diff --git a/npc/CMakeLists.txt b/npc/CMakeLists.txt
index db9d0a1..bd56708 100644
--- a/npc/CMakeLists.txt
+++ b/npc/CMakeLists.txt
@@ -1,91 +1,157 @@
-cmake_minimum_required(VERSION 3.20)
+cmake_minimum_required(VERSION 3.26)
 
 project(flow)
-set (CMAKE_CXX_STANDARD 11)
+set (CMAKE_CXX_STANDARD 14)
 cmake_policy(SET CMP0144 NEW)
 
-execute_process(
-  COMMAND ${CMAKE_SOURCE_DIR}/../git_commit.sh "configure(npc)"
-  WORKING_DIRECTORY ${CMAKE_SOURCE_DIR}/..
-)
+include(CMakeDependentOption)
+enable_testing()
 
-find_package(SDL2 REQUIRED)
-find_package(SDL2_image REQUIRED)
+# -- Build options
+option(BUILD_USE_BLOOP "Whether to use bloop to spped up elaborate" ON)
+option(BUILD_SIM_TARGET "Whether to build verilator simulation binary" ON)
+cmake_dependent_option(BUILD_SIM_NVBOARD_TARGET "Whether to build nvboard target" OFF "BUILD_SIM_TARGET" OFF)
+option(ENABLE_YSYX_GIT_TRACKER "Ysyx tracker support" OFF)
+set(TOPMODULE "Flow" CACHE STRING "Topmodule name in chisel")
 
-find_package(verilator REQUIRED)
+# -- Ysyx tracker, configure
+if(ENABLE_YSYX_GIT_TRACKER)
+  execute_process(
+    COMMAND ${CMAKE_SOURCE_DIR}/../git_commit.sh "configure(npc)"
+    WORKING_DIRECTORY ${CMAKE_SOURCE_DIR}/..
+  )
+endif()
+
+# -- Check dependencies
+if(BUILD_SIM_TARGET)
+  find_package(verilator REQUIRED)
+endif()
+if(BUILD_SIM_NVBOARD_TARGET)
+  find_package(SDL2 REQUIRED)
+  find_package(SDL2_image REQUIRED)
+endif()
 
 find_library(NVBOARD_LIBRARY NAMES nvboard)
 find_path(NVBOARD_INCLUDE_DIR NAMES nvboard.h)
 
-set(TOPMODULES "Flow")
+# FIXME: all scala source file are tracked here, cause all files to rebuild 
+# after a source update. 
+set(SCALA_CORE "${CMAKE_CURRENT_SOURCE_DIR}/core")
+set(CHISEL_MODULE_CLASS "${CMAKE_PROJECT_NAME}.${TOPMODULE}")
 
-foreach(TOPMODULE IN LISTS TOPMODULES)
- 
-  # FIXME: all scala source file are tracked here, cause all files to rebuild 
-  # after a source update. 
-  set(SCALA_CORE "${CMAKE_CURRENT_SOURCE_DIR}/core")
-  set(CHISEL_MODULE_CLASS "${CMAKE_PROJECT_NAME}.${TOPMODULE}")
-  file(GLOB_RECURSE SCALA_CORE_SOURCES "${SCALA_CORE}/src/main/scala/*.scala")
-  file(GLOB_RECURSE SCALA_CORE_TEST_SOURCES "${SCALA_CORE}/src/test/scala/*.scala")
+# Verilog files are generted in CHISEL_OUTPUT_TMP_DIR and copy to 
+# CHISEL_OUTPUT_DIR if content changes
+set(CHISEL_OUTPUT_DIR ${CMAKE_CURRENT_BINARY_DIR}/${TOPMODULE}/vsrc/)
+set(CHISEL_OUTPUT_TMP_DIR ${CMAKE_CURRENT_BINARY_DIR}/${TOPMODULE}/vsrc_tmp/)
 
