New FP encoding; improved FP implementation
This commit is contained in:
Andrew Waterman
parent
00bc1a2293
commit
a0389645b7
@ -25,8 +25,6 @@ class Core(implicit conf: RocketConfiguration) extends Module
|
||||
val fpu = Module(new FPU(conf.fpu.get))
|
||||
dpath.io.fpu <> fpu.io.dpath
|
||||
ctrl.io.fpu <> fpu.io.ctrl
|
||||
fpu.io.sfma.valid := Bool(false)
|
||||
fpu.io.dfma.valid := Bool(false)
|
||||
}
|
||||
|
||||
ctrl.io.dpath <> dpath.io.ctrl
|
||||
|
||||
@ -10,40 +10,25 @@ case class FPUConfig(sfmaLatency: Int = 2, dfmaLatency: Int = 3)
|
||||
|
||||
object FPConstants
|
||||
{
|
||||
val FCMD_ADD = Bits("b000000")
|
||||
val FCMD_SUB = Bits("b000001")
|
||||
val FCMD_MUL = Bits("b000010")
|
||||
val FCMD_DIV = Bits("b000011")
|
||||
val FCMD_SQRT = Bits("b000100")
|
||||
val FCMD_SGNJ = Bits("b000101")
|
||||
val FCMD_SGNJN = Bits("b000110")
|
||||
val FCMD_SGNJX = Bits("b000111")
|
||||
val FCMD_CVT_L_FMT = Bits("b001000")
|
||||
val FCMD_CVT_LU_FMT = Bits("b001001")
|
||||
val FCMD_CVT_W_FMT = Bits("b001010")
|
||||
val FCMD_CVT_WU_FMT = Bits("b001011")
|
||||
val FCMD_CVT_FMT_L = Bits("b001100")
|
||||
val FCMD_CVT_FMT_LU = Bits("b001101")
|
||||
val FCMD_CVT_FMT_W = Bits("b001110")
|
||||
val FCMD_CVT_FMT_WU = Bits("b001111")
|
||||
val FCMD_CVT_FMT_S = Bits("b010000")
|
||||
val FCMD_CVT_FMT_D = Bits("b010001")
|
||||
val FCMD_EQ = Bits("b010101")
|
||||
val FCMD_LT = Bits("b010110")
|
||||
val FCMD_LE = Bits("b010111")
|
||||
val FCMD_MIN = Bits("b011000")
|
||||
val FCMD_MAX = Bits("b011001")
|
||||
val FCMD_MFTX = Bits("b011100")
|
||||
val FCMD_CLASS = Bits("b011101")
|
||||
val FCMD_MXTF = Bits("b011110")
|
||||
val FCMD_MADD = Bits("b100100")
|
||||
val FCMD_MSUB = Bits("b100101")
|
||||
val FCMD_NMSUB = Bits("b100110")
|
||||
val FCMD_NMADD = Bits("b100111")
|
||||
val FCMD_LOAD = Bits("b111000")
|
||||
val FCMD_STORE = Bits("b111001")
|
||||
val FCMD_X = Bits("b??????")
|
||||
val FCMD_WIDTH = 6
|
||||
val FCMD_ADD = Bits("b0??00")
|
||||
val FCMD_SUB = Bits("b0??01")
|
||||
val FCMD_MUL = Bits("b0??10")
|
||||
val FCMD_MADD = Bits("b1??00")
|
||||
val FCMD_MSUB = Bits("b1??01")
|
||||
val FCMD_NMSUB = Bits("b1??10")
|
||||
val FCMD_NMADD = Bits("b1??11")
|
||||
val FCMD_DIV = Bits("b?0?11")
|
||||
val FCMD_SQRT = Bits("b?1?11")
|
||||
val FCMD_SGNJ = Bits("b??1?0")
|
||||
val FCMD_MINMAX = Bits("b?01?1")
|
||||
val FCMD_CVT_FF = Bits("b??0??")
|
||||
val FCMD_CVT_IF = Bits("b?10??")
|
||||
val FCMD_CMP = Bits("b?01??")
|
||||
val FCMD_MV_XF = Bits("b?11??")
|
||||
val FCMD_CVT_FI = Bits("b??0??")
|
||||
val FCMD_MV_FX = Bits("b??1??")
|
||||
val FCMD_X = Bits("b?????")
|
||||
val FCMD_WIDTH = 5
|
||||
|
||||
val RM_SZ = 3
|
||||
val FLAGS_SZ = 5
|
||||
@ -52,10 +37,12 @@ object FPConstants
|
||||
class FPUCtrlSigs extends Bundle
|
||||
{
|
||||
val cmd = Bits(width = FCMD_WIDTH)
|
||||
val ldst = Bool()
|
||||
val wen = Bool()
|
||||
val ren1 = Bool()
|
||||
val ren2 = Bool()
|
||||
val ren3 = Bool()
|
||||
val swap23 = Bool()
|
||||
val single = Bool()
|
||||
val fromint = Bool()
|
||||
val toint = Bool()
|
||||
@ -75,79 +62,69 @@ class FPUDecoder extends Module
|
||||
val Y = Bool(true)
|
||||
val X = Bool(false)
|
||||
val decoder = DecodeLogic(io.inst,
|
||||
List (FCMD_X, X,X,X,X,X,X,X,X,X,X),
|
||||
Array(FLW -> List(FCMD_LOAD, Y,N,N,N,Y,N,N,N,N,N),
|
||||
FLD -> List(FCMD_LOAD, Y,N,N,N,N,N,N,N,N,N),
|
||||
FSW -> List(FCMD_STORE, N,N,Y,N,Y,N,Y,N,N,N),
|
||||
FSD -> List(FCMD_STORE, N,N,Y,N,N,N,Y,N,N,N),
|
||||
FMV_S_X -> List(FCMD_MXTF, Y,N,N,N,Y,Y,N,N,N,Y),
|
||||
FMV_D_X -> List(FCMD_MXTF, Y,N,N,N,N,Y,N,N,N,Y),
|
||||
FCVT_S_W -> List(FCMD_CVT_FMT_W, Y,N,N,N,Y,Y,N,N,N,Y),
|
||||
FCVT_S_WU-> List(FCMD_CVT_FMT_WU,Y,N,N,N,Y,Y,N,N,N,Y),
|
||||
FCVT_S_L -> List(FCMD_CVT_FMT_L, Y,N,N,N,Y,Y,N,N,N,Y),
|
||||
FCVT_S_LU-> List(FCMD_CVT_FMT_LU,Y,N,N,N,Y,Y,N,N,N,Y),
|
||||
FCVT_D_W -> List(FCMD_CVT_FMT_W, Y,N,N,N,N,Y,N,N,N,Y),
|
||||
FCVT_D_WU-> List(FCMD_CVT_FMT_WU,Y,N,N,N,N,Y,N,N,N,Y),
|
||||
FCVT_D_L -> List(FCMD_CVT_FMT_L, Y,N,N,N,N,Y,N,N,N,Y),
|
||||
FCVT_D_LU-> List(FCMD_CVT_FMT_LU,Y,N,N,N,N,Y,N,N,N,Y),
|
||||
FMV_X_S -> List(FCMD_MFTX, N,Y,N,N,Y,N,Y,N,N,Y),
|
||||
FMV_X_D -> List(FCMD_MFTX, N,Y,N,N,N,N,Y,N,N,Y),
|
||||
FCLASS_S -> List(FCMD_CLASS, N,Y,N,N,Y,N,Y,N,N,Y),
|
||||
FCLASS_D -> List(FCMD_CLASS, N,Y,N,N,N,N,Y,N,N,Y),
|
||||
FCVT_W_S -> List(FCMD_CVT_W_FMT, N,Y,N,N,Y,N,Y,N,N,Y),
|
||||
FCVT_WU_S-> List(FCMD_CVT_WU_FMT,N,Y,N,N,Y,N,Y,N,N,Y),
|
||||
FCVT_L_S -> List(FCMD_CVT_L_FMT, N,Y,N,N,Y,N,Y,N,N,Y),
|
||||
FCVT_LU_S-> List(FCMD_CVT_LU_FMT,N,Y,N,N,Y,N,Y,N,N,Y),
|
||||
FCVT_W_D -> List(FCMD_CVT_W_FMT, N,Y,N,N,N,N,Y,N,N,Y),
|
||||
FCVT_WU_D-> List(FCMD_CVT_WU_FMT,N,Y,N,N,N,N,Y,N,N,Y),
|
||||
FCVT_L_D -> List(FCMD_CVT_L_FMT, N,Y,N,N,N,N,Y,N,N,Y),
|
||||
