fpu added exact div/sqrt implementations using iterative approaches

src/main/scala/vexriscv/ip/fpu/FpuCore.scala

@@ -3,6 +3,7 @@ package vexriscv.ip.fpu
 import spinal.core._
 import spinal.lib._
 import spinal.lib.eda.bench.{Bench, Rtl, XilinxStdTargets}
+import spinal.lib.math.UnsignedDivider
 
 import scala.collection.mutable.ArrayBuffer
 
@@ -24,8 +25,8 @@ case class FpuCore( portCount : Int, p : FpuParameter) extends Component{
   val exponentF32Infinity = exponentOne+127+1
   val exponentF64Infinity = exponentOne+1023+1
 
-  val rfLockCount = 5
+
-  val lockIdType = HardType(UInt(log2Up(rfLockCount) bits))
+  val lockIdType = HardType(UInt(log2Up(p.rfLockCount) bits))
 
   def whenDouble(format : FpuFormat.C)(yes : => Unit)(no : => Unit): Unit ={
     if(p.withDouble) when(format === FpuFormat.DOUBLE) { yes } otherwise{ no }
@@ -106,6 +107,25 @@ case class FpuCore( portCount : Int, p : FpuParameter) extends Component{
     val format = p.withDouble generate FpuFormat()
   }
 
+  case class DivInput() extends Bundle{
+    val source = Source()
+    val rs1, rs2 = p.internalFloating()
+    val rd = p.rfAddress()
+    val lockId = lockIdType()
+    val roundMode = FpuRoundMode()
+    val format = p.withDouble generate FpuFormat()
+  }
+
+
+  case class SqrtInput() extends Bundle{
+    val source = Source()
+    val rs1 = p.internalFloating()
+    val rd = p.rfAddress()
+    val lockId = lockIdType()
+    val roundMode = FpuRoundMode()
+    val format = p.withDouble generate FpuFormat()
+  }
+
 
   case class AddInput() extends Bundle{
     val source = Source()
@@ -145,11 +165,11 @@ case class FpuCore( portCount : Int, p : FpuParameter) extends Component{
       val boxed = p.withDouble generate Bool()
     }
     val ram = Mem(Entry(), 32*portCount)
-    val lock = for(i <- 0 until rfLockCount) yield new Area{
+    val lock = for(i <- 0 until p.rfLockCount) yield new Area{
       val valid = RegInit(False)
       val source = Reg(Source())
       val address = Reg(p.rfAddress)
-      val id = Reg(UInt(log2Up(rfLockCount) bits))
+      val id = Reg(UInt(log2Up(p.rfLockCount+1) bits))
       val commited = Reg(Bool)
       val write = Reg(Bool)
     }
@@ -184,7 +204,7 @@ case class FpuCore( portCount : Int, p : FpuParameter) extends Component{
 
   val commitLogic = for(source <- 0 until portCount) yield new Area{
     val fire = False
-    val target, hit = Reg(UInt(log2Up(rfLockCount+1) bits)) init(0)
+    val target, hit = Reg(UInt(log2Up(p.rfLockCount+1) bits)) init(0)
     val full = target + 1 === hit
     when(fire){
       hit := hit + 1
@@ -241,7 +261,7 @@ case class FpuCore( portCount : Int, p : FpuParameter) extends Component{
           commitLogic(i).target := commitLogic(i).target + 1
         }
       }
-      for(i <- 0 until rfLockCount){
+      for(i <- 0 until p.rfLockCount){
         when(rf.lockFreeId(i)){
           rf.lock(i).valid := True
           rf.lock(i).source := s0.source
@@ -317,10 +337,31 @@ case class FpuCore( portCount : Int, p : FpuParameter) extends Component{
       divSqrt.div := input.opcode === p.Opcode.DIV
     }
 
+    val divHit = input.opcode === p.Opcode.DIV
+    val div = Stream(DivInput())
+    if(p.withDiv) {
+      input.ready setWhen (divHit && div.ready)
+      div.valid := input.valid && divHit
+      div.payload.assignSomeByName(input.payload)
+    }
+
+    val sqrtHit = input.opcode === p.Opcode.SQRT
+    val sqrt = Stream(SqrtInput())
+    if(p.withSqrt) {
+      input.ready setWhen (sqrtHit && sqrt.ready)
+      sqrt.valid := input.valid && sqrtHit
+      sqrt.payload.assignSomeByName(input.payload)
+    }
+
+
     val fmaHit = input.opcode === p.Opcode.FMA
     val mulHit = input.opcode === p.Opcode.MUL || fmaHit
     val mul = Stream(new MulInput())
     val divSqrtToMul = Stream(new MulInput())
+    if(!p.withDivSqrt){
+      divSqrtToMul.valid := False
+      divSqrtToMul.payload.assignDontCare()
+    }
 