-  # Configure time verilog source generation for verilator
-  execute_process(
-    COMMAND sbt "runMain circt.stage.ChiselMain --target-dir ${CMAKE_CURRENT_BINARY_DIR}/${TOPMODULE}/vsrc --module ${CHISEL_MODULE_CLASS} --target verilog"
-    WORKING_DIRECTORY ${SCALA_CORE}
-  )
+set(CHISEL_OUTPUT_VERILATOR_CONF ${CHISEL_OUTPUT_DIR}/conf.vlt)
+set(CHISEL_OUTPUT_TOPMODULE ${CHISEL_OUTPUT_DIR}/${TOPMODULE}.sv)
+set(CHISEL_EMIT_ARGS "--target-dir ${CHISEL_OUTPUT_TMP_DIR}")
 
+# -- Add an always run target to generate verilog files with sbt/bloop,
+#    as we don't know if the result files will be different from cmake
+# NOTE: Must reconfigure if we add new files in SCALA_CORE directory
+file(GLOB_RECURSE SCALA_CORE_SOURCES "${SCALA_CORE}/src/main/scala/*.scala")
+file(GLOB_RECURSE SCALA_CORE_RESOURCES "${SCALA_CORE}/src/resource/*")
+set(CHISEL_DEPENDENCY ${SCALA_CORE_SOURCES} ${SCALA_CORE_RESOURCE} ${SCALA_CORE}/build.sbt)
+
+if(BUILD_USE_BLOOP)
+  set(CHISEL_TARGET bloop_${TOPMODULE})
+  set(CHISEL_TEST_TARGET bloop_${TOPMODULE}_test)
+  # Export sbt build config to bloop
+  if(NOT EXISTS ${SCALA_CORE}/.bloop)
+    execute_process(
+      COMMAND sbt bloopInstall
+      WORKING_DIRECTORY ${SCALA_CORE}
+    )
+  endif()
+  string(REPLACE " " ";" CHISEL_EMIT_ARGS_LIST ${CHISEL_EMIT_ARGS})
+  list(TRANSFORM CHISEL_EMIT_ARGS_LIST PREPEND "--args;")
   add_custom_command(
-    OUTPUT ${CMAKE_CURRENT_BINARY_DIR}/${TOPMODULE}/vsrc/${TOPMODULE}.v
-    COMMAND sbt "runMain circt.stage.ChiselMain --target-dir ${CMAKE_CURRENT_BINARY_DIR}/${TOPMODULE}/vsrc --module ${CHISEL_MODULE_CLASS} --target verilog"
+    OUTPUT ${CHISEL_OUTPUT_TOPMODULE}
+    COMMAND bloop run root ${CHISEL_EMIT_ARGS_LIST}
+    COMMAND ${CMAKE_COMMAND} -E copy_directory_if_different ${CHISEL_OUTPUT_TMP_DIR} ${CHISEL_OUTPUT_DIR}
     WORKING_DIRECTORY ${SCALA_CORE}
-    DEPENDS ${SCALA_CORE_SOURCES}
+    DEPENDS ${CHISEL_DEPENDENCY}
+    COMMAND_EXPAND_LISTS
   )
-
-  add_custom_target(
-      ChiselBuild_${TOPMODULE}
-      DEPENDS ${CMAKE_CURRENT_BINARY_DIR}/${TOPMODULE}/vsrc/${TOPMODULE}.v
+  add_test(
+    NAME bloop_${TOPMODULE}_test
+    COMMAND bloop test
+    WORKING_DIRECTORY ${SCALA_CORE}
   )
+else()
+  set(CHISEL_TARGET sbt_${TOPMODULE})
+  set(CHISEL_TEST_TARGET sbt_${TOPMODULE}_test)
+  add_custom_command(
+    OUTPUT ${CHISEL_OUTPUT_TOPMODULE}
+    COMMAND sbt "run ${CHISEL_EMIT_ARGS}"
+    COMMAND ${CMAKE_COMMAND} -E copy_directory_if_different ${CHISEL_OUTPUT_TMP_DIR} ${CHISEL_OUTPUT_DIR}
+    WORKING_DIRECTORY ${SCALA_CORE}
+    DEPENDS ${CHISEL_DEPENDENCY}
+    VERBATIM
+  )
+  add_test(
+    NAME sbt_${TOPMODULE}_test
+    COMMAND sbt test
+    WORKING_DIRECTORY ${SCALA_CORE}
+  )
+endif()
 