FCVT_LU_D-> List(FCMD_CVT_LU_FMT,N,Y,N,N,N,N,Y,N,N,Y),
|
||||
FCVT_S_D -> List(FCMD_CVT_FMT_D, Y,Y,N,N,Y,N,N,Y,N,Y),
|
||||
FCVT_D_S -> List(FCMD_CVT_FMT_S, Y,Y,N,N,N,N,N,Y,N,Y),
|
||||
FEQ_S -> List(FCMD_EQ, N,Y,Y,N,Y,N,Y,N,N,Y),
|
||||
FLT_S -> List(FCMD_LT, N,Y,Y,N,Y,N,Y,N,N,Y),
|
||||
FLE_S -> List(FCMD_LE, N,Y,Y,N,Y,N,Y,N,N,Y),
|
||||
FEQ_D -> List(FCMD_EQ, N,Y,Y,N,N,N,Y,N,N,Y),
|
||||
FLT_D -> List(FCMD_LT, N,Y,Y,N,N,N,Y,N,N,Y),
|
||||
FLE_D -> List(FCMD_LE, N,Y,Y,N,N,N,Y,N,N,Y),
|
||||
FSGNJ_S -> List(FCMD_SGNJ, Y,Y,Y,N,Y,N,N,Y,N,Y),
|
||||
FSGNJN_S -> List(FCMD_SGNJN, Y,Y,Y,N,Y,N,N,Y,N,Y),
|
||||
FSGNJX_S -> List(FCMD_SGNJX, Y,Y,Y,N,Y,N,N,Y,N,Y),
|
||||
FSGNJ_D -> List(FCMD_SGNJ, Y,Y,Y,N,N,N,N,Y,N,Y),
|
||||
FSGNJN_D -> List(FCMD_SGNJN, Y,Y,Y,N,N,N,N,Y,N,Y),
|
||||
FSGNJX_D -> List(FCMD_SGNJX, Y,Y,Y,N,N,N,N,Y,N,Y),
|
||||
FMIN_S -> List(FCMD_MIN, Y,Y,Y,N,Y,N,Y,Y,N,Y),
|
||||
FMAX_S -> List(FCMD_MAX, Y,Y,Y,N,Y,N,Y,Y,N,Y),
|
||||
FMIN_D -> List(FCMD_MIN, Y,Y,Y,N,N,N,Y,Y,N,Y),
|
||||
FMAX_D -> List(FCMD_MAX, Y,Y,Y,N,N,N,Y,Y,N,Y),
|
||||
FADD_S -> List(FCMD_ADD, Y,Y,Y,N,Y,N,N,N,Y,Y),
|
||||
FSUB_S -> List(FCMD_SUB, Y,Y,Y,N,Y,N,N,N,Y,Y),
|
||||
FMUL_S -> List(FCMD_MUL, Y,Y,Y,N,Y,N,N,N,Y,Y),
|
||||
FADD_D -> List(FCMD_ADD, Y,Y,Y,N,N,N,N,N,Y,Y),
|
||||
FSUB_D -> List(FCMD_SUB, Y,Y,Y,N,N,N,N,N,Y,Y),
|
||||
FMUL_D -> List(FCMD_MUL, Y,Y,Y,N,N,N,N,N,Y,Y),
|
||||
FMADD_S -> List(FCMD_MADD, Y,Y,Y,Y,Y,N,N,N,Y,Y),
|
||||
FMSUB_S -> List(FCMD_MSUB, Y,Y,Y,Y,Y,N,N,N,Y,Y),
|
||||
FNMADD_S -> List(FCMD_NMADD, Y,Y,Y,Y,Y,N,N,N,Y,Y),
|
||||
FNMSUB_S -> List(FCMD_NMSUB, Y,Y,Y,Y,Y,N,N,N,Y,Y),
|
||||
FMADD_D -> List(FCMD_MADD, Y,Y,Y,Y,N,N,N,N,Y,Y),
|
||||
FMSUB_D -> List(FCMD_MSUB, Y,Y,Y,Y,N,N,N,N,Y,Y),
|
||||
FNMADD_D -> List(FCMD_NMADD, Y,Y,Y,Y,N,N,N,N,Y,Y),
|
||||
FNMSUB_D -> List(FCMD_NMSUB, Y,Y,Y,Y,N,N,N,N,Y,Y)
|
||||
List (FCMD_X, X,X,X,X,X,X,X,X,X,X,X,X),
|
||||
Array(FLW -> List(FCMD_X, Y,Y,N,N,N,X,Y,N,N,N,N,N),
|
||||
FLD -> List(FCMD_X, Y,Y,N,N,N,X,N,N,N,N,N,N),
|
||||
FSW -> List(FCMD_MV_XF, Y,N,N,Y,N,X,Y,N,Y,N,N,N),
|
||||
FSD -> List(FCMD_MV_XF, Y,N,N,Y,N,X,N,N,Y,N,N,N),
|
||||
FMV_S_X -> List(FCMD_MV_FX, N,Y,N,N,N,X,Y,Y,N,N,N,Y),
|
||||
FMV_D_X -> List(FCMD_MV_FX, N,Y,N,N,N,X,N,Y,N,N,N,Y),
|
||||
FCVT_S_W -> List(FCMD_CVT_FI, N,Y,N,N,N,X,Y,Y,N,N,N,Y),
|
||||
FCVT_S_WU-> List(FCMD_CVT_FI, N,Y,N,N,N,X,Y,Y,N,N,N,Y),
|
||||
FCVT_S_L -> List(FCMD_CVT_FI, N,Y,N,N,N,X,Y,Y,N,N,N,Y),
|
||||
FCVT_S_LU-> List(FCMD_CVT_FI, N,Y,N,N,N,X,Y,Y,N,N,N,Y),
|
||||
FCVT_D_W -> List(FCMD_CVT_FI, N,Y,N,N,N,X,N,Y,N,N,N,Y),
|
||||
FCVT_D_WU-> List(FCMD_CVT_FI, N,Y,N,N,N,X,N,Y,N,N,N,Y),
|
||||
FCVT_D_L -> List(FCMD_CVT_FI, N,Y,N,N,N,X,N,Y,N,N,N,Y),
|
||||
FCVT_D_LU-> List(FCMD_CVT_FI, N,Y,N,N,N,X,N,Y,N,N,N,Y),
|
||||
FMV_X_S -> List(FCMD_MV_XF, N,N,Y,N,N,X,Y,N,Y,N,N,Y),
|
||||
FMV_X_D -> List(FCMD_MV_XF, N,N,Y,N,N,X,N,N,Y,N,N,Y),
|
||||
FCLASS_S -> List(FCMD_MV_XF, N,N,Y,N,N,X,Y,N,Y,N,N,Y),
|
||||
FCLASS_D -> List(FCMD_MV_XF, N,N,Y,N,N,X,N,N,Y,N,N,Y),
|
||||
FCVT_W_S -> List(FCMD_CVT_IF, N,N,Y,N,N,X,Y,N,Y,N,N,Y),
|
||||
FCVT_WU_S-> List(FCMD_CVT_IF, N,N,Y,N,N,X,Y,N,Y,N,N,Y),
|
||||
FCVT_L_S -> List(FCMD_CVT_IF, N,N,Y,N,N,X,Y,N,Y,N,N,Y),
|
||||
FCVT_LU_S-> List(FCMD_CVT_IF, N,N,Y,N,N,X,Y,N,Y,N,N,Y),
|
||||
FCVT_W_D -> List(FCMD_CVT_IF, N,N,Y,N,N,X,N,N,Y,N,N,Y),
|
||||
FCVT_WU_D-> List(FCMD_CVT_IF, N,N,Y,N,N,X,N,N,Y,N,N,Y),
|
||||
FCVT_L_D -> List(FCMD_CVT_IF, N,N,Y,N,N,X,N,N,Y,N,N,Y),
|
||||
FCVT_LU_D-> List(FCMD_CVT_IF, N,N,Y,N,N,X,N,N,Y,N,N,Y),
|
||||
FCVT_S_D -> List(FCMD_CVT_FF, N,Y,Y,N,N,X,Y,N,N,Y,N,Y),
|
||||
FCVT_D_S -> List(FCMD_CVT_FF, N,Y,Y,N,N,X,N,N,N,Y,N,Y),
|
||||
FEQ_S -> List(FCMD_CMP, N,N,Y,Y,N,N,Y,N,Y,N,N,N),
|
||||
FLT_S -> List(FCMD_CMP, N,N,Y,Y,N,N,Y,N,Y,N,N,N),
|
||||
FLE_S -> List(FCMD_CMP, N,N,Y,Y,N,N,Y,N,Y,N,N,N),
|
||||
FEQ_D -> List(FCMD_CMP, N,N,Y,Y,N,N,N,N,Y,N,N,N),
|
||||
FLT_D -> List(FCMD_CMP, N,N,Y,Y,N,N,N,N,Y,N,N,N),
|
||||
FLE_D -> List(FCMD_CMP, N,N,Y,Y,N,N,N,N,Y,N,N,N),
|
||||
FSGNJ_S -> List(FCMD_SGNJ, N,Y,Y,Y,N,N,Y,N,N,Y,N,N),
|
||||
FSGNJN_S -> List(FCMD_SGNJ, N,Y,Y,Y,N,N,Y,N,N,Y,N,N),
|
||||
FSGNJX_S -> List(FCMD_SGNJ, N,Y,Y,Y,N,N,Y,N,N,Y,N,N),
|
||||
FSGNJ_D -> List(FCMD_SGNJ, N,Y,Y,Y,N,N,N,N,N,Y,N,N),
|
||||
FSGNJN_D -> List(FCMD_SGNJ, N,Y,Y,Y,N,N,N,N,N,Y,N,N),
|
||||
FSGNJX_D -> List(FCMD_SGNJ, N,Y,Y,Y,N,N,N,N,N,Y,N,N),
|
||||
FMIN_S -> List(FCMD_MINMAX, N,Y,Y,Y,N,N,Y,N,Y,Y,N,N),
|
||||
FMAX_S -> List(FCMD_MINMAX, N,Y,Y,Y,N,N,Y,N,Y,Y,N,N),
|
||||
FMIN_D -> List(FCMD_MINMAX, N,Y,Y,Y,N,N,N,N,Y,Y,N,N),
|
||||
FMAX_D -> List(FCMD_MINMAX, N,Y,Y,Y,N,N,N,N,Y,Y,N,N),
|
||||
FADD_S -> List(FCMD_ADD, N,Y,Y,Y,N,Y,Y,N,N,N,Y,Y),
|
||||
FSUB_S -> List(FCMD_SUB, N,Y,Y,Y,N,Y,Y,N,N,N,Y,Y),
|
||||
FMUL_S -> List(FCMD_MUL, N,Y,Y,Y,N,N,Y,N,N,N,Y,Y),
|
||||
FADD_D -> List(FCMD_ADD, N,Y,Y,Y,N,Y,N,N,N,N,Y,Y),
|
||||