     if(p.withMul) {
       input.ready setWhen (mulHit && mul.ready && !divSqrtToMul.valid)
@@ -910,7 +951,7 @@ case class FpuCore( portCount : Int, p : FpuParameter) extends Component{
 //      val flag = io.port(input.source).completion.flag
       when(forceNan) {
         output.setNanQuiet
-        NV setWhen(input.valid && (infinitynan || input.rs1.isNanSignaling || input.rs2.isNanSignaling))
+        NV setWhen(infinitynan || input.rs1.isNanSignaling || input.rs2.isNanSignaling)
       } elsewhen(forceOverflow) {
         output.setInfinity
       } elsewhen(forceZero) {
@@ -958,6 +999,145 @@ case class FpuCore( portCount : Int, p : FpuParameter) extends Component{
     }
   }
 
+
+  val div = p.withDiv generate new Area{
+    val input = decode.div.halfPipe()
+    val haltIt = True
+    val output = input.haltWhen(haltIt).swapPayload(new MergeInput())
+
+    val dividerShift = if(p.withDouble) 0 else 1
+    val divider = FpuDiv(p.internalMantissaSize + dividerShift)
+    divider.io.input.a := input.rs1.mantissa << dividerShift
+    divider.io.input.b := input.rs2.mantissa << dividerShift
+    val dividerResult = divider.io.output.result >> dividerShift
+    val dividerScrap = divider.io.output.remain =/= 0 || divider.io.output.result(0, dividerShift bits) =/= 0
+
+    val cmdSent = RegInit(False) setWhen(divider.io.input.fire) clearWhen(!haltIt)
+    divider.io.input.valid := input.valid && !cmdSent
+    divider.io.output.ready := input.ready
+    output.payload.assignSomeByName(input.payload)
+
+    val needShift = !dividerResult.msb
+    val mantissa = needShift ? dividerResult(0, p.internalMantissaSize + 1 bits) |  dividerResult(1, p.internalMantissaSize + 1 bits)
+    val scrap = dividerScrap || !needShift && dividerResult(0)
+    val exponentOffset = 1 << (p.internalExponentSize + (if(p.withDouble) 0 else 1))
+    val exponent = input.rs1.exponent + U(exponentOffset | exponentOne) - input.rs2.exponent - U(needShift)
+
+    output.value.setNormal
+    output.value.sign := input.rs1.sign ^ input.rs2.sign
+    output.value.exponent := exponent.resized
+    output.value.mantissa := mantissa
+    output.scrap := scrap
+    if(!p.withDouble) when(exponent.takeHigh(2) === 3){ output.value.exponent(p.internalExponentSize-3, 3 bits) := 7} //Handle overflow
+
+
+
+    val underflowThreshold = muxDouble[UInt](input.format)(exponentOne + exponentOffset - 1023 - 53) (exponentOne + exponentOffset - 127 - 24)
+    val underflowExp = muxDouble[UInt](input.format)(exponentOne + exponentOffset - 1023 - 54) (exponentOne + exponentOffset - 127 - 25)
+    val forceUnderflow = exponent <  underflowThreshold
+    val forceOverflow = input.rs1.isInfinity || input.rs2.isZero
+    val infinitynan = input.rs1.isZero && input.rs2.isZero
+    val forceNan = input.rs1.isNan || input.rs2.isNan || infinitynan
+    val forceZero = input.rs1.isZero
+
+
+
+    output.NV := False
+    output.DZ := !forceNan && input.rs2.isZero
+
+    when(exponent(exponent.getWidth-3, 3 bits) === 7) { output.value.exponent(p.internalExponentSize-2, 2 bits) := 3 }
+
+    when(forceNan) {
+      output.value.setNanQuiet
+      output.NV setWhen((infinitynan || input.rs1.isNanSignaling || input.rs2.isNanSignaling))
+    } elsewhen(forceOverflow) {
+      output.value.setInfinity
+    } elsewhen(forceZero) {
+      output.value.setZero
+    } elsewhen(forceUnderflow) {
+      output.value.exponent := underflowExp.resized
+    }
+
+
+    haltIt clearWhen(divider.io.output.valid)
+  }
+
+
+
+  val sqrt = p.withSqrt generate new Area{
+    val input = decode.sqrt.halfPipe()
+    val haltIt = True
+    val output = input.haltWhen(haltIt).swapPayload(new MergeInput())