-  # -- Build NVBoard executable
+if(NOT EXISTS ${CHISEL_OUTPUT_TOPMODULE})
+  # Probably cold build, generate verilog at configure time to produce top module file
+  execute_process(
+    COMMAND sbt "run ${CHISEL_EMIT_ARGS}"
+    WORKING_DIRECTORY ${SCALA_CORE}
+  )
+  execute_process(
+    COMMAND ${CMAKE_COMMAND} -E copy_directory_if_different ${CHISEL_OUTPUT_TMP_DIR} ${CHISEL_OUTPUT_DIR}
+  )
+endif()
 
+# -- Build NVBoard executable
+if(BUILD_SIM_NVBOARD_TARGET)
   add_custom_command(
     OUTPUT ${CMAKE_CURRENT_BINARY_DIR}/${TOPMODULE}/auto_bind.cpp
     COMMAND auto_pin_bind ${CMAKE_SOURCE_DIR}/constr/${TOPMODULE}.nxdc ${CMAKE_CURRENT_BINARY_DIR}/${TOPMODULE}/auto_bind.cpp
     DEPENDS ${CMAKE_SOURCE_DIR}/constr/${TOPMODULE}.nxdc 
   )
 
-  unset(SOURCES)
   file(GLOB_RECURSE SOURCES csrc_nvboard/${TOPMODULE}/*.cpp)
   add_executable(V${TOPMODULE}_nvboard ${SOURCES} ${CMAKE_CURRENT_BINARY_DIR}/${TOPMODULE}/auto_bind.cpp)
 
   verilate(V${TOPMODULE}_nvboard TRACE THREADS
     TOP_MODULE ${TOPMODULE}
     PREFIX V${TOPMODULE}
-    SOURCES ${CMAKE_CURRENT_BINARY_DIR}/${TOPMODULE}/vsrc/${TOPMODULE}.v)
+    SOURCES ${CHISEL_OUTPUT_TOPMODULE}
+    INCLUDE_DIRS ${CMAKE_CURRENT_BINARY_DIR}/${TOPMODULE}/vsrc)
 
   add_dependencies(V${TOPMODULE}_nvboard ChiselBuild_${TOPMODULE})
   target_include_directories(V${TOPMODULE}_nvboard PRIVATE ${NVBOARD_INCLUDE_DIR} ${SDL2_INCLUDE_DIRS})
   target_link_libraries(V${TOPMODULE}_nvboard PRIVATE ${NVBOARD_LIBRARY} SDL2::SDL2 SDL2_image::SDL2_image)
 
   install(TARGETS V${TOPMODULE}_nvboard)
+endif()
 
-  # -- Build Verilator executable and add to test
+# -- Build Verilator executable and add to test
+file(GLOB_RECURSE SOURCES csrc/${TOPMODULE}/*.cpp)
+add_executable(V${TOPMODULE} ${SOURCES})
 
-  unset(SOURCES)
-  file(GLOB_RECURSE SOURCES csrc/${TOPMODULE}/*.cpp)
-  add_executable(V${TOPMODULE} ${SOURCES})
+verilate(V${TOPMODULE} TRACE COVERAGE THREADS
+  TOP_MODULE ${TOPMODULE}
+  PREFIX V${TOPMODULE}
+  SOURCES ${CHISEL_OUTPUT_TOPMODULE} ${CHISEL_OUTPUT_VERILATOR_CONF}
+  INCLUDE_DIRS ${CMAKE_CURRENT_BINARY_DIR}/${TOPMODULE}/vsrc
+  VERILATOR_ARGS
+    "--vpi"   # Enable VPI
+    
+)
 
-  verilate(V${TOPMODULE} TRACE COVERAGE THREADS
-    TOP_MODULE ${TOPMODULE}
-    PREFIX V${TOPMODULE}
-    SOURCES ${CMAKE_CURRENT_BINARY_DIR}/${TOPMODULE}/vsrc/${TOPMODULE}.v)
 
-  add_dependencies(V${TOPMODULE} ChiselBuild_${TOPMODULE})
+add_test(NAME V${TOPMODULE} COMMAND V${TOPMODULE})
 
