The hardware description of this CPU is done by using an very software oriented approach
(without any overhead in the generated hardware). There is a list of software concepts used :
- There is very few fixed things. Nearly everything is plugin based. The PC manager is a plugin, the register file is a plugin, the hazard controller is a plugin ...
- There is an automatic a tool which allow plugins to insert data in the pipeline at a given stage, and allow other plugins to read it in another stages through automatic pipelining.
- There is an service system which provide a very dynamic framework. As instance, a plugin could provide an exception service which could then be used by others plugins to emit exceptions from the pipeline.
The following number where obtains by synthesis the CPU as toplevel without any specific synthesis option to save area or to get better maximal frequency (neutral).<br>
The clock constraint is set to a unattainable value, which tends to increase the design area.<br>
All the cached configuration have some cache trashing during the dhrystone benchmark except the `VexRiscv full max perf` one. This of course reduce the performance. It is possible to produce dhrystone binaries which fit inside a 4KB I$ and 4KB D$ (I already had this case once) but currently it isn't the case.<br>
VexRiscv full max perf -> (RV32IM, 1.44 DMIPS/Mhz, 16KB-I$,16KB-D$, single cycle barrel shifter, debug module, catch exceptions, dynamic branch prediction in the fetch stage, branch and shift operations done in the Execute stage) ->
The VexRiscv could need the unreleased master-head of SpinalHDL. If it fail to compile, just get the SpinalHDL repository and do a "sbt clean compile publish-local" in it as described in the dependencies chapter.
- ISA tests from https://github.com/riscv/riscv-tests/tree/master/isa
- Dhrystone benchmark
- 24 tests FreeRTOS tests
- Some handwritten tests to check the CSR, debug module and MMU plugins
You can enable FreeRTOS tests by adding 'FREERTOS=yes' in the command line, will take time. Also, it use THREAD_COUNT host CPU threads to run multiple regression in parallel.
You can use the eclipse + Zylin embedded CDT plugin to do it (http://opensource.zylin.com/embeddedcdt.html). Tested with Helios Service Release 2 (http://www.eclipse.org/downloads/download.php?file=/technology/epp/downloads/release/helios/SR2/eclipse-cpp-helios-SR2-linux-gtk-x86_64.tar.gz) and the corresponding zylin plugin.
See https://drive.google.com/drive/folders/1NseNHH05B6lmIXqQFVwK8xRjWE4ydeG-?usp=sharing to import a makefile project and create a debug configuration.
Note that sometime this eclipse need to be restarted in order to be able to place new breakpoints.
### By using FreedomStudio
You can get FreedomStudio (which is package with eclipse and some plugins) there https://www.sifive.com/products/tools/
See https://drive.google.com/drive/folders/1a7FyMOYgFc9UDhfsWUSCjyqDCvOrts2J?usp=sharing to import a makefile project and create a debug configuration.
You can find some FPGA project which instantiate the Briey SoC there (DE1-SoC, DE0-Nano): https://drive.google.com/drive/folders/0B-CqLXDTaMbKZGdJZlZ5THAxRTQ?usp=sharing
Note that now a toplevel simulation testbench with the same feature + a GUI is implemented with SpinalSim. You can find it in src/test/scala/vexriscv/MuraxSim.scala.
To run it :
```sh
#This will generate the Murax RTL + run its testbench. You need Verilator 3.9xx installated.
You can find many example of different config in the https://github.com/SpinalHDL/VexRiscv/tree/master/src/main/scala/vexriscv/demo folder. There is one :
//Specify the IS_SIMD_ADD default value when instruction are decoded
decoderService.addDefault(IS_SIMD_ADD, False)
//Specify the instruction decoding which should be applied when the instruction match the 'key' parttern
decoderService.add(
//Bit pattern of the new SIMD_ADD instruction
key = M"0000011----------000-----0110011",
//Decoding specification when the 'key' pattern is recognized in the instruction
List(
IS_SIMD_ADD -> True,
REGFILE_WRITE_VALID -> True, //Enable the register file write
BYPASSABLE_EXECUTE_STAGE -> True, //Notify the hazard management unit that the instruction result is already accessible in the EXECUTE stage (Bypass ready)
BYPASSABLE_MEMORY_STAGE -> True, //Same as above but for the memory stage
RS1_USE -> True, //Notify the hazard management unit that this instruction use the RS1 value
As a demonstrator, this SimdAddPlugin was integrated in the src/main/scala/vexriscv/demo/GenCustomSimdAdd.scala CPU configuration and is self tested by the src/test/cpp/custom/simd_add application by running the following commands :
The first one (CustomCsrDemoPlugin) is adding an instruction counter and an clock cycle counter into the CSR mapping (and also do tricky stuff as a demonstration).<br>
While the second one (CustomCsrDemoGpioPlugin) is creating an GPIO peripheral directly mapped into the CSR.