+
+    val needShift = !input.rs1.exponent.lsb
+    val sqrt = FpuSqrt(p.internalMantissaSize)
+    sqrt.io.input.a := (needShift ? (U"1" @@ input.rs1.mantissa @@ U"0") | (U"01" @@ input.rs1.mantissa))
+
+    val cmdSent = RegInit(False) setWhen(sqrt.io.input.fire) clearWhen(!haltIt)
+    sqrt.io.input.valid := input.valid && !cmdSent
+    sqrt.io.output.ready := input.ready
+    output.payload.assignSomeByName(input.payload)
+
+
+    val scrap = sqrt.io.output.remain =/= 0
+    val exponent =   RegNext(exponentOne-exponentOne/2 -1 +^ (input.rs1.exponent >> 1) + U(input.rs1.exponent.lsb))
+
+    output.value.setNormal
+    output.value.sign := input.rs1.sign
+    output.value.exponent := exponent
+    output.value.mantissa := sqrt.io.output.result
+    output.scrap := scrap
+    output.NV := False
+    output.DZ := False
+
+    val negative  = !input.rs1.isNan && !input.rs1.isZero && input.rs1.sign
+
+    when(input.rs1.isInfinity){
+      output.value.setInfinity
+    }
+    when(negative){
+      output.value.setNanQuiet
+      output.NV := True
+    }
+    when(input.rs1.isNan){
+      output.value.setNanQuiet
+      output.NV := !input.rs1.isQuiet
+    }
+    when(input.rs1.isZero){
+      output.value.setZero
+    }
+
+
+//    val underflowThreshold = muxDouble[UInt](input.format)(exponentOne + exponentOffset - 1023 - 53) (exponentOne + exponentOffset - 127 - 24)
+//    val underflowExp = muxDouble[UInt](input.format)(exponentOne + exponentOffset - 1023 - 54) (exponentOne + exponentOffset - 127 - 25)
+//    val forceUnderflow = exponent <  underflowThreshold
+//    val forceOverflow = input.rs1.isInfinity// || input.rs2.isInfinity
+//    val infinitynan = input.rs1.isZero && input.rs2.isZero
+//    val forceNan = input.rs1.isNan || input.rs2.isNan || infinitynan
+//    val forceZero = input.rs1.isZero
+//
+//
+//
+//    output.NV := False
+//    output.DZ := !forceNan && input.rs2.isZero
+//
+//    when(exponent(exponent.getWidth-3, 3 bits) === 7) { output.value.exponent(p.internalExponentSize-2, 2 bits) := 3 }
+//
+//    when(forceNan) {
+//      output.value.setNanQuiet
+//      output.NV setWhen((infinitynan || input.rs1.isNanSignaling || input.rs2.isNanSignaling))
+//    } elsewhen(forceOverflow) {
+//      output.value.setInfinity
+//    } elsewhen(forceZero) {
+//      output.value.setZero
+//    } elsewhen(forceUnderflow) {
+//      output.value.exponent := underflowExp.resized
+//    }
+
+
+    haltIt clearWhen(sqrt.io.output.valid)
+  }
+
   val divSqrt = p.withDivSqrt generate new Area {
     val input = decode.divSqrt.halfPipe()
 
@@ -1263,7 +1443,7 @@ case class FpuCore( portCount : Int, p : FpuParameter) extends Component{
 
 
 //      val flag = io.port(input.source).completion.flag
-      output.NV := (input.valid && (infinityNan || input.rs1.isNanSignaling || input.rs2.isNanSignaling))
+      output.NV := infinityNan || input.rs1.isNanSignaling || input.rs2.isNanSignaling
       output.DZ := False
       when(forceNan) {
         output.value.setNanQuiet
@@ -1286,6 +1466,8 @@ case class FpuCore( portCount : Int, p : FpuParameter) extends Component{
     //TODO maybe load can bypass merge and round.
     val inputs = ArrayBuffer[Stream[MergeInput]]()
     inputs += load.s1.output.stage()
+    if(p.withSqrt) (inputs += sqrt.output)
+    if(p.withDiv) (inputs += div.output)
     if(p.withAdd) (inputs += add.result.output)
     if(p.withMul) (inputs += mul.result.output)
     if(p.withShortPipMisc) (inputs += shortPip.rfOutput.pipelined(m2s = true))
@@ -1422,7 +1604,7 @@ case class FpuCore( portCount : Int, p : FpuParameter) extends Component{
     }
 