-  enable_testing()
-  add_test(NAME V${TOPMODULE} COMMAND V${TOPMODULE})
-
-  # -- Add build tracking
+# -- Add build tracking
+if(ENABLE_YSYX_GIT_TRACKER)
   add_custom_command(
       TARGET V${TOPMODULE}_nvboard PRE_BUILD
       COMMAND ${CMAKE_SOURCE_DIR}/../git_commit.sh "build_${CMAKE_PROJECT_NAME}_V${TOPMODULE}_nvboard"
@@ -97,5 +163,4 @@ foreach(TOPMODULE IN LISTS TOPMODULES)
       COMMAND ${CMAKE_SOURCE_DIR}/../git_commit.sh "build_${CMAKE_PROJECT_NAME}_V${TOPMODULE}"
       WORKING_DIRECTORY ${CMAKE_SOURCE_DIR}/..
   )
-
-endforeach()
+endif()
diff --git a/npc/core/build.sbt b/npc/core/build.sbt
index e04b603..3d0f4f8 100644
--- a/npc/core/build.sbt
+++ b/npc/core/build.sbt
@@ -3,6 +3,7 @@ ThisBuild / version          := "0.1.0"
 
 
 val chiselVersion = "6.2.0"
+val circeVersion  = "0.14.1"
 
 lazy val root = (project in file("."))
   .settings(
@@ -10,8 +11,13 @@ lazy val root = (project in file("."))
     libraryDependencies ++= Seq(
       "org.chipsalliance" %% "chisel" % chiselVersion,
       "edu.berkeley.cs" %% "chiseltest" % "6.0.0" % "test",
-      "com.chuusai" %% "shapeless" % "2.3.3"
-    ),
+      "com.chuusai" %% "shapeless" % "2.3.3",
+      "com.github.scopt" %% "scopt" % "4.1.0",
+    ) ++ Seq(
+      "io.circe" %% "circe-core",
+      "io.circe" %% "circe-generic",
+      "io.circe" %% "circe-parser"
+    ).map(_ % circeVersion),
     scalacOptions ++= Seq(
       "-language:reflectiveCalls",
       "-deprecation",
@@ -20,4 +26,4 @@ lazy val root = (project in file("."))
       "-Ymacro-annotations",
     ),
     addCompilerPlugin("org.chipsalliance" % "chisel-plugin" % chiselVersion cross CrossVersion.full),
-  )
+  )
\ No newline at end of file
diff --git a/npc/core/src/main/scala/Main.scala b/npc/core/src/main/scala/FlowMain.scala
similarity index 93%
rename from npc/core/src/main/scala/Main.scala
rename to npc/core/src/main/scala/FlowMain.scala
index 1ede7e1..ae39e24 100644
--- a/npc/core/src/main/scala/Main.scala
+++ b/npc/core/src/main/scala/FlowMain.scala
@@ -5,16 +5,17 @@ import chisel3._
 import chisel3.util.{MuxLookup, Fill, Decoupled, Counter, Queue, Reverse}
 import chisel3.util.{SRAM}
 import chisel3.util.experimental.decode.{decoder, TruthTable}
-import chisel3.stage.ChiselOption
 import chisel3.util.log2Ceil
 import chisel3.util.BitPat
 import chisel3.util.Enum
-import chisel3.experimental.prefix
+import chisel3.experimental.Trace._
 import shapeless.{HNil, ::}
 import shapeless.HList
 import shapeless.ops.coproduct.Prepend
 import chisel3.util.{ BinaryMemoryFile, HexMemoryFile }
 
+import chisel3.experimental.Trace
+
 object RV32Inst {
   private val bp = BitPat
   val addi = this.bp("b???????_?????_?????_000_?????_00100_11")
@@ -73,7 +74,7 @@ class Control(width: Int) extends Module {
     })
 }
 
-import flow.components.{RegisterFile, RegFileInterface, ProgramCounter, ALU}
+import flow.components.{RegisterFile, ProgramCounter, ALU}
 import chisel3.util.experimental.loadMemoryFromFileInline
 class Flow extends Module {
   val dataType = UInt(32.W)
@@ -87,7 +88,7 @@ class Flow extends Module {
     memoryFile = HexMemoryFile("./resource/addi.txt")
   )
   val control = Module(new Control(32))
-  val reg = RegisterFile(32, dataType, 2, 2)
+  val reg = Module(new RegisterFile(dataType, 32, 2))
   val pc = Module(new ProgramCounter(dataType))
   val alu = Module(new ALU(dataType))
 