Without the debug plugin, the CPU will have `clk` input and a `reset` input, which is very standard. But with the debug plugin the situation is the following :
- clk : As before, the clock which drive the whole CPU design, including the debug logic
- reset : Reset all the CPU states excepted the debug logics
- debugReset : Reset the debug logic of the CPU
- debug_resetOut : It is a CPU output signal which allow the JTAG to reset the CPU + the memory interconnect + the peripherals
So there is the reset interconnect in case you use the debug plugin :
VexRiscv is implemented via an 5 stages in order pipeline on which many optional and complementary plugins will add functionalities to provide a functional RISC-V CPU. This approach is completely unconventional and only possible on meta hardware description languages (SpinalHDL in the current case) but had proved its advantages via the VexRiscv implementation :
- You can swap/turn on/turn off parts of the CPU directly via the plugin system
- You can add new functionalities/instruction without having to modify any sources code of the CPU
- It allow the CPU configuration to cover a very large spectrum of implementation without cooking spagetti code
- To resume it allow your code base to truly produce a parametrized CPU design
So again, if you generate the CPU without any plugin, it will only contain the 5 stages definition and their basic arbitration, but nothing else, as everything else, including the program counter is added into the CPU via plugins.
This plugin implement the programme counter and over an jump service to all plugins.
| Parameters | type | description |
| ------ | ----------- | ------ |
| resetVector | BigInt | Address of the program counter after the reset |
| relaxedPcCalculation | Boolean | By default jump have an asynchronous immediate effect on the program counter, which allow to reduce the branch penalties by one cycle but could reduce the FMax as it will combinatorialy drive the instruction bus address signal. To avoid this you can set this parameter to true, which will make the jump affecting the programm counter in a sequancial way, which will cut the combinatorial path but add one additional cycle of penalty when a jump occur. |
The jump interface implemented by this plugin allow all other plugin to request jumps. The stage argument specify from which stage the jump is asked, which will allow the PcManagerSimplePlugin plugin to manage priorities between jump requests.
| cacheSize | Int | Total storage capacity of the cache |
| bytePerLine | Int | Number of byte per cache line |
| wayCount | Int | Number of cache way |
| twoCycleRam | Boolean | Check the tags values in the decode stage instead of the fetch stage to relax timings |
| asyncTagMemory | Boolean | Read the cache tags in a asyncronus manner instead of syncronous one |
| addressWidth | Int | Address width, should be 32 |
| cpuDataWidth | Int | Cpu data width, should be 32 |
| memDataWidth | Int | Memory data width, could potentialy be something else than 32, but only 32 is currently tested |
| catchIllegalAccess | Boolean | Catch when an memory access is done on non valid memory address (MMU) |
| catchAccessFault | Boolean | Catch when the memeory bus is responding with an error |
| catchMemoryTranslationMiss | Boolean | Catch when the MMU miss a TLB |
Note : If you enable the twoCycleRam and and the wayCount is bigger than one, then the register file plugin should be configured to read the regFile in a asyncronus manner.