     when(input.valid){
-      for(i <- 0 until rfLockCount) when(input.lockId === i){
+      for(i <- 0 until p.rfLockCount) when(input.lockId === i){
         rf.lock(i).valid := False
       }
     }
@@ -1516,19 +1698,40 @@ object FpuSynthesisBench extends App{
     SpinalVerilog(new Component{
       val a = Delay(in UInt(width bits), 3)
       val sel = Delay(in UInt(log2Up(width) bits),3)
-//      val result =
+      //      val result =
-//      val output = Delay(result, 3)
+      //      val output = Delay(result, 3)
       setDefinitionName(Rotate3.this.getName())
     })
   }
 
+  class Div(width : Int) extends Rtl{
+    override def getName(): String = "div_" + width
+    override def getRtlPath(): String = getName() + ".v"
+    SpinalVerilog(new UnsignedDivider(width,width, false).setDefinitionName(Div.this.getName()))
+  }
 
+  class Add(width : Int) extends Rtl{
+    override def getName(): String = "add_" + width
+    override def getRtlPath(): String = getName() + ".v"
+    SpinalVerilog(new Component{
+      val a, b = in UInt(width bits)
+      val result = out(a + b)
+      setDefinitionName(Add.this.getName())
+    })
+  }
+
+  class DivSqrtRtl(width : Int) extends Rtl{
+    override def getName(): String = "DivSqrt_" + width
+    override def getRtlPath(): String = getName() + ".v"
+    SpinalVerilog(new FpuDiv(width).setDefinitionName(DivSqrtRtl.this.getName()))
+  }
 
   val rtls = ArrayBuffer[Rtl]()
   rtls += new Fpu(
     "32",
     portCount = 1,
     FpuParameter(
+//      withDivSqrt = false,
       withDouble = false
     )
   )
@@ -1536,11 +1739,18 @@ object FpuSynthesisBench extends App{
     "64",
     portCount = 1,
     FpuParameter(
+//      withDivSqrt = false,
       withDouble = true
     )
   )
 
-//  rtls += new Shifter(24)
+//  rtls += new Div(52)
+//  rtls += new Div(23)
+//  rtls += new Add(64)
+//  rtls += new DivSqrtRtl(52)
+//  rtls += new DivSqrtRtl(23)
+
+  //  rtls += new Shifter(24)
 //  rtls += new Shifter(32)
 //  rtls += new Shifter(52)
 //  rtls += new Shifter(64)
@@ -1557,4 +1767,28 @@ object FpuSynthesisBench extends App{
 
 
   Bench(rtls, targets)
-}
+}
+
+//Fpu_32 ->
+//Artix 7 -> 136 Mhz 1471 LUT 1336 FF
+//Artix 7 -> 196 Mhz 1687 LUT 1371 FF
+//Fpu_64 ->
+//Artix 7 -> 105 Mhz 2822 LUT 2132 FF
+//Artix 7 -> 161 Mhz 3114 LUT 2272 FF
+//
+//
+//
+//Fpu_32 ->
+//Artix 7 -> 128 Mhz 1693 LUT 1481 FF
+//Artix 7 -> 203 Mhz 1895 LUT 1481 FF
+//Fpu_64 ->
+//Artix 7 -> 99 Mhz 3073 LUT 2396 FF
+//Artix 7 -> 164 Mhz 3433 LUT 2432 FF
+
+
+//Fpu_32 ->
+//Artix 7 -> 112 Mhz 1790 LUT 1666 FF
+//Artix 7 -> 158 Mhz 1989 LUT 1701 FF
+//Fpu_64 ->
+//Artix 7 -> 100 Mhz 3294 LUT 2763 FF
+//Artix 7 -> 151 Mhz 3708 LUT 2904 FF