FSUB_D -> List(FCMD_SUB, N,Y,Y,Y,N,Y,N,N,N,N,Y,Y),
|
||||
FMUL_D -> List(FCMD_MUL, N,Y,Y,Y,N,N,N,N,N,N,Y,Y),
|
||||
FMADD_S -> List(FCMD_MADD, N,Y,Y,Y,Y,N,Y,N,N,N,Y,Y),
|
||||
FMSUB_S -> List(FCMD_MSUB, N,Y,Y,Y,Y,N,Y,N,N,N,Y,Y),
|
||||
FNMADD_S -> List(FCMD_NMADD, N,Y,Y,Y,Y,N,Y,N,N,N,Y,Y),
|
||||
FNMSUB_S -> List(FCMD_NMSUB, N,Y,Y,Y,Y,N,Y,N,N,N,Y,Y),
|
||||
FMADD_D -> List(FCMD_MADD, N,Y,Y,Y,Y,N,N,N,N,N,Y,Y),
|
||||
FMSUB_D -> List(FCMD_MSUB, N,Y,Y,Y,Y,N,N,N,N,N,Y,Y),
|
||||
FNMADD_D -> List(FCMD_NMADD, N,Y,Y,Y,Y,N,N,N,N,N,Y,Y),
|
||||
FNMSUB_D -> List(FCMD_NMSUB, N,Y,Y,Y,Y,N,N,N,N,N,Y,Y)
|
||||
))
|
||||
val cmd :: wen :: ren1 :: ren2 :: ren3 :: single :: fromint :: toint :: fastpipe :: fma :: round :: Nil = decoder
|
||||
|
||||
io.sigs.cmd := cmd
|
||||
io.sigs.wen := wen.toBool
|
||||
io.sigs.ren1 := ren1.toBool
|
||||
io.sigs.ren2 := ren2.toBool
|
||||
io.sigs.ren3 := ren3.toBool
|
||||
io.sigs.single := single.toBool
|
||||
io.sigs.fromint := fromint.toBool
|
||||
io.sigs.toint := toint.toBool
|
||||
io.sigs.fastpipe := fastpipe.toBool
|
||||
io.sigs.fma := fma.toBool
|
||||
io.sigs.round := round.toBool
|
||||
val s = io.sigs
|
||||
Vec(s.cmd, s.ldst, s.wen, s.ren1, s.ren2, s.ren3, s.swap23, s.single, s.fromint,
|
||||
s.toint, s.fastpipe, s.fma, s.round) := decoder
|
||||
}
|
||||
|
||||
class DpathFPUIO extends Bundle {
|
||||
@ -179,18 +156,25 @@ class CtrlFPUIO extends Bundle {
|
||||
val sboard_clra = UInt(INPUT, 5)
|
||||
}
|
||||
|
||||
class FPToInt extends Module
|
||||
class FPResult extends Bundle
|
||||
{
|
||||
class Input extends Bundle {
|
||||
val single = Bool()
|
||||
val cmd = Bits(width = FCMD_WIDTH)
|
||||
val data = Bits(width = 65)
|
||||
val exc = Bits(width = 5)
|
||||
}
|
||||
|
||||
class FPInput extends FPUCtrlSigs {
|
||||
val rm = Bits(width = 3)
|
||||
val typ = Bits(width = 2)
|
||||
val in1 = Bits(width = 65)
|
||||
val in2 = Bits(width = 65)
|
||||
override def clone = new Input().asInstanceOf[this.type]
|
||||
val in3 = Bits(width = 65)
|
||||
override def clone = new FPInput().asInstanceOf[this.type]
|
||||
}
|
||||
|
||||
class FPToInt extends Module
|
||||
{
|
||||
val io = new Bundle {
|
||||
val in = Valid(new Input).flip
|
||||
val in = Valid(new FPInput).flip
|
||||
val out = Valid(new Bundle {
|
||||
val lt = Bool()
|
||||
val store = Bits(width = 64)
|
||||
@ -199,77 +183,54 @@ class FPToInt extends Module
|
||||
})
|
||||
}
|
||||
|
||||
val in = Reg(new Input)
|
||||
val in = Reg(new FPInput)
|
||||
val valid = Reg(next=io.in.valid)
|
||||
when (io.in.valid) {
|
||||
def upconvert(x: UInt) = hardfloat.recodedFloatNToRecodedFloatM(x, Bits(0), 23, 9, 52, 12)._1
|
||||
when (io.in.bits.cmd === FCMD_STORE) {
|
||||
in.in1 := io.in.bits.in2
|
||||
}.otherwise {
|
||||
val doUpconvert = io.in.bits.single && io.in.bits.cmd != FCMD_MFTX && io.in.bits.cmd != FCMD_CLASS
|
||||
in.in1 := Mux(doUpconvert, upconvert(io.in.bits.in1), io.in.bits.in1)
|
||||
in.in2 := Mux(doUpconvert, upconvert(io.in.bits.in2), io.in.bits.in2)
|
||||
in := io.in.bits
|
||||
when (io.in.bits.single && !io.in.bits.ldst && io.in.bits.cmd != FCMD_MV_XF) {
|
||||
in.in1 := upconvert(io.in.bits.in1)
|
||||
in.in2 := upconvert(io.in.bits.in2)
|
||||
}
|
||||
in.single := io.in.bits.single
|
||||
in.cmd := io.in.bits.cmd
|
||||
in.rm := io.in.bits.rm
|
||||
}
|
||||
|
||||
val unrec_s = hardfloat.recodedFloatNToFloatN(in.in1, 23, 9)
|
||||
val unrec_d = hardfloat.recodedFloatNToFloatN(in.in1, 52, 12)
|
||||
val unrec_out = Mux(in.single, Cat(Fill(32, unrec_s(31)), unrec_s), unrec_d)
|
||||
|
||||
val classify_s = hardfloat.recodedFloatNClassify(in.in1, 23, 9)
|
||||
val classify_d = hardfloat.recodedFloatNClassify(in.in1, 52, 12)
|
||||
val classify_out = Mux(in.single, classify_s, classify_d)
|
||||
|
||||
val dcmp = Module(new hardfloat.recodedFloatNCompare(52, 12))
|
||||
dcmp.io.a := in.in1
|
||||
dcmp.io.b := in.in2
|
||||
val dcmp_out = (in.cmd & Cat(dcmp.io.a_lt_b, dcmp.io.a_eq_b)).orR
|
||||
val dcmp_exc = (in.cmd & Cat(dcmp.io.a_lt_b_invalid, dcmp.io.a_eq_b_invalid)).orR << UInt(4)
|
||||
val dcmp_out = (~in.rm & Cat(dcmp.io.a_lt_b, dcmp.io.a_eq_b)).orR
|
||||
val dcmp_exc = (~in.rm & Cat(dcmp.io.a_lt_b_invalid, dcmp.io.a_eq_b_invalid)).orR << UInt(4)
|
||||
|
||||
val d2i = hardfloat.recodedFloatNToAny(in.in1, in.rm, ~in.cmd(1,0), 52, 12, 64)
|
||||
val d2i = hardfloat.recodedFloatNToAny(in.in1, in.rm, in.typ ^ 1, 52, 12, 64)
|
||||
|
||||
io.out.bits.toint := unrec_out
|
||||
io.out.bits.toint := Mux(in.rm(0), classify_out, unrec_out)
|
||||
io.out.bits.store := unrec_out
|
||||
io.out.bits.exc := Bits(0)
|
||||
|
||||
when (in.cmd === FCMD_CVT_W_FMT || in.cmd === FCMD_CVT_WU_FMT) {
|
||||
io.out.bits.toint := Cat(Fill(32, d2i._1(31)), d2i._1(31,0))
|
||||
io.out.bits.exc := d2i._2
|
||||
}
|
||||
