xref: /XiangShan/src/main/scala/xiangshan/backend/fu/wrapper/VFALU.scala (revision b1e920234888fd3e5463ceb2a99c9bdca087f585)
1package xiangshan.backend.fu.wrapper
2
3import org.chipsalliance.cde.config.Parameters
4import chisel3._
5import chisel3.util._
6import utils.XSError
7import xiangshan.backend.fu.FuConfig
8import xiangshan.backend.fu.vector.Bundles.{VLmul, VSew, ma}
9import xiangshan.backend.fu.vector.utils.VecDataSplitModule
10import xiangshan.backend.fu.vector.{Mgu, VecInfo, VecPipedFuncUnit}
11import xiangshan.ExceptionNO
12import yunsuan.{VfaluType, VfpuType}
13import yunsuan.vector.VectorFloatAdder
14
15class VFAlu(cfg: FuConfig)(implicit p: Parameters) extends VecPipedFuncUnit(cfg) {
16  XSError(io.in.valid && io.in.bits.ctrl.fuOpType === VfpuType.dummy, "Vfalu OpType not supported")
17
18  // params alias
19  private val dataWidth = cfg.dataBits
20  private val dataWidthOfDataModule = 64
21  private val numVecModule = dataWidth / dataWidthOfDataModule
22
23  // io alias
24  private val opcode  = fuOpType(4,0)
25  private val resWiden  = fuOpType(5)
26  private val opbWiden  = fuOpType(6)
27
28  // modules
29  private val vfalus = Seq.fill(numVecModule)(Module(new VectorFloatAdder))
30  private val vs2Split = Module(new VecDataSplitModule(dataWidth, dataWidthOfDataModule))
31  private val vs1Split = Module(new VecDataSplitModule(dataWidth, dataWidthOfDataModule))
32  private val oldVdSplit  = Module(new VecDataSplitModule(dataWidth, dataWidthOfDataModule))
33  private val mgu = Module(new Mgu(dataWidth))
34
35  /**
36    * In connection of [[vs2Split]], [[vs1Split]] and [[oldVdSplit]]
37    */
38  vs2Split.io.inVecData := vs2
39  vs1Split.io.inVecData := vs1
40  oldVdSplit.io.inVecData := oldVd
41
42  /**
43    * [[vfalus]]'s in connection
44    */
45  // Vec(vs2(31,0), vs2(63,32), vs2(95,64), vs2(127,96)) ==>
46  // Vec(
47  //   Cat(vs2(95,64),  vs2(31,0)),
48  //   Cat(vs2(127,96), vs2(63,32)),
49  // )
50  private val vs2GroupedVec: Vec[UInt] = VecInit(vs2Split.io.outVec32b.zipWithIndex.groupBy(_._2 % 2).map(x => x._1 -> x._2.map(_._1)).values.map(x => Cat(x.reverse)).toSeq)
51  private val vs1GroupedVec: Vec[UInt] = VecInit(vs1Split.io.outVec32b.zipWithIndex.groupBy(_._2 % 2).map(x => x._1 -> x._2.map(_._1)).values.map(x => Cat(x.reverse)).toSeq)
52  private val resultData = Wire(Vec(numVecModule,UInt(dataWidthOfDataModule.W)))
53  private val fflagsData = Wire(Vec(numVecModule,UInt(20.W)))
54  private val srcMaskRShiftForReduction = Wire(UInt((8 * numVecModule).W))
55  // for reduction
56  val isFirstGroupUop = vuopIdx === 0.U ||
57    (vuopIdx === 1.U && (vlmul === VLmul.m4 || vlmul === VLmul.m8)) ||
58    ((vuopIdx === 2.U || vuopIdx === 3.U) && vlmul === VLmul.m8)
59  val maskRshiftWidthForReduction = Wire(UInt(6.W))
60  maskRshiftWidthForReduction := Mux(fuOpType === VfaluType.vfredosum,
61    vuopIdx,
62    Mux1H(Seq(
63      (vsew === VSew.e16) -> (vuopIdx(1, 0) << 4),
64      (vsew === VSew.e32) -> (vuopIdx(1, 0) << 3),
65      (vsew === VSew.e64) -> (vuopIdx(1, 0) << 2),
66    ))
67  )