@@ -95,6 +96,9 @@ class Flow extends Module {
   ram.readPorts(0).address := pc.out - 0x80000000L.U
   val inst = ram.readPorts(0).data
 
+  Trace.traceName(reg.control.writeEnable)
+  dontTouch(reg.control.writeEnable)
+
   import control.pc.SrcSelect._
 
   pc.in.pcSrcs(pStaticNpc.litValue.toInt) := pc.out + 4.U
@@ -125,5 +129,6 @@ class Flow extends Module {
   alu.in.a(aSrcImm.litValue.toInt) := inst(31, 20)
   alu.in.b := reg.out.src(1)
 
+
   dontTouch(control.out)
 }
diff --git a/npc/core/src/main/scala/Mem.scala b/npc/core/src/main/scala/Mem.scala
new file mode 100644
index 0000000..2a39fce
--- /dev/null
+++ b/npc/core/src/main/scala/Mem.scala
@@ -0,0 +1 @@
+package flow.components
\ No newline at end of file
diff --git a/npc/core/src/main/scala/RegisterFile.scala b/npc/core/src/main/scala/RegisterFile.scala
index 509ceaa..28ec02c 100644
--- a/npc/core/src/main/scala/RegisterFile.scala
+++ b/npc/core/src/main/scala/RegisterFile.scala
@@ -4,6 +4,7 @@ import chisel3._
 import chisel3.util.log2Ceil
 import chisel3.util.UIntToOH
 import chisel3.util.MuxLookup
+import chisel3.experimental.Trace._
 import shapeless.{ HNil, :: }
 
 class RegControl extends Bundle {
@@ -18,69 +19,34 @@ class RegControl extends Bundle {
   def ctrlBindPorts: CtrlTypes = {
     writeEnable :: writeSelect :: HNil
   }
+  traceName(writeEnable)
 }
 
-class RegFileData[T <: Data](size:Int, tpe: T, numReadPorts: Int, numWritePorts: Int) extends Bundle {
-  val write = new Bundle {
-    val addr = Input(UInt(size.W))
-    val data = Vec(numWritePorts, Input(tpe))
-  }
-  val read = Vec(numReadPorts, new Bundle {
-    val rs = Input(UInt(size.W))
-    val src = Output(tpe)
-  })
-}
-
-class RegFileInterface[T <: Data](size: Int, tpe: T, numReadPorts: Int, numWritePorts: Int) extends Bundle {
-  val control = new RegControl
-  // val data = new RegFileData(size, tpe, numReadPorts, numWritePorts)
-  val in = new Bundle {
-    val writeAddr = Input(UInt(size.W))
-    val writeData = Input(Vec(numWritePorts, tpe))
-    val rs = Input(Vec(numReadPorts, UInt(size.W)))
-  }
-  val out = new Bundle {
-    val src = Output(Vec(numReadPorts, tpe))
-  }
-}
-
-class RegisterFileCore[T <: Data](size: Int, tpe: T, numReadPorts: Int) extends Module {
+class RegisterFile[T <: Data](tpe: T, regCount: Int, numReadPorts: Int) extends Module {
   require(numReadPorts >= 0)
-  val writePort = IO(new Bundle {
-    val enable = Input(Bool())
-    val addr = Input(UInt(log2Ceil(size).W))
-    val data = Input(tpe)
+  val control = IO(new RegControl)
+  val dataAddrWidth = log2Ceil(tpe.getWidth).W
+  val in = IO(new Bundle {
+    val writeAddr = Input(UInt(dataAddrWidth))
+    val writeData = Input(Vec(control.WriteSelect.all.length, tpe))
+    val rs = Input(Vec(numReadPorts, UInt(dataAddrWidth)))
+  })
+  val out = IO(new Bundle {
+    val src = Output(Vec(numReadPorts, tpe))
   })
-  val readPorts = IO(Vec(numReadPorts, new Bundle {
-    val addr = Input(UInt(log2Ceil(size).W))
-    val data = Output(tpe)
-  }))
 