when (in.cmd === FCMD_CVT_L_FMT || in.cmd === FCMD_CVT_LU_FMT) {
|
||||
io.out.bits.toint := d2i._1
|
||||
io.out.bits.exc := d2i._2
|
||||
}
|
||||
when (in.cmd === FCMD_CLASS) {
|
||||
val classify_s = hardfloat.recodedFloatNClassify(in.in1, 23, 9)
|
||||
val classify_d = hardfloat.recodedFloatNClassify(in.in1, 52, 12)
|
||||
io.out.bits.toint := Mux(in.single, classify_s, classify_d)
|
||||
}
|
||||
when (in.cmd === FCMD_EQ || in.cmd === FCMD_LT || in.cmd === FCMD_LE) {
|
||||
when (in.cmd === FCMD_CMP) {
|
||||
io.out.bits.toint := dcmp_out
|
||||
io.out.bits.exc := dcmp_exc
|
||||
}
|
||||
when (in.cmd === FCMD_CVT_IF) {
|
||||
io.out.bits.toint := Mux(in.typ(1), d2i._1, d2i._1(31,0).toSInt)
|
||||
io.out.bits.exc := d2i._2
|
||||
}
|
||||
|
||||
io.out.valid := valid
|
||||
io.out.bits.lt := dcmp.io.a_lt_b
|
||||
}
|
||||
|
||||
class FPResult extends Bundle
|
||||
{
|
||||
val data = Bits(width = 65)
|
||||
val exc = Bits(width = 5)
|
||||
}
|
||||
|
||||
class IntToFP(val latency: Int) extends Module
|
||||
{
|
||||
class Input extends Bundle {
|
||||
val single = Bool()
|
||||
val cmd = Bits(width = FCMD_WIDTH)
|
||||
val rm = Bits(width = 3)
|
||||
val data = Bits(width = 64)
|
||||
override def clone = new Input().asInstanceOf[this.type]
|
||||
}
|
||||
val io = new Bundle {
|
||||
val in = Valid(new Input).flip
|
||||
val in = Valid(new FPInput).flip
|
||||
val out = Valid(new FPResult)
|
||||
}
|
||||
|
||||
@ -277,19 +238,18 @@ class IntToFP(val latency: Int) extends Module
|
||||
|
||||
val mux = new FPResult
|
||||
mux.exc := Bits(0)
|
||||
mux.data := hardfloat.floatNToRecodedFloatN(in.bits.data, 52, 12)
|
||||
mux.data := hardfloat.floatNToRecodedFloatN(in.bits.in1, 52, 12)
|
||||
when (in.bits.single) {
|
||||
mux.data := Cat(SInt(-1, 32), hardfloat.floatNToRecodedFloatN(in.bits.data, 23, 9))
|
||||
mux.data := Cat(SInt(-1, 32), hardfloat.floatNToRecodedFloatN(in.bits.in1, 23, 9))
|
||||
}
|
||||
|
||||
when (in.bits.cmd === FCMD_CVT_FMT_W || in.bits.cmd === FCMD_CVT_FMT_WU ||
|
||||
in.bits.cmd === FCMD_CVT_FMT_L || in.bits.cmd === FCMD_CVT_FMT_LU) {
|
||||
when (in.bits.cmd === FCMD_CVT_FI) {
|
||||
when (in.bits.single) {
|
||||
val u = hardfloat.anyToRecodedFloatN(in.bits.data, in.bits.rm, ~in.bits.cmd(1,0), 23, 9, 64)
|
||||
val u = hardfloat.anyToRecodedFloatN(in.bits.in1(63,0), in.bits.rm, in.bits.typ ^ 1, 23, 9, 64)
|
||||
mux.data := Cat(SInt(-1, 32), u._1)
|
||||
mux.exc := u._2
|
||||
}.otherwise {
|
||||
val u = hardfloat.anyToRecodedFloatN(in.bits.data, in.bits.rm, ~in.bits.cmd(1,0), 52, 12, 64)
|
||||
val u = hardfloat.anyToRecodedFloatN(in.bits.in1(63,0), in.bits.rm, in.bits.typ ^ 1, 52, 12, 64)
|
||||
mux.data := u._1
|
||||
mux.exc := u._2
|
||||
}
|
||||
@ -300,16 +260,8 @@ class IntToFP(val latency: Int) extends Module
|
||||
|
||||
class FPToFP(val latency: Int) extends Module
|
||||
{
|
||||
class Input extends Bundle {
|
||||
val single = Bool()
|
||||
val cmd = Bits(width = FCMD_WIDTH)
|
||||
val rm = Bits(width = 3)
|
||||
val in1 = Bits(width = 65)
|
||||
val in2 = Bits(width = 65)
|
||||
override def clone = new Input().asInstanceOf[this.type]
|
||||
}
|
||||
val io = new Bundle {
|
||||
val in = Valid(new Input).flip
|
||||
val in = Valid(new FPInput).flip
|
||||
val out = Valid(new FPResult)
|
||||
val lt = Bool(INPUT) // from FPToInt
|
||||
}
|
||||
@ -317,34 +269,31 @@ class FPToFP(val latency: Int) extends Module
|
||||
val in = Pipe(io.in)
|
||||
|
||||
// fp->fp units
|
||||
val sign_s = Mux(in.bits.cmd === FCMD_SGNJ, in.bits.in2(32),
|
||||
Mux(in.bits.cmd === FCMD_SGNJN, ~in.bits.in2(32),
|
||||
in.bits.in1(32) ^ in.bits.in2(32))) // FCMD_SGNJX
|
||||
val sign_d = Mux(in.bits.cmd === FCMD_SGNJ, in.bits.in2(64),
|
||||
Mux(in.bits.cmd === FCMD_SGNJN, ~in.bits.in2(64),
|
||||
in.bits.in1(64) ^ in.bits.in2(64))) // FCMD_SGNJX
|
||||
val fsgnj = Cat(Mux(in.bits.single, in.bits.in1(64), sign_d), in.bits.in1(63,33),
|
||||
Mux(in.bits.single, sign_s, in.bits.in1(32)), in.bits.in1(31,0))
|
||||
val isSgnj = in.bits.cmd === FCMD_SGNJ
|
||||
def fsgnjSign(in1: Bits, in2: Bits, pos: Int, en: Bool, rm: Bits) =
|
||||
Mux(rm(1) || !en, in1(pos), rm(0)) ^ (en && in2(pos))
|
||||
val sign_s = fsgnjSign(in.bits.in1, in.bits.in2, 32, in.bits.single && isSgnj, in.bits.rm)
|
||||
val sign_d = fsgnjSign(in.bits.in1, in.bits.in2, 64, !in.bits.single && isSgnj, in.bits.rm)
|
||||
val fsgnj = Cat(sign_d, in.bits.in1(63,33), sign_s, in.bits.in1(31,0))
|
||||
|
||||
val s2d = hardfloat.recodedFloatNToRecodedFloatM(in.bits.in1, in.bits.rm, 23, 9, 52, 12)
|
||||
val d2s = hardfloat.recodedFloatNToRecodedFloatM(in.bits.in1, in.bits.rm, 52, 12, 23, 9)
|
||||
|
||||
val isnan1 = Mux(in.bits.single, in.bits.in1(31,29) === Bits("b111"), in.bits.in1(63,61) === Bits("b111"))
|
||||