This plugin will provide instruction decoding capabilities to others plugins. <br>
As instance, the pipeline hazard plugin will need to know, for a given instruction, if it is using the register file source 1/2 in order stall the pipeline until the hazard is gone. So to provide this kind of information, each plugin which implement an instruction will document to the DecoderSimplePlugin plugin this kind of informations.
| Parameters | type | description |
| ------ | ----------- | ------ |
| catchIllegalInstruction | Boolean | If set to true, instruction which have no decoding specification will generate an trap exception |
There is an usage example :
```scala
//Specify the instruction decoding which should be applied when the instruction match the 'key' pattern
decoderService.add(
//Bit pattern of the new instruction
key = M"0000011----------000-----0110011",
//Decoding specification when the 'key' pattern is recognized in the instruction
List(
IS_SIMD_ADD -> True,
REGFILE_WRITE_VALID -> True, //Enable the register file write
BYPASSABLE_EXECUTE_STAGE -> True, //Notify the hazard management unit that the instruction result is already accessible in the EXECUTE stage (Bypass ready)
BYPASSABLE_MEMORY_STAGE -> True, //Same as above but for the memory stage
RS1_USE -> True, //Notify the hazard management unit that this instruction use the RS1 value
RS2_USE -> True //Same than above but for RS2.
)
)
}
```
This plugin operate in the Decode stage
#### RegFilePlugin
This plugin implement the register file.
| Parameters | type | description |
| ------ | ----------- | ------ |
| regFileReadyKind | RegFileReadKind | Can bet set to ASYNC or SYNC. Specify the kind of memory read used to implement the register file. ASYNC mean zero cycle latency memory read, while SYNC mean one cycle latency memory read which can be mapped into standard FPGA memory blocks |
| zeroBoot | Boolean | Load all registers with zeroes at the beginning of simulations to keep everything deterministic in logs/traces|
This register file use an `don't care` read during write policy, so the bypassing/hazard plugin should take care of this.
#### HazardSimplePlugin
This plugin check the pipeline instruction dependencies and depending them, it will stop the instruction in the decoding stage or bypass the instruction results from the following stages to the decode stage.
As the register file is implemented with a `don't care` read during write policy, this plugin also have to manage hazard comming from this.
| Parameters | type | description |
| ------ | ----------- | ------ |
| bypassExecute | Boolean | Enable the bypassing of instruction results comming from the Execute stage |
| bypassMemory | Boolean | Enable the bypassing of instruction results comming from the Memory stage |
| bypassWriteBack | Boolean | Enable the bypassing of instruction results comming from the WriteBack stage |
| bypassWriteBackBuffer | Boolean | Enable the bypassing of the previous cycle register file written value |
#### SrcPlugin
This plugin muxes different inputs values to produce SRC1/SRC2/SRC_ADD/SRC_SUB/SRC_LESS values which are common values used by many plugins in the exectue stage (ALU / Branch / Load / Store).
| Parameters | type | description |
| ------ | ----------- | ------ |
| separatedAddSub | RegFileReadKind | By default SRC_ADD/SRC_SUB are generated from a single controllable adder/substractor, but if this is set to true, it use separated adder/substractors |
| executeInsertion | Boolean | By default SRC1/SRC2 are generated in the Decode stage, but if this parameter is true, it is done in the Execute stage (It will relax the bypassing network) |
Excepted SRC1/SRC2, this plugin do everything at the begining of Execute stage.
#### IntAluPlugin
This plugin implement all ADD/SUB/SLT/SLTU/XOR/OR/AND/LUI/AUIPC instructions in the execute stage by using the SrcPlugin outputs. It is a realy simple plugin.
The result is injected into the pipeline directly at the end of the execute stage.
#### LightShifterPlugin
Implement SLL/SRL/SRA instructions by using an iterative shifter register, whill use one cycle per bit shift.
The result is injected into the pipeline directly at the end of the execute stage.
| earlyInjection | Boolean | By default the result of the shift is injected into the pipeline in the Memory stage to relax timings, but if this option is true it will be done in the Execute stage |
#### BranchPlugin
This plugin implement all branch/jump instructions (JAL/JALR/BEQ/BNE/BLT/BGE/BLTU/BGEU) with some optional branch prediction. Each of those branch prediction could have been implemented into separated plugins.
| Parameters | type | description |
| ------ | ----------- | ------ |
| earlyBranch | Boolean | By default the branch is done in the Memory stage to relax timings, but if this option is set it's done in the Execute stage|
| catchAddressMisaligned | Boolean | If a jump/branch is done in an unaligned PC address, it will fire an trap exception |
| prediction | BranchPrediction | Can be set to NONE/STATIC/DYNAMIC/DYNAMIC_TARGET to specify the branch predictor implementation, see bellow for more descriptions |
| historyRamSizeLog2 | Int | Specify the number of entries in the direct mapped prediction cache of DYNAMIC/DYNAMIC_TARGET implementation. 2 pow historyRamSizeLog2 entries |
Each miss predicted jumps will produce between 2 and 4 cycles penalty depending the `earlyBranch` and the `PcManagerSimplePlugin.relaxedPcCalculation` configurations
##### Prediction NONE
No prediction, each PC changes due to a jump/branch will produce a penalty.