68  val vlMaskForReduction = (~(Fill(VLEN, 1.U) << vl)).asUInt
69  srcMaskRShiftForReduction := ((srcMask & vlMaskForReduction) >> maskRshiftWidthForReduction)(8 * numVecModule - 1, 0)
70
71  def genMaskForReduction(inmask: UInt, sew: UInt, i: Int): UInt = {
72    val f64MaskNum = dataWidth / 64 * 2
73    val f32MaskNum = dataWidth / 32 * 2
74    val f16MaskNum = dataWidth / 16 * 2
75    val f64Mask = inmask(f64MaskNum - 1, 0)
76    val f32Mask = inmask(f32MaskNum - 1, 0)
77    val f16Mask = inmask(f16MaskNum - 1, 0)
78    // vs2 reordered, so mask use high bits
79    val f64FirstFoldMask = Mux1H(
80      Seq(
81        vecCtrl.fpu.isFoldTo1_2 -> Cat(0.U(3.W), f64Mask(0), 0.U(3.W), f64Mask(1)),
82      )
83    )
84    val f32FirstFoldMask = Mux1H(
85      Seq(
86        vecCtrl.fpu.isFoldTo1_2 -> Cat(0.U(2.W), f32Mask(1), f32Mask(0), 0.U(2.W), f32Mask(3), f32Mask(2)),
87        vecCtrl.fpu.isFoldTo1_4 -> Cat(0.U(3.W), f32Mask(0), 0.U(3.W), f32Mask(1)),
88      )
89    )
90    val f16FirstFoldMask = Mux1H(
91      Seq(
92        vecCtrl.fpu.isFoldTo1_2 -> Cat(f16Mask(7,4), f16Mask(3,0)),
93        vecCtrl.fpu.isFoldTo1_4 -> Cat(0.U(2.W), f16Mask(1), f16Mask(0), 0.U(2.W), f16Mask(3), f16Mask(2)),
94        vecCtrl.fpu.isFoldTo1_8 -> Cat(0.U(3.W), f16Mask(0), 0.U(3.W), f16Mask(1)),
95      )
96    )
97    val f64FoldMask = Mux1H(
98      Seq(
99        vecCtrl.fpu.isFoldTo1_2 -> "b00010001".U,
100      )
101    )
102    val f32FoldMask = Mux1H(
103      Seq(
104        vecCtrl.fpu.isFoldTo1_2 -> "b00110011".U,
105        vecCtrl.fpu.isFoldTo1_4 -> "b00010001".U,
106      )
107    )
108    val f16FoldMask = Mux1H(
109      Seq(
110        vecCtrl.fpu.isFoldTo1_2 -> "b11111111".U,
111        vecCtrl.fpu.isFoldTo1_4 -> "b00110011".U,
112        vecCtrl.fpu.isFoldTo1_8 -> "b00010001".U,
113      )
114    )
115    // low 4 bits for vs2(fp_a), high 4 bits for vs1(fp_b),
116    val isFold = vecCtrl.fpu.isFoldTo1_2 || vecCtrl.fpu.isFoldTo1_4 || vecCtrl.fpu.isFoldTo1_8
117    val f64FirstNotFoldMask = Cat(0.U(3.W), f64Mask(i+2), 0.U(3.W), f64Mask(i))
118    val f32FirstNotFoldMask = Cat(0.U(2.W), f32Mask(i + 5, i+4), 0.U(2.W), Cat(f32Mask(i + 1, i)))
119    val f16FirstNotFoldMask = Cat(f16Mask(i+11,i+8), f16Mask(i+3,0))
120    val f64MaskI = Mux(isFirstGroupUop, Mux(isFold, f64FirstFoldMask, f64FirstNotFoldMask), Mux(isFold, f64FoldMask, Fill(8,1.U)))
121    val f32MaskI = Mux(isFirstGroupUop, Mux(isFold, f32FirstFoldMask, f32FirstNotFoldMask), Mux(isFold, f32FoldMask, Fill(8,1.U)))
122    val f16MaskI = Mux(isFirstGroupUop, Mux(isFold, f16FirstFoldMask, f16FirstNotFoldMask), Mux(isFold, f16FoldMask, Fill(8,1.U)))
123    val outMask = Mux1H(
124      Seq(
125        (sew === 3.U) -> f64MaskI,
126        (sew === 2.U) -> f32MaskI,
127        (sew === 1.U) -> f16MaskI,
128      )
129    )
130    Mux(fuOpType === VfaluType.vfredosum || fuOpType === VfaluType.vfwredosum, outMask(0),outMask)
131  }
132  def genMaskForMerge(inmask:UInt, sew:UInt, i:Int): UInt = {
133    val f64MaskNum = dataWidth / 64
134    val f32MaskNum = dataWidth / 32