val isnan2 = Mux(in.bits.single, in.bits.in2(31,29) === Bits("b111"), in.bits.in2(63,61) === Bits("b111"))
|
||||
val isnan1 = Mux(in.bits.single, in.bits.in1(31,29).andR, in.bits.in1(63,61).andR)
|
||||
val isnan2 = Mux(in.bits.single, in.bits.in2(31,29).andR, in.bits.in2(63,61).andR)
|
||||
val issnan1 = isnan1 && ~Mux(in.bits.single, in.bits.in1(22), in.bits.in1(51))
|
||||
val issnan2 = isnan2 && ~Mux(in.bits.single, in.bits.in2(22), in.bits.in2(51))
|
||||
val minmax_exc = Cat(issnan1 || issnan2, Bits(0,4))
|
||||
val min = in.bits.cmd === FCMD_MIN
|
||||
val minmax = Mux(isnan2 || !isnan1 && (min === io.lt), in.bits.in1, in.bits.in2)
|
||||
val isMax = in.bits.rm(0)
|
||||
val isLHS = isnan2 || isMax != io.lt && !isnan1
|
||||
|
||||
val mux = new FPResult
|
||||
mux.data := fsgnj
|
||||
mux.exc := Bits(0)
|
||||
mux.exc := minmax_exc
|
||||
mux.data := in.bits.in2
|
||||
|
||||
when (in.bits.cmd === FCMD_MIN || in.bits.cmd === FCMD_MAX) {
|
||||
mux.data := minmax
|
||||
}
|
||||
when (in.bits.cmd === FCMD_CVT_FMT_S || in.bits.cmd === FCMD_CVT_FMT_D) {
|
||||
when (isSgnj) { mux.exc := UInt(0) }
|
||||
when (isSgnj || isLHS) { mux.data := fsgnj }
|
||||
when (in.bits.cmd === FCMD_CVT_FF) {
|
||||
when (in.bits.single) {
|
||||
mux.data := Cat(SInt(-1, 32), d2s._1)
|
||||
mux.exc := d2s._2
|
||||
@ -357,89 +306,39 @@ class FPToFP(val latency: Int) extends Module
|
||||
io.out <> Pipe(in.valid, mux, latency-1)
|
||||
}
|
||||
|
||||
class ioFMA(width: Int) extends Bundle {
|
||||
val valid = Bool(INPUT)
|
||||
val cmd = Bits(INPUT, FCMD_WIDTH)
|
||||
val rm = Bits(INPUT, 3)
|
||||
val in1 = Bits(INPUT, width)
|
||||
val in2 = Bits(INPUT, width)
|
||||
val in3 = Bits(INPUT, width)
|
||||
val out = Bits(OUTPUT, width)
|
||||
val exc = Bits(OUTPUT, 5)
|
||||
}
|
||||
|
||||
class FPUSFMAPipe(val latency: Int) extends Module
|
||||
class FPUFMAPipe(val latency: Int, sigWidth: Int, expWidth: Int) extends Module
|
||||
{
|
||||
val io = new ioFMA(33)
|
||||
|
||||
val cmd = Reg(Bits())
|
||||
val rm = Reg(Bits())
|
||||
val in1 = Reg(Bits())
|
||||
val in2 = Reg(Bits())
|
||||
val in3 = Reg(Bits())
|
||||
|
||||
val cmd_fma = io.cmd === FCMD_MADD || io.cmd === FCMD_MSUB ||
|
||||
io.cmd === FCMD_NMADD || io.cmd === FCMD_NMSUB
|
||||
val cmd_addsub = io.cmd === FCMD_ADD || io.cmd === FCMD_SUB
|
||||
|
||||
val one = Bits("h80000000")
|
||||
val zero = Cat(io.in1(32) ^ io.in2(32), Bits(0, 32))
|
||||
|
||||
val valid = Reg(next=io.valid)
|
||||
when (io.valid) {
|
||||
cmd := Cat(io.cmd(1) & (cmd_fma || cmd_addsub), io.cmd(0))
|
||||
rm := io.rm
|
||||
in1 := io.in1
|
||||
in2 := Mux(cmd_addsub, one, io.in2)
|
||||
in3 := Mux(cmd_fma, io.in3, Mux(cmd_addsub, io.in2, zero))
|
||||
val io = new Bundle {
|
||||
val in = Valid(new FPInput).flip
|
||||
val out = Valid(new FPResult)
|
||||
}
|
||||
|
||||
val fma = Module(new hardfloat.mulAddSubRecodedFloatN(23, 9))
|
||||
fma.io.op := cmd
|
||||
fma.io.roundingMode := rm
|
||||
fma.io.a := in1
|
||||
fma.io.b := in2
|
||||
fma.io.c := in3
|
||||
val width = sigWidth + expWidth
|
||||
val one = UInt(1) << (width-1)
|
||||
val zero = (io.in.bits.in1(width) ^ io.in.bits.in2(width)) << width
|
||||
|
||||
io.out := Pipe(valid, fma.io.out, latency-1).bits
|
||||
io.exc := Pipe(valid, fma.io.exceptionFlags, latency-1).bits
|
||||
val valid = Reg(next=io.in.valid)
|
||||
val in = Reg(new FPInput)
|
||||
when (io.in.valid) {
|
||||
in := io.in.bits
|
||||
val cmd_fma = io.in.bits.ren3
|
||||
val cmd_addsub = io.in.bits.swap23
|
||||
in.cmd := Cat(io.in.bits.cmd(1) & (cmd_fma || cmd_addsub), io.in.bits.cmd(0))
|
||||
when (cmd_addsub) { in.in2 := one }
|
||||
unless (cmd_fma || cmd_addsub) { in.in3 := zero }
|
||||
}
|
||||
|
||||
class FPUDFMAPipe(val latency: Int) extends Module
|
||||
{
|
||||
val io = new ioFMA(65)
|
||||
val fma = Module(new hardfloat.mulAddSubRecodedFloatN(sigWidth, expWidth))
|
||||
fma.io.op := in.cmd
|
||||
fma.io.roundingMode := in.rm
|
||||
fma.io.a := in.in1
|
||||
fma.io.b := in.in2
|
||||
fma.io.c := in.in3
|
||||
|
||||
val cmd = Reg(Bits())
|
||||
val rm = Reg(Bits())
|
||||
val in1 = Reg(Bits())
|
||||
val in2 = Reg(Bits())
|
||||
val in3 = Reg(Bits())
|
||||
|
||||
val cmd_fma = io.cmd === FCMD_MADD || io.cmd === FCMD_MSUB ||
|
||||
io.cmd === FCMD_NMADD || io.cmd === FCMD_NMSUB
|
||||
val cmd_addsub = io.cmd === FCMD_ADD || io.cmd === FCMD_SUB
|
||||
|
||||
val one = Bits("h8000000000000000")
|
||||
val zero = Cat(io.in1(64) ^ io.in2(64), Bits(0, 64))
|
||||
|
||||
val valid = Reg(next=io.valid)
|
||||
when (io.valid) {
|
||||
cmd := Cat(io.cmd(1) & (cmd_fma || cmd_addsub), io.cmd(0))
|
||||
rm := io.rm
|
||||
in1 := io.in1
|
||||
in2 := Mux(cmd_addsub, one, io.in2)
|
||||
in3 := Mux(cmd_fma, io.in3, Mux(cmd_addsub, io.in2, zero))
|
||||
}
|