##### Prediction STATIC
In the decode stage, if the instruction is an conditional branch pointing backward or an JAL, it branch it speculatively. If the speculation is right it the branch penality is reduced to a single cycle, else the standard penalty is applied.
##### Prediction DYNAMIC
It is the same than the STATIC prediction, excepted that to do the prediction, it use a direct mapped 2 bit history cache (BHT) which remember if the branch is more likely to be taken or not.
##### Prediction DYNAMIC_TARGET
This predictor is using a direct mapped branch target buffer (BTB) in the Fetch stage which store the PC of the instruction, the target PC of the instruction and a 2 bit history to remember if the branch is more likely to be taken or not. This is the most efficient branch predictor actualy implemented on VexRiscv as when the branch prediction is right, is produce no branch penalty. The down side is that this predictor has a long combinatorial path comming from the prediction cache read port to the programm counter by passing through the jump interface.
#### DBusSimplePlugin
This plugin implement the load and store instructions (LB/LH/LW/LBU/LHU/LWU/SB/SH/SW) via a simple and neutral memory bus going out of the CPU.
| Parameters | type | description |
| ------ | ----------- | ------ |
| catchAddressMisaligned | Boolean | If a memory access is done in an unaligned memory address, it will fire an trap exception |
| catchAccessFault | Boolean | If a memory read return an error, it will fire an trap exception |
| earlyInjection | Boolean | By default, the memory read values are injected into the pipeline in the WriteBack stage to relax the timings, if this parameter is true it's done in the Memory stage |
There is the DBusSimpleBus
```scala
case class DBusSimpleCmd() extends Bundle{
val wr = Bool
val address = UInt(32 bits)
val data = Bits(32 bit)
val size = UInt(2 bit)
}
case class DBusSimpleRsp() extends Bundle with IMasterSlave{
val ready = Bool
val error = Bool
val data = Bits(32 bit)
override def asMaster(): Unit = {
out(ready,error,data)
}
}
case class DBusSimpleBus() extends Bundle with IMasterSlave{
val cmd = Stream(DBusSimpleCmd())
val rsp = DBusSimpleRsp()
override def asMaster(): Unit = {
master(cmd)
slave(rsp)
}
}
```
Note that bridges are implemented to convert this interface into AXI4 and Avalon
There is at least one cycle latency between que cmd and the rsp. the rsp.ready flag should be false after a read cmd until the rsp is present.
#### DBusCachedPlugin
Single way cache implementation with a victime buffer, documentation WIP
#### MulPlugin
Implement the multiplication instruction from the RISC-V M extension. Its implementation was done in a FPGA friendly way by using 4 multiplication of 17*17 bits. The processing is fully pipelined between the Execute/Memory/Writeback stage. The results of the instructions is always inserted in the WriteBack stage.
Implement the division/modulo instruction from the RISC-V M extension. It is done by a simple iterative manner which always take 34 cycles. The result is inserted into the Memory stage.
This plugin is now based on the MulDivIterativePlugin one.
#### MulDivIterativePlugin
This plugin implement the multiplication, division and modulo of the RISC-V M extension by an iterative manner, which is friendly for small FPGA which don't provide DSP blocks.