135    val f16MaskNum = dataWidth / 16
136    val f64Mask = inmask(f64MaskNum-1,0)
137    val f32Mask = inmask(f32MaskNum-1,0)
138    val f16Mask = inmask(f16MaskNum-1,0)
139    val f64MaskI = Cat(0.U(3.W),f64Mask(i))
140    val f32MaskI = Cat(0.U(2.W),f32Mask(2*i+1,2*i))
141    val f16MaskI = f16Mask(4*i+3,4*i)
142    val outMask = Mux1H(
143      Seq(
144        (sew === 3.U) -> f64MaskI,
145        (sew === 2.U) -> f32MaskI,
146        (sew === 1.U) -> f16MaskI,
147      )
148    )
149    outMask
150  }
151  val isScalarMove = (fuOpType === VfaluType.vfmv_f_s) || (fuOpType === VfaluType.vfmv_s_f)
152  val srcMaskRShift = Wire(UInt((4 * numVecModule).W))
153  val maskRshiftWidth = Wire(UInt(6.W))
154  maskRshiftWidth := Mux1H(
155    Seq(
156      (vsew === VSew.e16) -> (vuopIdx(2,0) << 3),
157      (vsew === VSew.e32) -> (vuopIdx(2,0) << 2),
158      (vsew === VSew.e64) -> (vuopIdx(2,0) << 1),
159    )
160  )
161  srcMaskRShift := (srcMask >> maskRshiftWidth)(4 * numVecModule - 1, 0)
162  val fp_aIsFpCanonicalNAN = Wire(Vec(numVecModule,Bool()))
163  val fp_bIsFpCanonicalNAN = Wire(Vec(numVecModule,Bool()))
164  vfalus.zipWithIndex.foreach {
165    case (mod, i) =>
166      mod.io.fp_a             := Mux(opbWiden, vs1Split.io.outVec64b(i), vs2Split.io.outVec64b(i))  // very dirty TODO
167      mod.io.fp_b             := Mux(opbWiden, vs2Split.io.outVec64b(i), vs1Split.io.outVec64b(i))  // very dirty TODO
168      mod.io.widen_a          := Cat(vs2Split.io.outVec32b(i+numVecModule), vs2Split.io.outVec32b(i))
169      mod.io.widen_b          := Cat(vs1Split.io.outVec32b(i+numVecModule), vs1Split.io.outVec32b(i))
170      mod.io.frs1             := 0.U     // already vf -> vv
171      mod.io.is_frs1          := false.B // already vf -> vv
172      mod.io.mask             := Mux(isScalarMove, !vuopIdx.orR, genMaskForMerge(inmask = srcMaskRShift, sew = vsew, i = i))
173      mod.io.maskForReduction := genMaskForReduction(inmask = srcMaskRShiftForReduction, sew = vsew, i = i)
174      mod.io.uop_idx          := Mux(fuOpType === VfaluType.vfwredosum, 0.U, vuopIdx(0))
175      mod.io.is_vec           := true.B // Todo
176      mod.io.round_mode       := frm
177      mod.io.fp_format        := Mux(resWiden, vsew + 1.U, vsew)
178      mod.io.opb_widening     := opbWiden || (fuOpType === VfaluType.vfwredosum)
179      mod.io.res_widening     := resWiden
180      mod.io.op_code          := opcode
181      resultData(i)           := mod.io.fp_result
182      fflagsData(i)           := mod.io.fflags
183      fp_aIsFpCanonicalNAN(i) := vecCtrl.fpu.isFpToVecInst & (
184          ((vsew === VSew.e32) & (!vs2Split.io.outVec64b(i).head(32).andR)) |
185          ((vsew === VSew.e16) & (!vs2Split.io.outVec64b(i).head(48).andR))
186        )
187      fp_bIsFpCanonicalNAN(i) := vecCtrl.fpu.isFpToVecInst & (
188          ((vsew === VSew.e32) & (!vs1Split.io.outVec64b(i).head(32).andR)) |
189          ((vsew === VSew.e16) & (!vs1Split.io.outVec64b(i).head(48).andR))
190        )
191      mod.io.fp_aIsFpCanonicalNAN := fp_aIsFpCanonicalNAN(i)