src/main/scala/vexriscv/ip/fpu/FpuDiv.scala

@@ -0,0 +1,128 @@
+package vexriscv.ip.fpu
+
+
+import spinal.core._
+import spinal.lib.math.{UnsignedDividerCmd, UnsignedDividerRsp}
+import spinal.lib._
+import spinal.lib.sim.{StreamDriver, StreamMonitor, StreamReadyRandomizer}
+
+import scala.collection.mutable
+import scala.util.Random
+
+case class FpuDivCmd(mantissaWidth : Int) extends Bundle{
+  val a,b = UInt(mantissaWidth bits)
+}
+
+case class FpuDivRsp(mantissaWidth : Int) extends Bundle{
+  val result = UInt(mantissaWidth+1 + 2 bits)
+  val remain = UInt(mantissaWidth+1 bits)
+}
+
+case class FpuDiv(val mantissaWidth : Int) extends Component {
+  assert(mantissaWidth % 2 == 0)
+  val io = new Bundle{
+    val input = slave Stream(FpuDivCmd(mantissaWidth))
+    val output = master Stream(FpuDivRsp(mantissaWidth))
+  }
+
+  val iterations = (mantissaWidth+2+2)/2
+  val counter = Reg(UInt(log2Up(iterations) bits))
+  val busy = RegInit(False) clearWhen(io.output.fire)
+  val done = RegInit(False) setWhen(busy && counter === iterations-1) clearWhen(io.output.fire)
+
+  val shifter = Reg(UInt(mantissaWidth + 3 bits))
+  val result = Reg(UInt(mantissaWidth+1+2 bits))
+
+  val div1, div3 = Reg(UInt(mantissaWidth+3 bits))
+  val div2 = div1 |<< 1
+
+  val sub1 = shifter -^ div1
+  val sub2 = shifter -^ div2
+  val sub3 = shifter -^ div3
+
+  io.output.valid := done
+  io.output.result := (result << 0).resized
+  io.output.remain := (shifter >> 2).resized
+  io.input.ready := !busy
+
+  when(!done){
+    counter := counter + 1
+    val sel = CombInit(shifter)
+    result := result |<< 2
+    when(!sub1.msb){
+      sel := sub1.resized
+      result(1 downto 0) := 1
+    }
+    when(!sub2.msb){
+      sel := sub2.resized
+      result(1 downto 0) := 2
+    }
+    when(!sub3.msb){
+      sel := sub3.resized
+      result(1 downto 0) := 3
+    }
+    shifter := sel |<< 2
+  }
+
+  when(!busy){
+    counter := 0
+    shifter := (U"1" @@ io.input.a @@ U"").resized
+    div1    := (U"1" @@ io.input.b).resized
+    div3    := (U"1" @@ io.input.b) +^ (((U"1" @@ io.input.b)) << 1)
+    busy := io.input.valid
+  }
+}
+
+
+object FpuDivTester extends App{
+  import spinal.core.sim._
+
+  for(w <- List(16, 20)) {
+    val config = SimConfig
+    config.withFstWave
+    config.compile(new FpuDiv(w)).doSim(seed=2){dut =>
+      dut.clockDomain.forkStimulus(10)
+
+
+      val (cmdDriver, cmdQueue) = StreamDriver.queue(dut.io.input, dut.clockDomain)
+      val rspQueue = mutable.Queue[FpuDivRsp => Unit]()
+      StreamMonitor(dut.io.output, dut.clockDomain)( rspQueue.dequeue()(_))
+      StreamReadyRandomizer(dut.io.output, dut.clockDomain)
+
+      def test(a : Int, b : Int): Unit ={
+        cmdQueue +={p =>
+          p.a #= a
+          p.b #= b
+        }
+        rspQueue += {p =>
+          val x = (a | (1 << dut.mantissaWidth)).toLong
+          val y = (b | (1 << dut.mantissaWidth)).toLong
+          val result = (x << dut.mantissaWidth+2) / y
+          val remain = (x << dut.mantissaWidth+2) % y
+
+          assert(p.result.toLong == result, f"$x%x/$y%x=${p.result.toLong}%x instead of $result%x")
+          assert(p.remain.toLong == remain, f"$x%x %% $y%x=${p.remain.toLong}%x instead of $remain%x")
+        }
+      }
+
+      val s = dut.mantissaWidth-16
+      val f = (1 << dut.mantissaWidth)-1
+      test(0xE000 << s, 0x8000 << s)
+      test(0xC000 << s, 0x4000 << s)
+      test(0xC835 << s, 0x4742 << s)
+      test(0,0)
+      test(0,f)
+      test(f,0)
+      test(f,f)
+
+      for(i <- 0 until 10000){
+        test(Random.nextInt(1 << dut.mantissaWidth), Random.nextInt(1 << dut.mantissaWidth))
+      }
+
+      waitUntil(rspQueue.isEmpty)
+
+      dut.clockDomain.waitSampling(100)
+
+    }
+  }
+}