-  val regFile = RegInit(VecInit(Seq.fill(size)(0.U(tpe.getWidth.W))))
-  val writeAddrOH = UIntToOH(writePort.addr)
+  val regResetValue = 0.U(tpe.getWidth.W)
+  val regFile = RegInit(VecInit(Seq.fill(regCount)(regResetValue)))
+  val writeAddrOH = UIntToOH(in.writeAddr)
+
   for ((reg, i) <- regFile.zipWithIndex.tail) {
-    reg := Mux(writeAddrOH(i) && writePort.enable, writePort.data, reg)
+    reg := Mux(writeAddrOH(i.U(log2Ceil(regCount).W)) && control.writeEnable, in.writeData(control.writeSelect.asUInt), reg)
   }
   regFile(0) := 0.U
 
-  for (readPort <- readPorts) {
-    readPort.data := regFile(readPort.addr)
+  for (port <- 0 until numReadPorts) {
+    out.src(port) := regFile(in.rs(port).asUInt)
   }
+  traceName(regFile)
   dontTouch(regFile)
 }
-
-object RegisterFile {
-  def apply[T <: Data](size: Int, tpe: T, numReadPorts: Int, numWritePorts: Int): RegFileInterface[T] = {
-    val core = Module(new RegisterFileCore(size, tpe, numReadPorts))
-    val _out = Wire(new RegFileInterface(size, tpe, numReadPorts, numWritePorts))
-    val clock = core.clock
-    for (i <- 0 until numReadPorts) {
-      core.readPorts(i).addr := _out.in.rs(i)
-      _out.out.src(i) := core.readPorts(i).data
-    }
-    core.writePort.addr := _out.in.writeAddr
-    core.writePort.data := _out.in.writeData(_out.control.writeSelect.asUInt)
-    core.writePort.enable := _out.control.writeEnable
-    _out
-  }
-}
diff --git a/npc/core/src/main/scala/top/ArgParse.scala b/npc/core/src/main/scala/top/ArgParse.scala
new file mode 100644
index 0000000..b430a70
--- /dev/null
+++ b/npc/core/src/main/scala/top/ArgParse.scala
@@ -0,0 +1,47 @@
+package flow
+
+import scopt.{ OParser, DefaultOEffectSetup }
+import java.io.File
+
+case class CliOptions(
+    targetDir: File = new File("."),
+    configFile: Option[File] = None,
+    argsFile: Option[File] = None,
+    verilatorConfigFileOut: File = new File("conf.vlt"),
+) {
+  val builder = OParser.builder[CliOptions]
+  val parser =  {
+    import builder._
+    OParser.sequence(
+      programName("flow"),
+      help("help"),
+      opt[Option[File]]('c', "config")
+        .action((x, c) => c.copy(configFile = x))
+        .text("JSON Configuration file for generation"),
+      opt[Option[File]]("args-file")
+        .action((x, c) => c.copy(argsFile = x))
+        .text("Passing file content as args when emit"),
+      opt[File]('o', "target-dir")
+        .action((x, c) => c.copy(targetDir = x))
+        .text("Output files relative to this path"),
+      opt[File]("out-verilator-conf")
+        .action((x, c) => c.copy(verilatorConfigFileOut = x))
+        .text("Options needed when simulating with verilator")
+    )
+  }
+
+  def parse(args: Array[String]): CliOptions = {
+    OParser.runParser(parser, args, CliOptions()) match {
+      case (result, effects) =>
+        OParser.runEffects(effects, new DefaultOEffectSetup {
+          // ignore terminate
+          override def terminate(exitState: Either[String, Unit]): Unit = ()
+        })
+
+        result match {
+          case Some(cliOptions: CliOptions) => { return cliOptions }
+          case _ => { throw new IllegalArgumentException("Wrong command line argument") }
+        }
+    }
+  }
+}
diff --git a/npc/core/src/main/scala/top/Config.scala b/npc/core/src/main/scala/top/Config.scala
new file mode 100644
index 0000000..aa58107
--- /dev/null
+++ b/npc/core/src/main/scala/top/Config.scala
@@ -0,0 +1,16 @@
+package flow
+
+import io.circe.generic.JsonCodec
+
+// Which group of signals to trace
+@JsonCodec case class TraceConfig (
+  enable: Boolean = false,