192      mod.io.fp_bIsFpCanonicalNAN := fp_bIsFpCanonicalNAN(i)
193  }
194  val resultDataUInt = resultData.asUInt
195  val cmpResultWidth = dataWidth / 16
196  val cmpResult = Wire(Vec(cmpResultWidth, Bool()))
197  for (i <- 0 until cmpResultWidth) {
198    if(i == 0) {
199      cmpResult(i) := resultDataUInt(0)
200    }
201    else if(i < dataWidth / 64) {
202      cmpResult(i) := Mux1H(
203        Seq(
204          (outVecCtrl.vsew === 1.U) -> resultDataUInt(i*16),
205          (outVecCtrl.vsew === 2.U) -> resultDataUInt(i*32),
206          (outVecCtrl.vsew === 3.U) -> resultDataUInt(i*64)
207        )
208      )
209    }
210    else if(i < dataWidth / 32) {
211      cmpResult(i) := Mux1H(
212        Seq(
213          (outVecCtrl.vsew === 1.U) -> resultDataUInt(i * 16),
214          (outVecCtrl.vsew === 2.U) -> resultDataUInt(i * 32),
215          (outVecCtrl.vsew === 3.U) -> false.B
216        )
217      )
218    }
219    else if(i <  dataWidth / 16) {
220      cmpResult(i) := Mux(outVecCtrl.vsew === 1.U, resultDataUInt(i*16), false.B)
221    }
222  }
223
224  val outEew = Mux(RegNext(resWiden), outVecCtrl.vsew + 1.U, outVecCtrl.vsew)
225  val outVuopidx = outVecCtrl.vuopIdx(2, 0)
226  val vlMax = ((VLEN/8).U >> outEew).asUInt
227  val lmulAbs = Mux(outVecCtrl.vlmul(2), (~outVecCtrl.vlmul(1,0)).asUInt + 1.U, outVecCtrl.vlmul(1,0))
228  //  vfmv_f_s need vl=1, reduction last uop need vl=1, other uop need vl=vlmax
229  val numOfUopVFRED = {
230    // addTime include add frs1
231    val addTime = MuxLookup(outVecCtrl.vlmul, 1.U(4.W), Array(
232      VLmul.m2 -> 2.U,
233      VLmul.m4 -> 4.U,
234      VLmul.m8 -> 8.U,
235    ))
236    val foldLastVlmul = MuxLookup(outVecCtrl.vsew, "b000".U, Array(
237      VSew.e16 -> VLmul.mf8,
238      VSew.e32 -> VLmul.mf4,
239      VSew.e64 -> VLmul.mf2,
240    ))
241    // lmul < 1, foldTime = vlmul - foldFastVlmul
242    // lmul >= 1, foldTime = 0.U - foldFastVlmul
243    val foldTime = Mux(outVecCtrl.vlmul(2), outVecCtrl.vlmul, 0.U) - foldLastVlmul
244    addTime + foldTime
245  }
246  val reductionVl = Mux((outVecCtrl.vuopIdx ===  numOfUopVFRED - 1.U) || (outCtrl.fuOpType === VfaluType.vfredosum || outCtrl.fuOpType === VfaluType.vfwredosum), 1.U, vlMax)
247  val outIsResuction = outCtrl.fuOpType === VfaluType.vfredusum ||
248    outCtrl.fuOpType === VfaluType.vfredmax ||
249    outCtrl.fuOpType === VfaluType.vfredmin ||
250    outCtrl.fuOpType === VfaluType.vfredosum ||
251    outCtrl.fuOpType === VfaluType.vfwredosum
252  val outVlFix = Mux(
253    outVecCtrl.fpu.isFpToVecInst || (outCtrl.fuOpType === VfaluType.vfmv_f_s),
254    1.U,
255    Mux(
256      outCtrl.fuOpType === VfaluType.vfmv_s_f,
257      outVl.orR,
258      Mux(outIsResuction, reductionVl, outVl)
259    )
260  )
261  val vlMaxAllUop = Wire(outVl.cloneType)
262  vlMaxAllUop := Mux(outVecCtrl.vlmul(2), vlMax >> lmulAbs, vlMax << lmulAbs).asUInt
263  val vlMaxThisUop = Mux(outVecCtrl.vlmul(2), vlMax >> lmulAbs, vlMax).asUInt
264  val vlSetThisUop = Mux(outVlFix > outVuopidx*vlMaxThisUop, outVlFix - outVuopidx*vlMaxThisUop, 0.U)