src/main/scala/vexriscv/ip/fpu/FpuSqrt.scala

@@ -0,0 +1,116 @@
+package vexriscv.ip.fpu
+
+import spinal.core._
+import spinal.lib._
+import spinal.lib.sim.{StreamDriver, StreamMonitor, StreamReadyRandomizer}
+
+import scala.collection.mutable
+import scala.util.Random
+
+case class FpuSqrtCmd(mantissaWidth : Int) extends Bundle{
+  val a = UInt(mantissaWidth+2 bits)
+}
+
+case class FpuSqrtRsp(mantissaWidth : Int) extends Bundle{
+  val result = UInt(mantissaWidth+1 bits)
+  val remain = UInt(mantissaWidth+5 bits)
+}
+
+case class FpuSqrt(val mantissaWidth : Int) extends Component {
+  val io = new Bundle{
+    val input = slave Stream(FpuSqrtCmd(mantissaWidth))
+    val output = master Stream(FpuSqrtRsp(mantissaWidth))
+  }
+
+  val iterations = mantissaWidth+2
+  val counter = Reg(UInt(log2Up(iterations ) bits))
+  val busy = RegInit(False) clearWhen(io.output.fire)
+  val done = RegInit(False) setWhen(busy && counter === iterations-1) clearWhen(io.output.fire)
+
+  val a = Reg(UInt(mantissaWidth+5 bits))
+  val x = Reg(UInt(mantissaWidth bits))
+  val q = Reg(UInt(mantissaWidth+1 bits))
+  val t = a-(q @@ U"01")
+
+
+  io.output.valid := done
+  io.output.result := (q << 0).resized
+  io.output.remain := a
+  io.input.ready := !busy
+
+  when(!done){
+    counter := counter + 1
+    val sel = CombInit(a)
+    when(!t.msb){
+      sel := t.resized
+    }
+    q := (q @@ !t.msb).resized
+    a := (sel @@ x(widthOf(x)-2,2 bits)).resized
+    x := x |<< 2
+  }
+
+  when(!busy){
+    q := 0
+    a := io.input.a(widthOf(io.input.a)-2,2 bits).resized
+    x := (io.input.a).resized
+    counter := 0
+    when(io.input.valid){
+      busy := True
+    }
+  }
+}
+
+
+object FpuSqrtTester extends App{
+  import spinal.core.sim._
+
+  for(w <- List(16)) {
+    val config = SimConfig
+    config.withFstWave
+    config.compile(new FpuSqrt(w)).doSim(seed=2){dut =>
+      dut.clockDomain.forkStimulus(10)
+
+
+      val (cmdDriver, cmdQueue) = StreamDriver.queue(dut.io.input, dut.clockDomain)
+      val rspQueue = mutable.Queue[FpuSqrtRsp => Unit]()
+      StreamMonitor(dut.io.output, dut.clockDomain)( rspQueue.dequeue()(_))
+      StreamReadyRandomizer(dut.io.output, dut.clockDomain)
+
+      def test(a : Int): Unit ={
+        cmdQueue +={p =>
+          p.a #= a
+        }
+        rspQueue += {p =>
+//          val x = (a * (1l << dut.mantissaWidth)).toLong
+//          val result = Math.sqrt(x).toLong/(1 << dut.mantissaWidth/2)
+//          val remain = a-x*x
+          val x = a.toDouble / (1 << dut.mantissaWidth)
+          val result = (Math.sqrt(x)*(1 << dut.mantissaWidth+1)).toLong
+          val filtred = result  % (1 << dut.mantissaWidth+1)
+//          val remain = (a-(result*result)).toLong
+          assert(p.result.toLong == filtred, f"$a%x=${p.result.toLong}%x instead of $filtred%x")
+//          assert(p.remain.toLong == remain, f"$a%x=${p.remain.toLong}%x instead of $remain%x")
+        }
+      }
+
+      val s = dut.mantissaWidth-16
+      val f = (1 << dut.mantissaWidth)-1
+//      test(121)
+      test(0x20000)
+      test(0x18000)
+//      test(0,0)
+//      test(0,f)
+//      test(f,0)
+//      test(f,f)
+
+      for(i <- 0 until 10000){
+        test(Random.nextInt(3 << dut.mantissaWidth) + (1 << dut.mantissaWidth))
+      }
+
+      waitUntil(rspQueue.isEmpty)
+
+      dut.clockDomain.waitSampling(100)
+
+    }
+  }
+}