+  registers: Array[Int] = Array(),
+  mem: Array[(Int, Int)] = Array(),
+)
+
+@JsonCodec case class Config(
+  // Whether to enable Difftest
+  enableDifftest: Boolean = true,
+  traceConfig: TraceConfig = TraceConfig(),
+)
diff --git a/npc/core/src/main/scala/top/Main.scala b/npc/core/src/main/scala/top/Main.scala
new file mode 100644
index 0000000..3c02ff3
--- /dev/null
+++ b/npc/core/src/main/scala/top/Main.scala
@@ -0,0 +1,72 @@
+package flow
+
+import flow._
+
+import chisel3._
+import chisel3.experimental.Trace._
+import chisel3.stage.{ChiselGeneratorAnnotation, DesignAnnotation}
+import chisel3.util.experimental.InlineInstance
+import circt.stage.ChiselStage
+import firrtl.AnnotationSeq
+import firrtl.annotations.TargetToken.{Instance, OfModule, Ref}
+import java.io.PrintWriter
+import scala.io.Source
+import java.io.File
+
+
+// TODO: Generate verilator config file
+
+object VerilogMain extends App {
+  val opt = CliOptions().parse(args)
+  val topName = "Flow"
+
+  val config: Config = opt.configFile match {
+    case Some(f) => {
+      val source = Source.fromFile(f)
+      val jsonString = source.mkString
+      source.close()
+      io.circe.parser.decode[Config](jsonString) match {
+        case Right(x) => x
+        case Left(e) => throw e
+      }
+    }
+    case None => Config(traceConfig = TraceConfig(enable = true))
+  }
+
+  val annos = (new ChiselStage).execute(
+    Array("--target-dir", opt.targetDir.toString, "--target", "systemverilog", "--split-verilog"),
+    Seq(
+
+    ) ++ (if(config.traceConfig.enable) Seq(ChiselGeneratorAnnotation(() => new Flow)) else Seq())
+  )
+
+  if(config.traceConfig.enable) {
+    val dut = annos.collectFirst { case DesignAnnotation(dut) => dut }.get.asInstanceOf[Flow]
+
+    val verilatorConfigSeq = finalTargetMap(annos)
+      .values
+      .flatten
+      .map(ct =>
+        s"""public_flat_rd -module "${
+          ct.tokens.collectFirst { case OfModule(m) => m }.get
+        }" -var "${ct.tokens.collectFirst { case Ref(r) => r }.get}"""")
+    finalTargetMap(annos)
+      .values
+      .flatten
+      .foreach(
+        ct => println(s"""TOP.${ct.circuit}.${ct.path.map { case (Instance(i), _) => i }.mkString(".")}.${ct.tokens.collectFirst {
+          case Ref(r) => r
+        }.get}""") 
+      )
+    
+    val verilatorConfigWriter = new PrintWriter(new File(opt.targetDir, opt.verilatorConfigFileOut.toString()))
+    verilatorConfigWriter.write("`verilator_config\n")
+    try {
+      for(ct <- verilatorConfigSeq) {
+        verilatorConfigWriter.println(ct)
+      }
+    } finally {
+      verilatorConfigWriter.close()
+    }
+  }
+}
\ No newline at end of file
diff --git a/npc/core/src/main/scala/utils/DPI.scala b/npc/core/src/main/scala/utils/DPI.scala
new file mode 100644
index 0000000..2c83be5
--- /dev/null
+++ b/npc/core/src/main/scala/utils/DPI.scala
@@ -0,0 +1,11 @@
+package flow.utils
+
+import chisel3._
+import chisel3.util.HasBlackBoxResource
+
+// class DiffTester extends BlackBox with HasBlackBoxResource {
+//     val io = IO(new Bundle {
+//         val regs = 
+//     })
+//     addResource("difftest.v");
+// }
diff --git a/npc/csrc/Flow/main.cpp b/npc/csrc/Flow/main.cpp
index c708610..465e1f0 100644
--- a/npc/csrc/Flow/main.cpp
+++ b/npc/csrc/Flow/main.cpp
@@ -1,43 +1,86 @@
+#include <vpi_user.h>
+#include <VFlow.h>
 #include <cstdlib>
-#include <cassert>
-#include <cstdlib>
+#include <vector>
+#include <memory>
 #include <verilated.h>
 #include <verilated_vcd_c.h>
-#include <VFlow.h>
+#include <verilated_vpi.h>
+#include <string>
 #define MAX_SIM_TIME 100
 #define VERILATOR_TRACE
 