265  val vlThisUop = Wire(UInt(3.W))
266  vlThisUop := Mux(vlSetThisUop < vlMaxThisUop, vlSetThisUop, vlMaxThisUop)
267  val vlMaskRShift = Wire(UInt((4 * numVecModule).W))
268  vlMaskRShift := Fill(4 * numVecModule, 1.U(1.W)) >> ((4 * numVecModule).U - vlThisUop)
269
270  private val needNoMask = outCtrl.fuOpType === VfaluType.vfmerge ||
271    outCtrl.fuOpType === VfaluType.vfmv_s_f ||
272    outIsResuction ||
273    outVecCtrl.fpu.isFpToVecInst
274  val maskToMgu = Mux(needNoMask, allMaskTrue, outSrcMask)
275  val allFFlagsEn = Wire(Vec(4*numVecModule,Bool()))
276  val outSrcMaskRShift = Wire(UInt((4*numVecModule).W))
277  outSrcMaskRShift := (maskToMgu >> (outVecCtrl.vuopIdx(2,0) * vlMax))(4*numVecModule-1,0)
278  val f16FFlagsEn = outSrcMaskRShift
279  val f32FFlagsEn = Wire(Vec(numVecModule,UInt(4.W)))
280  for (i <- 0 until numVecModule){
281    f32FFlagsEn(i) := Cat(Fill(2, 0.U),outSrcMaskRShift(2*i+1,2*i))
282  }
283  val f64FFlagsEn = Wire(Vec(numVecModule, UInt(4.W)))
284  for (i <- 0 until numVecModule) {
285    f64FFlagsEn(i) := Cat(Fill(3, 0.U), outSrcMaskRShift(i))
286  }
287  val fflagsEn= Mux1H(
288    Seq(
289      (outEew === 1.U) -> f16FFlagsEn.asUInt,
290      (outEew === 2.U) -> f32FFlagsEn.asUInt,
291      (outEew === 3.U) -> f64FFlagsEn.asUInt
292    )
293  )
294  allFFlagsEn := Mux(outIsResuction, Fill(4*numVecModule, 1.U), (fflagsEn & vlMaskRShift)).asTypeOf(allFFlagsEn)
295
296  val allFFlags = fflagsData.asTypeOf(Vec(4*numVecModule,UInt(5.W)))
297  val outFFlags = allFFlagsEn.zip(allFFlags).map{
298    case(en,fflags) => Mux(en, fflags, 0.U(5.W))
299  }.reduce(_ | _)
300  io.out.bits.res.fflags.get := outFFlags
301
302
303  val cmpResultOldVd = Wire(UInt(cmpResultWidth.W))
304  val cmpResultOldVdRshiftWidth = Wire(UInt(6.W))
305  cmpResultOldVdRshiftWidth := Mux1H(
306    Seq(
307      (outVecCtrl.vsew === VSew.e16) -> (outVecCtrl.vuopIdx(2, 0) << 3),
308      (outVecCtrl.vsew === VSew.e32) -> (outVecCtrl.vuopIdx(2, 0) << 2),
309      (outVecCtrl.vsew === VSew.e64) -> (outVecCtrl.vuopIdx(2, 0) << 1),
310    )
311  )
312  cmpResultOldVd := (outOldVd >> cmpResultOldVdRshiftWidth)(4*numVecModule-1,0)
313  val cmpResultForMgu = Wire(Vec(cmpResultWidth, Bool()))
314  for (i <- 0 until cmpResultWidth) {
315    cmpResultForMgu(i) := Mux(outSrcMaskRShift(i), cmpResult(i), Mux(outVecCtrl.vma, true.B, cmpResultOldVd(i)))
316  }
317  val outIsFold = outVecCtrl.fpu.isFoldTo1_2 || outVecCtrl.fpu.isFoldTo1_4 || outVecCtrl.fpu.isFoldTo1_8
318  val outOldVdForREDO = Mux1H(Seq(
319    (outVecCtrl.vsew === VSew.e16) -> (outOldVd >> 16),
320    (outVecCtrl.vsew === VSew.e32) -> (outOldVd >> 32),
321    (outVecCtrl.vsew === VSew.e64) -> (outOldVd >> 64),
322  ))
323  val outOldVdForWREDO = Mux(
324    !outIsFold,
325    Mux(outVecCtrl.vsew === VSew.e16, Cat(outOldVd(VLEN-1-16,16), 0.U(32.W)), Cat(outOldVd(VLEN-1-32,32), 0.U(64.W))),
326    Mux(outVecCtrl.vsew === VSew.e16,
327      // Divide vuopIdx by 8 and the remainder is 1