src/main/scala/vexriscv/ip/fpu/Interface.scala

@@ -50,7 +50,7 @@ case class FpuFloat(exponentSize: Int,
   def isInfinity  =  special && exponent(1 downto 0) === FpuFloat.INFINITY
   def isNan       =  special && exponent(1 downto 0) === FpuFloat.NAN
   def isQuiet     =  mantissa.msb
-  def isNanSignaling  =  special && exponent(1 downto 0) === FpuFloat.NAN && !isQuiet
+  def isNanSignaling  = special && exponent(1 downto 0) === FpuFloat.NAN && !isQuiet
   def isCanonical = exponent(FpuFloat.NAN_CANONICAL_BIT)
 
   def setNormal    =  { special := False }
@@ -118,10 +118,13 @@ object FpuRoundModeInstr extends SpinalEnum(){
 case class FpuParameter( withDouble : Boolean,
                          mulWidthA : Int = 18,
                          mulWidthB : Int = 18,
+                         rfLockCount : Int = 8,
                          sim : Boolean = false,
                          withAdd : Boolean = true,
                          withMul : Boolean = true,
-                         withDivSqrt : Boolean = true,
+                         withDivSqrt : Boolean = false,
+                         withDiv : Boolean = true,
+                         withSqrt : Boolean = true,
                          withShortPipMisc : Boolean = true){
 
   val internalMantissaSize = if(withDouble) 52 else 23

src/test/scala/vexriscv/ip/fpu/FpuTest.scala

@@ -55,11 +55,11 @@ class FpuTest extends FunSuite{
   }
 
   def testP(p : FpuParameter){
-    val portCount = 4
+    val portCount = 1
 
     val config = SimConfig
     config.allOptimisation
-//    if(p.withDouble) config.withFstWave
+//    config.withFstWave
     config.compile(new FpuCore(portCount, p){
       for(i <- 0 until portCount) out(Bits(5 bits)).setName(s"flagAcc$i") := io.port(i).completion.flags.asBits
       setDefinitionName("FpuCore"+ (if(p.withDouble) "Double" else  ""))
@@ -724,58 +724,34 @@ class FpuTest extends FunSuite{
           }
         }
 
-        def testSqrtExact(a : Float, ref : Float, flag : Int, rounding : FpuRoundMode.E): Unit ={
-          val rs = new RegAllocator()
-          val rs1, rs2, rs3 = rs.allocate()
-          val rd = Random.nextInt(32)
-          load(rs1, a)
-
-          sqrt(rd,rs1,  FpuRoundMode.RNE, FpuFormat.FLOAT)
-          storeFloat(rd){v =>
-            val error = Math.abs(ref-v)/ref
-            assert(checkFloat(ref, v), f"sqrt($a) = $v, $ref $error $rounding")
-          }
-        }
-
-        def testDivExact(a : Float, b : Float, ref : Float, flag : Int, rounding : FpuRoundMode.E): Unit ={
-          val rs = new RegAllocator()
-          val rs1, rs2, rs3 = rs.allocate()
-          val rd = Random.nextInt(32)
-          load(rs1, a)
-          load(rs2, b)
-
-          div(rd,rs1, rs2, FpuRoundMode.RNE, FpuFormat.FLOAT)
-          storeFloat(rd){v =>
-            val error = Math.abs(ref-v)/ref
-            assert(checkFloat(ref, v), f"div($a, $b) = $v, $ref $error $rounding")
-          }
-        }
-
         def testSqrtF64Exact(a : Double, ref : Double, flag : Int, rounding : FpuRoundMode.E): Unit ={
           val rs = new RegAllocator()
           val rs1, rs2, rs3 = rs.allocate()
           val rd = Random.nextInt(32)
           load(rs1, a)
 
-          sqrt(rd,rs1,  FpuRoundMode.RNE, FpuFormat.DOUBLE)
+          sqrt(rd,rs1,  rounding, FpuFormat.DOUBLE)
+
           store(rd){v =>
-            val error = Math.abs(ref-v)/ref
+            assert(d2b(v) == d2b(ref), f"## sqrt${a}   = $v, $ref $rounding, ${d2b(a).toString(16)} ${d2b(ref).toString(16)}")
-            assert(checkDouble(ref, v), f"sqrt($a) = $v, $ref $error $rounding")
           }
+
+          flagMatch(flag, ref, f"## sqrt${a} $ref $rounding")
         }
 