-int main(int argc, char **argv, char **env) {
-    int sim_time = 0;
-    Verilated::commandArgs(argc, argv);
+std::vector<vpiHandle> regsHandle;
+int regs[32];
 
-    VFlow *top = new VFlow;
-
-    Verilated::traceEverOn(true);
-    VerilatedVcdC *m_trace = new VerilatedVcdC; 
-#ifdef VERILATOR_TRACE
-    top->trace(m_trace, 5);
-    m_trace->open("waveform.vcd");
-#endif
-    for (sim_time = 0; sim_time < 10; sim_time++) {
-        top->eval();
-        top->clock = !top->clock;
-        top->reset = 1;
-#ifdef VERILATOR_TRACE
-        m_trace->dump(sim_time);
-#endif
-    }
-    top->reset = 0;
-    for (sim_time = 10; sim_time < MAX_SIM_TIME; sim_time++) {
-        top->eval();
-        top->clock = !top->clock;
-#ifdef VERILATOR_TRACE
-        m_trace->dump(sim_time);
-#endif
-    }
-#ifdef VERILATOR_TRACE
-    m_trace->close();
-#endif
-    delete top;
-    exit(EXIT_SUCCESS);
+static void init_vpi_regs() {
+  std::string regfile = "TOP.Flow.reg_0.regFile_";
+  for(int i = 0; i < 32; i++) {
+    std::string regname = regfile + std::to_string(i);
+    vpiHandle vh = vpi_handle_by_name((PLI_BYTE8 *)regname.c_str(), NULL);
+    regsHandle.push_back(vh);
+  }
+}
+
+static void init_vpi() {
+  init_vpi_regs();
+}
+
+static int vpi_get_int(vpiHandle vh) {
+  s_vpi_value v;
+  v.format = vpiIntVal;
+  vpi_get_value(vh, &v);
+  return v.value.integer;
+}
+
+static void update_regs() {
+  for(int i = 0; i < 32; i++) {
+    regs[i] = vpi_get_int(regsHandle[i]);
+  }
+}
+
+static void print_regs() {
+  for(int i = 0; i < 32; i++) {
+    printf("%d: %d\t", i, regs[i]);
+    if(i % 8 == 7) putchar('\n');
+  }
+  putchar('\n');
+}
+
+static int sim_time = 0;
+
+int main(int argc, char **argv, char **env) {
+  int sim_time = 0;
+  int posedge_cnt = 0;
+  Verilated::commandArgs(argc, argv);
+
+  std::unique_ptr<VFlow> top{new VFlow};
+  Verilated::traceEverOn(true);
+  VerilatedVcdC *m_trace = new VerilatedVcdC;
+#ifdef VERILATOR_TRACE
+  top->trace(m_trace, 5);
+  m_trace->open("waveform.vcd");
+#endif
+
+  init_vpi();
+
+  top->reset = 0;
+  for (sim_time = 10; sim_time < MAX_SIM_TIME; sim_time++) {
+    top->eval();
+    top->clock = !top->clock;
+    if(top->clock == 1) {
+      // Posedge
+      ++posedge_cnt;
+      update_regs();
+      print_regs();
+    }
+
+#ifdef VERILATOR_TRACE
+    m_trace->dump(sim_time);
+#endif
+  }
+#ifdef VERILATOR_TRACE
+  m_trace->close();
+#endif
+  exit(EXIT_SUCCESS);
 }
diff --git a/npc/flake.nix b/npc/flake.nix
index 15b910f..00a98ba 100644
--- a/npc/flake.nix
+++ b/npc/flake.nix
@@ -20,9 +20,10 @@
           CHISEL_FIRTOOL_PATH = "${nixpkgs-circt162.legacyPackages.${system}.circt}/bin";
           packages = [
             clang-tools
-            # rnix-lsp
+            cmake
             coursier
             espresso
+            bloop
 
             gdb
             jre