||||
|
||||
val fma = Module(new hardfloat.mulAddSubRecodedFloatN(52, 12))
|
||||
fma.io.op := cmd
|
||||
fma.io.roundingMode := rm
|
||||
fma.io.a := in1
|
||||
fma.io.b := in2
|
||||
fma.io.c := in3
|
||||
|
||||
io.out := Pipe(valid, fma.io.out, latency-1).bits
|
||||
io.exc := Pipe(valid, fma.io.exceptionFlags, latency-1).bits
|
||||
val res = new FPResult
|
||||
res.data := fma.io.out
|
||||
res.exc := fma.io.exceptionFlags
|
||||
io.out := Pipe(valid, res, latency-1)
|
||||
}
|
||||
|
||||
class FPU(conf: FPUConfig) extends Module
|
||||
@ -447,24 +346,21 @@ class FPU(conf: FPUConfig) extends Module
|
||||
val io = new Bundle {
|
||||
val ctrl = (new CtrlFPUIO).flip
|
||||
val dpath = (new DpathFPUIO).flip
|
||||
val sfma = new ioFMA(33)
|
||||
val dfma = new ioFMA(65)
|
||||
}
|
||||
|
||||
val ex_reg_inst = Reg(Bits())
|
||||
when (io.ctrl.valid) {
|
||||
ex_reg_inst := io.dpath.inst
|
||||
}
|
||||
val ex_reg_valid = Reg(next=io.ctrl.valid, init=Bool(false))
|
||||
val ex_reg_inst = RegEnable(io.dpath.inst, io.ctrl.valid)
|
||||
val mem_reg_valid = Reg(next=ex_reg_valid && !io.ctrl.killx, init=Bool(false))
|
||||
val mem_reg_inst = RegEnable(ex_reg_inst, ex_reg_valid)
|
||||
val killm = io.ctrl.killm || io.ctrl.nack_mem
|
||||
val wb_reg_valid = Reg(next=mem_reg_valid && !killm, init=Bool(false))
|
||||
|
||||
val fp_decoder = Module(new FPUDecoder)
|
||||
fp_decoder.io.inst := io.dpath.inst
|
||||
|
||||
val ctrl = RegEnable(fp_decoder.io.sigs, io.ctrl.valid)
|
||||
val mem_ctrl = RegEnable(ctrl, ex_reg_valid)
|
||||
val id_ctrl = fp_decoder.io.sigs
|
||||
val ex_ctrl = RegEnable(id_ctrl, io.ctrl.valid)
|
||||
val mem_ctrl = RegEnable(ex_ctrl, ex_reg_valid)
|
||||
val wb_ctrl = RegEnable(mem_ctrl, mem_reg_valid)
|
||||
|
||||
// load response
|
||||
@ -480,64 +376,58 @@ class FPU(conf: FPUConfig) extends Module
|
||||
val regfile = Mem(Bits(width = 65), 32)
|
||||
when (load_wb) { regfile(load_wb_tag) := load_wb_data_recoded }
|
||||
|
||||
val ex_rs1 = regfile(ex_reg_inst(19,15))
|
||||
val ex_rs2 = regfile(ex_reg_inst(24,20))
|
||||
val ex_rs3 = regfile(ex_reg_inst(31,27))
|
||||
val ex_ra1::ex_ra2::ex_ra3::Nil = List.fill(3)(Reg(UInt()))
|
||||
when (io.ctrl.valid) {
|
||||
when (id_ctrl.ren1) { ex_ra1 := io.dpath.inst(19,15) }
|
||||
when (id_ctrl.ren3) { ex_ra3 := io.dpath.inst(31,27) }
|
||||
when (id_ctrl.ren2) {
|
||||
when ( id_ctrl.ldst) { ex_ra1 := io.dpath.inst(24,20) }
|
||||
when (!id_ctrl.ldst && !id_ctrl.swap23) { ex_ra2 := io.dpath.inst(24,20) }
|
||||
when (!id_ctrl.ldst && id_ctrl.swap23) { ex_ra3 := io.dpath.inst(24,20) }
|
||||
}
|
||||
}
|
||||
val ex_rs1::ex_rs2::ex_rs3::Nil = Seq(ex_ra1, ex_ra2, ex_ra3).map(regfile(_))
|
||||
val ex_rm = Mux(ex_reg_inst(14,12) === Bits(7), io.dpath.fcsr_rm, ex_reg_inst(14,12))
|
||||
|
||||
val fpiu = Module(new FPToInt)
|
||||
fpiu.io.in.valid := ex_reg_valid && ctrl.toint
|
||||
fpiu.io.in.bits := ctrl
|
||||
fpiu.io.in.bits.rm := ex_rm
|
||||
fpiu.io.in.bits.in1 := ex_rs1
|
||||
fpiu.io.in.bits.in2 := ex_rs2
|
||||
val req = new FPInput
|
||||
req := ex_ctrl
|
||||
req.rm := ex_rm
|
||||
req.in1 := ex_rs1
|
||||
req.in2 := ex_rs2
|
||||
req.in3 := ex_rs3
|
||||
req.typ := ex_reg_inst(21,20)
|
||||
|
||||
val sfma = Module(new FPUFMAPipe(conf.sfmaLatency, 23, 9))
|
||||
sfma.io.in.valid := ex_reg_valid && ex_ctrl.fma && ex_ctrl.single
|
||||
sfma.io.in.bits := req
|
||||
|
||||
val dfma = Module(new FPUFMAPipe(conf.dfmaLatency, 52, 12))
|
||||
dfma.io.in.valid := ex_reg_valid && ex_ctrl.fma && !ex_ctrl.single
|
||||
dfma.io.in.bits := req
|
||||
|
||||
val fpiu = Module(new FPToInt)
|
||||
fpiu.io.in.valid := ex_reg_valid && ex_ctrl.toint
|
||||
fpiu.io.in.bits := req
|
||||
io.dpath.store_data := fpiu.io.out.bits.store
|
||||
io.dpath.toint_data := fpiu.io.out.bits.toint
|
||||
|
||||
val ifpu = Module(new IntToFP(3))
|
||||
ifpu.io.in.valid := ex_reg_valid && ctrl.fromint
|
||||
ifpu.io.in.bits := ctrl
|
||||
ifpu.io.in.bits.rm := ex_rm
|
||||
ifpu.io.in.bits.data := io.dpath.fromint_data
|
||||
ifpu.io.in.valid := ex_reg_valid && ex_ctrl.fromint
|
||||
ifpu.io.in.bits := req
|
||||
ifpu.io.in.bits.in1 := io.dpath.fromint_data
|
||||
|
||||
val fpmu = Module(new FPToFP(2))
|
||||
fpmu.io.in.valid := ex_reg_valid && ctrl.fastpipe
|
||||
fpmu.io.in.bits := ctrl
|
||||
fpmu.io.in.bits.rm := ex_rm
|
||||
fpmu.io.in.bits.in1 := ex_rs1
|
||||
fpmu.io.in.bits.in2 := ex_rs2
|
||||
fpmu.io.in.valid := ex_reg_valid && ex_ctrl.fastpipe
|
||||
fpmu.io.in.bits := req
|
||||
fpmu.io.lt := fpiu.io.out.bits.lt
|
||||
|
||||
val cmd_fma = mem_ctrl.cmd === FCMD_MADD || mem_ctrl.cmd === FCMD_MSUB ||
|
||||
mem_ctrl.cmd === FCMD_NMADD || mem_ctrl.cmd === FCMD_NMSUB
|
||||
val cmd_addsub = mem_ctrl.cmd === FCMD_ADD || mem_ctrl.cmd === FCMD_SUB
|
||||