-        def testDivF64Exact(a : Double, b : Double, ref : Double, flag : Int, rounding : FpuRoundMode.E): Unit ={
+        def testSqrtExact(a : Float, ref : Float, flag : Int, rounding : FpuRoundMode.E): Unit ={
           val rs = new RegAllocator()
           val rs1, rs2, rs3 = rs.allocate()
           val rd = Random.nextInt(32)
           load(rs1, a)
-          load(rs2, b)
 
-          div(rd,rs1, rs2, FpuRoundMode.RNE, FpuFormat.DOUBLE)
+          sqrt(rd,rs1,  rounding, FpuFormat.FLOAT)
-          store(rd){v =>
+
-            val error = Math.abs(ref-v)/ref
+          storeFloat(rd){v =>
-            assert(checkDouble(ref, v), f"div($a, $b) = $v, $ref $error $rounding")
+            assert(d2b(v) == d2b(ref), f"## sqrt${a}   = $v, $ref $rounding, ${f2b(a).toString()} ${f2b(ref).toString()}")
           }
+
+          flagMatch(flag, ref, f"## sqrt${a} $ref $rounding")
         }
 
 
@@ -1108,8 +1084,7 @@ class FpuTest extends FunSuite{
         def testDiv() : Unit = {
           val rounding = FpuRoundMode.elements.randomPick()
           val (a,b,r,f) = f32.div(rounding).f32_f32_f32
-          testDivExact(a, b, r, f, rounding)
+          testBinaryOp(div, a, b, r, f, rounding, "div")
-          flagClear()
         }
 
         def testSqrt() : Unit = {
@@ -1132,7 +1107,8 @@ class FpuTest extends FunSuite{
         def testDivF64() : Unit = {
           val rounding = FpuRoundMode.elements.randomPick()
           val (a,b,r,f) = f64.div(rounding).f64_f64_f64
-          testDivF64Exact(a, b, r, f, rounding)
+         // testDivF64Exact(a, b, r, f, rounding)
+          testBinaryOpF64(div, a, b, r, f,rounding, "div")
           flagClear()
         }
 
@@ -1280,22 +1256,34 @@ class FpuTest extends FunSuite{
 
         var fxxTests = f32Tests
         if(p.withDouble) fxxTests ++= f64Tests
-        
+
-        
+        for(_ <- 0 until 10000) testDiv()
+        println("f32 div done")
+
+        for(_ <- 0 until 10000) testSqrt()
+        println("f32 sqrt done")
+
+
+
+
         //TODO test boxing
         //TODO double <-> simple convertions
         if(p.withDouble) {
 
-          load(0, 1.0)
+          testSqrtF64Exact(1.25*1.25, 1.25, 0, FpuRoundMode.RNE)
-          load(0, 2.0)
+          testSqrtF64Exact(1.5*1.5, 1.5, 0, FpuRoundMode.RNE)
-          load(0, 2.5)
+
-          load(0, 0.75)
+          for(_ <- 0 until 10000) testSqrtF64()
-          load(0, -5)
+          println("f64 sqrt done")
-          load(0, 0)
+
-          load(0, Double.PositiveInfinity)
+//          testDivF64Exact(1.0, 8.0, 0.125, 0, FpuRoundMode.RNE)
-          load(0, Double.NaN)
+//          testDivF64Exact(4.0, 8.0, 0.5, 0, FpuRoundMode.RNE)
-          dut.clockDomain.waitSampling(200)
+//          testDivF64Exact(8.0, 8.0, 1.0, 0, FpuRoundMode.RNE)
-          simSuccess()
+//          testDivF64Exact(1.5, 2.0, 0.75, 0, FpuRoundMode.RNE)
+//          testDivF64Exact(1.875, 1.5, 1.25, 0, FpuRoundMode.RNE)
+
+          for(_ <- 0 until 10000) testDivF64()
+          println("f64 div done")
 
           for(_ <- 0 until 10000) testSgnjF64()
           println("f64 sgnj done")
@@ -1338,12 +1326,6 @@ class FpuTest extends FunSuite{
           println("f64 Cmp done")
 
 
-          for(_ <- 0 until 10000) testDivF64()
-          println("f64 div done")
-
-          for(_ <- 0 until 10000) testSqrtF64()
-          println("f64 sqrt done")
-
           for(_ <- 0 until 10000) testClassF64()
           println("f64 class done")
 //