This plugin is able to unrool the iterative calculation process to reduce the number of cycles used to execute mul/div instructions.
| Parameters | type | description |
| ------ | ----------- | ------ |
| genMul | Boolean | Enable the multiplication support, can be set to false if you wan for instance to use the MulPlugin instead |
| genDiv | Boolean | Enable the division support |
| mulUnroolFactor | Int | Number of combinatorial stages used to speed up the multiplication, should be > 0 |
| divUnroolFactor | Int | Number of combinatorial stages used to speed up the division, should be > 0 |
The number of cycles used to execute a multiplication is '32/mulUnroolFactor'
The number of cycles used to execute a division is '32/divUnroolFactor + 1'
Both mul/div are processed into the memory stage (late result)
Implement most of the Machine mode and a very little bit of the User mode specified in the RISC-V previlegied spec. The access mode of most of the CSR is parameterizable (NONE/READ_ONLY/WRITE_ONLY/READ_WRITE) to reduce the area usage of useless features.
(CsrAccess can be NONE/READ_ONLY/WRITE_ONLY/READ_WRITE)
| Parameters | type | description |
| ------ | ----------- | ------ |
| catchIllegalAccess | Boolean | |
| mvendorid | BigInt | |
| marchid | BigInt | |
| mimpid | BigInt | |
| mhartid | BigInt | |
| misaExtensionsInit | Int | |
| misaAccess | CsrAccess | |
| mtvecAccess | CsrAccess | |
| mtvecInit | BigInt | |
| mepcAccess | CsrAccess | |
| mscratchGen | Boolean | |
| mcauseAccess | CsrAccess | |
| mbadaddrAccess | CsrAccess | |
| mcycleAccess | CsrAccess | |
| minstretAccess | CsrAccess | |
| ucycleAccess | CsrAccess | |
| wfiGen | Boolean | |
| ecallGen | Boolean | |
If an interrupt occur, before jumping to mtvec, the plugin will stop the Prefetch stage and wait that all the instructions in the following stages end their execution.
If an exception occur, the plugin will kill the corresponding instruction, flush all previous instruction, and wait until the previously killed instruction reach the WriteBack stage before jumping to mtvec.
#### StaticMemoryTranslatorPlugin
Static memory translator plugin which allow to specify which range of the memory addresses is IO mapped and shouldn't be cached
#### MemoryTranslatorPlugin
Simple software refilled MMU implementation. Allow others plugins as DBusCachedPlugin/IBusCachedPlugin to instanciate memory address translation ports. Each port has a small dedicated fully associative TLB cache which is refilled from a larger software filled TLB cache via an query which will look up one entry per cycle.
#### DebugPlugin
This plugin implement enough CPU debug feature to allow a comfortable GDB/eclipse debugging. To access those debug feature it provide a simple memory bus interface, the JTAG interface is provided by another bridge, which allow to efficiently connect multiple CPU to the same JTAG.
| Parameters | type | description |
| ------ | ----------- | ------ |
| debugClockDomain | ClockDomain | As the debug unit is able to reset the CPU itself, it should use another clock domain to avoid killing itself (only the reset wire should differ) |
The internals of the debug plugin are done in a manner which reduce the area usage and the FMax impact of this plugin.
There is the simple bus to access it, the rsp come one cycle after the request :
```scala
case class DebugExtensionCmd() extends Bundle{
val wr = Bool
val address = UInt(8 bit)
val data = Bits(32 bit)
}
case class DebugExtensionRsp() extends Bundle{
val data = Bits(32 bit)
}
case class DebugExtensionBus() extends Bundle with IMasterSlave{
val cmd = Stream(DebugExtensionCmd())
val rsp = DebugExtensionRsp()
override def asMaster(): Unit = {
master(cmd)
in(rsp)
}
}
```
There is the register mapping :
```
Read address 0x00 ->
bit 0 : resetIt
bit 1 : haltIt
bit 2 : isPipBusy
bit 3 : haltedByBreak
bit 4 : stepIt
Write address 0x00 ->
bit 4 : stepIt
bit 16 : set resetIt
bit 17 : set haltIt
bit 24 : clear resetIt
bit 25 : clear haltIt and haltedByBreak
Read Address 0x04 ->
bits (31 downto 0) : Last value written into the register file
Write Address 0x04 ->
bits (31 downto 0) : Instruction that should be pushed into the CPU pipeline for debug purposes
```
The OpenOCD port is there :
https://github.com/SpinalHDL/openocd_riscv
#### YamlPlugin
This plugin offer a service to others plugin to generate an usefull Yaml file about the CPU configuration, it will contain, for instance, the sequence of instruction required to flush the data cache (information used by openocd)
