports of your core)
val debug_i = Wire(Bool())
val mtip_i = Wire(Bool())
// We create a bundle here and decode the interrupt.
val int_bundle = new TileInterrupts()
outer.decodeCoreInterrupts(int_bundle)
debug_i := int_bundle.debug
mtip_i := int_bundle.meip & int_bundle.msip & int_bundle.mtip
Also, the tile can also notify other cores or devices for some events by calling following functions in SourcesExternalNotifications
from the implementation class:
(These functions can be found in in the trait SourcesExternalNotifications
in the file
Interrupts.scala)
def reportHalt(could_halt: Option[Bool]) // Triggered when there is an unrecoverable hardware error (halt the machine)
def reportHalt(errors: Seq[CanHaveErrors]) // Varient for standard error bundle (Rocket specific: used only by cache when there's an ECC error)
def reportCease(could_cease: Option[Bool], quiescenceCycles: Int = 8) // Triggered when the core stop retiring instructions (like clock gating)
def reportWFI(could_wfi: Option[Bool]) // Triggered when a WFI instruction is executed
Here is an example on how to use these functions to raise interrupt.
// This is a demo. You should call these function according to your core
// Suppose that the following signal is from the decoder indicating a WFI instruction is received.
val wfi_o = Wire(Bool())
outer.reportWFI(Some(wfi_o))
// Suppose that the following signal indicate an unreconverable hardware error.
val halt_o = Wire(Bool())
outer.reportHalt(Some(halt_o))
// Suppose that our core never stall for a long time / stop retiring. Use None to indicate that this interrupt never fires.
outer.reportCease(None)
6.4.7. Create Config Fragments to Integrate the Core
To use your core in a Chipyard config, you will need a config fragment that will create a TileParams
object of your core in
the current config. An example of such config will be like this:
class WithNMyCores(n: Int = 1) extends Config((site, here, up) => {
case TilesLocated(InSubsystem) => {
// Calculate the next available hart ID (since hart ID cannot be duplicated)
val prev = up(TilesLocated(InSubsystem), site)
val idOffset = up(NumTiles)
// Create TileAttachParams for every core to be instantiated
(0 until n).map { i =>
MyTileAttachParams(
tileParams = MyTileParams(tileId = i + idOffset),
crossingParams = RocketCrossingParams()
)
} ++ prev
}
// Configurate # of bytes in one memory / IO transaction. For RV64, one load/store instruction can transfer 8 bytes at most.
case SystemBusKey => up(SystemBusKey, site).copy(beatBytes = 8)
case NumTiles => up(NumTiles) + n
})
Chipyard looks up the tile parameters in the field TilesLocated(InSubsystem)
, whose type is a list of InstantiableTileParams
.
This config fragment simply appends new tile parameters to the end of this list.
Now you have finished all the steps to prepare your cores for Chipyard! To generate the custom core, simply follow the instructions in Integrating Custom Chisel Projects into the Generator Build System to add your project to the build system, then create a config by following the steps in Heterogeneous SoCs. You can now run most desired workflows for the new config just as you would for the built-in cores (depending on the functionality your core supports).
If you would like to see an example of a complete third-party Verilog core integrated into Chipyard, generators/ariane/src/main/scala/CVA6Tile.scala
provides a concrete example of the CVA6 core. Note that this particular example includes additional nuances with respect to the interaction of the AXI
interface with the memory coherency system.