val sfma = Module(new FPUSFMAPipe(conf.sfmaLatency))
|
||||
sfma.io.valid := io.sfma.valid || ex_reg_valid && ctrl.fma && ctrl.single
|
||||
sfma.io.in1 := Mux(io.sfma.valid, io.sfma.in1, ex_rs1)
|
||||
sfma.io.in2 := Mux(io.sfma.valid, io.sfma.in2, ex_rs2)
|
||||
sfma.io.in3 := Mux(io.sfma.valid, io.sfma.in3, ex_rs3)
|
||||
sfma.io.cmd := Mux(io.sfma.valid, io.sfma.cmd, ctrl.cmd)
|
||||
sfma.io.rm := Mux(io.sfma.valid, io.sfma.rm, ex_rm)
|
||||
io.sfma.out := sfma.io.out
|
||||
io.sfma.exc := sfma.io.exc
|
||||
|
||||
val dfma = Module(new FPUDFMAPipe(conf.dfmaLatency))
|
||||
dfma.io.valid := io.dfma.valid || ex_reg_valid && ctrl.fma && !ctrl.single
|
||||
dfma.io.in1 := Mux(io.dfma.valid, io.dfma.in1, ex_rs1)
|
||||
dfma.io.in2 := Mux(io.dfma.valid, io.dfma.in2, ex_rs2)
|
||||
dfma.io.in3 := Mux(io.dfma.valid, io.dfma.in3, ex_rs3)
|
||||
dfma.io.cmd := Mux(io.dfma.valid, io.dfma.cmd, ctrl.cmd)
|
||||
dfma.io.rm := Mux(io.dfma.valid, io.dfma.rm, ex_rm)
|
||||
io.dfma.out := dfma.io.out
|
||||
io.dfma.exc := dfma.io.exc
|
||||
|
||||
// writeback arbitration
|
||||
case class Pipe(p: Module, lat: Int, cond: (FPUCtrlSigs) => Bool, wdata: Bits, wexc: Bits)
|
||||
val pipes = List(
|
||||
Pipe(fpmu, fpmu.latency, (c: FPUCtrlSigs) => c.fastpipe, fpmu.io.out.bits.data, fpmu.io.out.bits.exc),
|
||||
Pipe(ifpu, ifpu.latency, (c: FPUCtrlSigs) => c.fromint, ifpu.io.out.bits.data, ifpu.io.out.bits.exc),
|
||||
Pipe(sfma, sfma.latency, (c: FPUCtrlSigs) => c.fma && c.single, Cat(SInt(-1, 32), sfma.io.out), sfma.io.exc),
|
||||
Pipe(dfma, dfma.latency, (c: FPUCtrlSigs) => c.fma && !c.single, dfma.io.out, dfma.io.exc))
|
||||
Pipe(sfma, sfma.latency, (c: FPUCtrlSigs) => c.fma && c.single, Cat(SInt(-1, 32), sfma.io.out.bits.data), sfma.io.out.bits.exc),
|
||||
Pipe(dfma, dfma.latency, (c: FPUCtrlSigs) => c.fma && !c.single, dfma.io.out.bits.data, dfma.io.out.bits.exc))
|
||||
def latencyMask(c: FPUCtrlSigs, offset: Int) = {
|
||||
require(pipes.forall(_.lat >= offset))
|
||||
pipes.map(p => Mux(p.cond(c), UInt(1 << p.lat-offset), UInt(0))).reduce(_|_)
|
||||
@ -549,11 +439,8 @@ class FPU(conf: FPUConfig) extends Module
|
||||
val wen = Reg(init=Bits(0, maxLatency-1))
|
||||
val winfo = Vec.fill(maxLatency-1){Reg(Bits())}
|
||||
val mem_wen = mem_reg_valid && (mem_ctrl.fma || mem_ctrl.fastpipe || mem_ctrl.fromint)
|
||||
val (write_port_busy, mem_winfo) = (Reg(Bool()), Reg(Bits()))
|
||||
when (ex_reg_valid) {
|
||||
write_port_busy := mem_wen && (memLatencyMask & latencyMask(ctrl, 1)).orR || (wen & latencyMask(ctrl, 0)).orR
|
||||
mem_winfo := Cat(pipeid(ctrl), ex_reg_inst(11,7))
|
||||
}
|
||||
val write_port_busy = RegEnable(mem_wen && (memLatencyMask & latencyMask(ex_ctrl, 1)).orR || (wen & latencyMask(ex_ctrl, 0)).orR, ex_reg_valid)
|
||||
val mem_winfo = Cat(pipeid(mem_ctrl), mem_reg_inst(11,7))
|
||||
|
||||
for (i <- 0 until maxLatency-2) {
|
||||
when (wen(i+1)) { winfo(i) := winfo(i+1) }
|
||||
@ -584,7 +471,7 @@ class FPU(conf: FPUConfig) extends Module
|
||||
Mux(wen(0), wexc, UInt(0))
|
||||
|
||||
val fp_inflight = wb_reg_valid && wb_ctrl.toint || wen.orR
|
||||
val units_busy = mem_reg_valid && mem_ctrl.fma && Reg(next=Mux(ctrl.single, io.sfma.valid, io.dfma.valid))
|
||||
val units_busy = Bool(false) //mem_reg_valid && mem_ctrl.fma && Reg(next=Mux(ex_ctrl.single, io.sfma.valid, io.dfma.valid))
|
||||
io.ctrl.fcsr_rdy := !fp_inflight
|
||||
io.ctrl.nack_mem := units_busy || write_port_busy
|
||||
io.ctrl.dec <> fp_decoder.io.sigs
|
||||
@ -593,5 +480,5 @@ class FPU(conf: FPUConfig) extends Module
|
||||
io.ctrl.sboard_clr := wen(0) && useScoreboard(x => wsrc === UInt(x._2))
|
||||
io.ctrl.sboard_clra := waddr
|
||||
// we don't currently support round-max-magnitude (rm=4)
|
||||
io.ctrl.illegal_rm := ex_rm(2) && ctrl.round
|
||||
io.ctrl.illegal_rm := ex_rm(2) && ex_ctrl.round
|
||||
}
|
||||
|
||||
@ -104,51 +104,51 @@ object Instructions {
|
||||
def FSUB_S = Bits("b0000100??????????????????1010011")
|
||||
def FMUL_S = Bits("b0001000??????????????????1010011")
|
||||
def FDIV_S = Bits("b0001100??????????????????1010011")
|
||||
def FSQRT_S = Bits("b001000000000?????????????1010011")
|
||||
def FSGNJ_S = Bits("b0010100??????????000?????1010011")
|
||||
def FSGNJN_S = Bits("b0011000??????????000?????1010011")
|
||||
def FSGNJX_S = Bits("b0011100??????????000?????1010011")
|
||||
def FSGNJ_S = Bits("b0010000??????????000?????1010011")
|
||||
def FSGNJN_S = Bits("b0010000??????????001?????1010011")
|
||||
def FSGNJX_S = Bits("b0010000??????????010?????1010011")
|
||||
def FMIN_S = Bits("b0010100??????????000?????1010011")
|
||||
def FMAX_S = Bits("b0010100??????????001?????1010011")
|
||||
def FSQRT_S = Bits("b010110000000?????????????1010011")
|
||||
def FADD_D = Bits("b0000001??????????????????1010011")
|
||||