328      Mux(outVecCtrl.vuopIdx(2,0) === 1.U, outOldVd, outOldVd >> 16),
329      // Divide vuopIdx by 4 and the remainder is 1
330      Mux(outVecCtrl.vuopIdx(1,0) === 1.U, outOldVd, outOldVd >> 32)
331    ),
332  )
333  val outOldVdForRED = Mux(outCtrl.fuOpType === VfaluType.vfredosum, outOldVdForREDO, outOldVdForWREDO)
334  val numOfUopVFREDOSUM = {
335    val uvlMax = MuxLookup(outVecCtrl.vsew, 0.U, Array(
336      VSew.e16 -> 8.U,
337      VSew.e32 -> 4.U,
338      VSew.e64 -> 2.U,
339    ))
340    val vlMax = Mux(outVecCtrl.vlmul(2), uvlMax >> (-outVecCtrl.vlmul)(1, 0), uvlMax << outVecCtrl.vlmul(1, 0)).asUInt
341    vlMax
342  }
343  val isOutOldVdForREDO = (outCtrl.fuOpType === VfaluType.vfredosum && outIsFold) || outCtrl.fuOpType === VfaluType.vfwredosum
344  val taIsFalseForVFREDO = ((outCtrl.fuOpType === VfaluType.vfredosum) || (outCtrl.fuOpType === VfaluType.vfwredosum)) && (outVecCtrl.vuopIdx =/= numOfUopVFREDOSUM - 1.U)
345  mgu.io.in.vd := Mux(outVecCtrl.isDstMask, Cat(0.U((dataWidth / 16 * 15).W), cmpResultForMgu.asUInt), resultDataUInt)
346  mgu.io.in.oldVd := Mux(isOutOldVdForREDO, outOldVdForRED, outOldVd)
347  mgu.io.in.mask := maskToMgu
348  mgu.io.in.info.ta := Mux(outCtrl.fuOpType === VfaluType.vfmv_f_s, true.B , Mux(taIsFalseForVFREDO, false.B, outVecCtrl.vta))
349  mgu.io.in.info.ma := Mux(outCtrl.fuOpType === VfaluType.vfmv_s_f, true.B , outVecCtrl.vma)
350  mgu.io.in.info.vl := outVlFix
351  mgu.io.in.info.vstart := outVecCtrl.vstart
352  mgu.io.in.info.vlmul := outVecCtrl.vlmul
353  mgu.io.in.info.valid := io.out.valid
354  mgu.io.in.info.vstart := Mux(outVecCtrl.fpu.isFpToVecInst, 0.U, outVecCtrl.vstart)
355  mgu.io.in.info.eew := outEew
356  mgu.io.in.info.vsew := outVecCtrl.vsew
357  mgu.io.in.info.vdIdx := Mux(outIsResuction, 0.U, outVecCtrl.vuopIdx)
358  mgu.io.in.info.narrow := outVecCtrl.isNarrow
359  mgu.io.in.info.dstMask := outVecCtrl.isDstMask
360  mgu.io.in.isIndexedVls := false.B
361  val resultFpMask = Wire(UInt(VLEN.W))
362  val isFclass = outVecCtrl.fpu.isFpToVecInst && (outCtrl.fuOpType === VfaluType.vfclass)
363  val fpCmpFuOpType = Seq(VfaluType.vfeq, VfaluType.vflt, VfaluType.vfle)
364  val isCmp = outVecCtrl.fpu.isFpToVecInst && (fpCmpFuOpType.map(_ === outCtrl.fuOpType).reduce(_|_))
365  resultFpMask := Mux(isFclass || isCmp, Fill(16, 1.U(1.W)), Fill(VLEN, 1.U(1.W)))
366  io.out.bits.res.data := mgu.io.out.vd & resultFpMask
367  io.out.bits.ctrl.exceptionVec.get(ExceptionNO.illegalInstr) := mgu.io.out.illegal
368
369}
370
371class VFMgu(vlen:Int)(implicit p: Parameters) extends Module{
372  val io = IO(new VFMguIO(vlen))
373
374  val vd = io.in.vd
375  val oldvd = io.in.oldVd
376  val mask = io.in.mask
377  val vsew = io.in.info.eew
378  val num16bits = vlen / 16
379
380}
381
382class VFMguIO(vlen: Int)(implicit p: Parameters) extends Bundle {
383  val in = new Bundle {
384    val vd = Input(UInt(vlen.W))
385    val oldVd = Input(UInt(vlen.W))
386    val mask = Input(UInt(vlen.W))
387    val info = Input(new VecInfo)
388  }
389  val out = new Bundle {
390    val vd = Output(UInt(vlen.W))
391  }
392}