def FSUB_D = Bits("b0000101??????????????????1010011")
|
||||
def FMUL_D = Bits("b0001001??????????????????1010011")
|
||||
def FDIV_D = Bits("b0001101??????????????????1010011")
|
||||
def FSQRT_D = Bits("b001000100000?????????????1010011")
|
||||
def FSGNJ_D = Bits("b0010101??????????000?????1010011")
|
||||
def FSGNJN_D = Bits("b0011001??????????000?????1010011")
|
||||
def FSGNJX_D = Bits("b0011101??????????000?????1010011")
|
||||
def FCVT_L_S = Bits("b010000000000?????????????1010011")
|
||||
def FCVT_LU_S = Bits("b010010000000?????????????1010011")
|
||||
def FCVT_W_S = Bits("b010100000000?????????????1010011")
|
||||
def FCVT_WU_S = Bits("b010110000000?????????????1010011")
|
||||
def FCVT_L_D = Bits("b010000100000?????????????1010011")
|
||||
def FCVT_LU_D = Bits("b010010100000?????????????1010011")
|
||||
def FCVT_W_D = Bits("b010100100000?????????????1010011")
|
||||
def FCVT_WU_D = Bits("b010110100000?????????????1010011")
|
||||
def FCVT_S_L = Bits("b011000000000?????????????1010011")
|
||||
def FCVT_S_LU = Bits("b011010000000?????????????1010011")
|
||||
def FCVT_S_W = Bits("b011100000000?????????????1010011")
|
||||
def FCVT_S_WU = Bits("b011110000000?????????????1010011")
|
||||
def FCVT_D_L = Bits("b011000100000?????????????1010011")
|
||||
def FCVT_D_LU = Bits("b011010100000?????????????1010011")
|
||||
def FCVT_D_W = Bits("b011100100000?????????????1010011")
|
||||
def FCVT_D_WU = Bits("b011110100000?????????????1010011")
|
||||
def FCVT_S_D = Bits("b100010000000?????????????1010011")
|
||||
def FCVT_D_S = Bits("b100000100000?????????????1010011")
|
||||
def FEQ_S = Bits("b1010100??????????000?????1010011")
|
||||
def FLT_S = Bits("b1011000??????????000?????1010011")
|
||||
def FLE_S = Bits("b1011100??????????000?????1010011")
|
||||
def FEQ_D = Bits("b1010101??????????000?????1010011")
|
||||
def FLT_D = Bits("b1011001??????????000?????1010011")
|
||||
def FLE_D = Bits("b1011101??????????000?????1010011")
|
||||
def FMIN_S = Bits("b1100000??????????000?????1010011")
|
||||
def FMAX_S = Bits("b1100100??????????000?????1010011")
|
||||
def FMIN_D = Bits("b1100001??????????000?????1010011")
|
||||
def FMAX_D = Bits("b1100101??????????000?????1010011")
|
||||
def FSGNJ_D = Bits("b0010001??????????000?????1010011")
|
||||
def FSGNJN_D = Bits("b0010001??????????001?????1010011")
|
||||
def FSGNJX_D = Bits("b0010001??????????010?????1010011")
|
||||
def FMIN_D = Bits("b0010101??????????000?????1010011")
|
||||
def FMAX_D = Bits("b0010101??????????001?????1010011")
|
||||
def FCVT_S_D = Bits("b010000000001?????????????1010011")
|
||||
def FCVT_D_S = Bits("b010000100000?????????????1010011")
|
||||
def FSQRT_D = Bits("b010110100000?????????????1010011")
|
||||
def FLE_S = Bits("b1010000??????????000?????1010011")
|
||||
def FLT_S = Bits("b1010000??????????001?????1010011")
|
||||
def FEQ_S = Bits("b1010000??????????010?????1010011")
|
||||
def FLE_D = Bits("b1010001??????????000?????1010011")
|
||||
def FLT_D = Bits("b1010001??????????001?????1010011")
|
||||
def FEQ_D = Bits("b1010001??????????010?????1010011")
|
||||
def FCVT_W_S = Bits("b110000000000?????????????1010011")
|
||||
def FCVT_WU_S = Bits("b110000000001?????????????1010011")
|
||||
def FCVT_L_S = Bits("b110000000010?????????????1010011")
|
||||
def FCVT_LU_S = Bits("b110000000011?????????????1010011")
|
||||
def FMV_X_S = Bits("b111000000000?????000?????1010011")
|
||||
def FCLASS_S = Bits("b111000000000?????001?????1010011")
|
||||
def FCVT_W_D = Bits("b110000100000?????????????1010011")
|
||||
def FCVT_WU_D = Bits("b110000100001?????????????1010011")
|
||||
def FCVT_L_D = Bits("b110000100010?????????????1010011")
|
||||
def FCVT_LU_D = Bits("b110000100011?????????????1010011")
|
||||
def FMV_X_D = Bits("b111000100000?????000?????1010011")
|
||||
def FCLASS_S = Bits("b111010000000?????000?????1010011")
|
||||
def FCLASS_D = Bits("b111010100000?????000?????1010011")
|
||||
def FCLASS_D = Bits("b111000100000?????001?????1010011")
|
||||
def FCVT_S_W = Bits("b110100000000?????????????1010011")
|
||||
def FCVT_S_WU = Bits("b110100000001?????????????1010011")
|
||||
def FCVT_S_L = Bits("b110100000010?????????????1010011")
|
||||
def FCVT_S_LU = Bits("b110100000011?????????????1010011")
|
||||
def FMV_S_X = Bits("b111100000000?????000?????1010011")
|
||||
def FCVT_D_W = Bits("b110100100000?????????????1010011")
|
||||
def FCVT_D_WU = Bits("b110100100001?????????????1010011")
|
||||
def FCVT_D_L = Bits("b110100100010?????????????1010011")
|
||||
def FCVT_D_LU = Bits("b110100100011?????????????1010011")
|
||||
def FMV_D_X = Bits("b111100100000?????000?????1010011")
|
||||
def FLW = Bits("b?????????????????010?????0000111")
|
||||
def FLD = Bits("b?????????????????011?????0000111")
|
||||
|
||||
Loading…
Reference in New Issue
Block a user