fix bench

.github/cases/blastoise/default.json

@@ -26,7 +26,7 @@
   "codegen.vasubu_vv": 92300,
   "codegen.vasubu_vx": 300458,
   "codegen.vcompress_vm": 19981,
-  "codegen.vcpop_m": 1691,
+  "codegen.vcpop_m": 1661,
   "codegen.vdiv_vv": 32739,
   "codegen.vdiv_vx": 225056,
   "codegen.vdivu_vv": 32835,
@@ -499,4 +499,4 @@
   "codegen.vfredusum_vs": 99859,
   "codegen.vfredmax_vs": 99859,
   "codegen.vfredmin_vs": 99859
-}
+}

.github/cases/machamp/default.json

@@ -24,7 +24,7 @@
   "codegen.vasubu_vv": 112654,
   "codegen.vasubu_vx": 334220,
   "codegen.vcompress_vm": 31353,
-  "codegen.vcpop_m": 1967,
+  "codegen.vcpop_m": 1943,
   "codegen.vdiv_vv": 38992,
   "codegen.vdiv_vx": 236842,
   "codegen.vdivu_vv": 39002,
@@ -435,4 +435,4 @@
   "codegen.vxor_vx": 83889,
   "codegen.vzext_vf2": 40007,
   "codegen.vzext_vf4": 6592
-}
+}

.github/cases/sandslash/default.json

@@ -24,7 +24,7 @@
   "codegen.vasubu_vv": 278200,
   "codegen.vasubu_vx": 2090344,
   "codegen.vcompress_vm": 136872,
-  "codegen.vcpop_m": 4955,
+  "codegen.vcpop_m": 4931,
   "codegen.vdiv_vv": 193913,
   "codegen.vdiv_vx": 727252,
   "codegen.vdivu_vv": 193676,
@@ -435,4 +435,4 @@
   "codegen.vxor_vx": 523387,
   "codegen.vzext_vf2": 167243,
   "codegen.vzext_vf4": 26777
-}
+}

t1/src/LaneFFO.scala

@@ -21,7 +21,7 @@ class LaneFFO(datapathWidth: Int) extends Module {
   @public
   val maskType:     Bool = IO(Input(Bool()))
 
-  val truthMask: UInt = Mux(maskType, src.head, -1.S(datapathWidth.W).asUInt)
+  val truthMask: UInt = Mux(maskType, src.head, -1.S(datapathWidth.W).asUInt) & src(3)
   val srcData:   UInt = truthMask & src(1)
   val notZero:   Bool = srcData.orR
   val lo:        UInt = scanLeftOr(srcData)

t1/src/T1.scala

@@ -751,6 +751,7 @@ class T1(val parameter: T1Parameter) extends Module with SerializableModule[T1Pa
         completedVec.foreach(_ := false.B)
         WARRedResult.valid := false.B
         unOrderTypeInstruction := unOrderType
+        dataResult := 0.U.asTypeOf(dataResult)
       }.elsewhen(control.state.wLast && maskUnitIdle) {
         // 如果真需要执行的lane会wScheduler,不会提前发出last确认
         when(!mixedUnit) {
@@ -806,7 +807,7 @@ class T1(val parameter: T1Parameter) extends Module with SerializableModule[T1Pa
         csrRegForMaskUnit.vl > csrRegForMaskUnit.vStart
       mvToVRF.foreach(d => when(requestRegDequeue.fire){d := writeMv})
       // 读后写中的读
-      val needWAR = maskTypeInstruction || border || (reduce && !popCount) || readMv
+      val needWAR = (maskTypeInstruction || border || reduce || readMv) && !popCount
       val skipLaneData: Bool = decodeResultReg(Decoder.mv)
       mixedUnit := writeMv || readMv
       maskReadLaneSelect.head := UIntToOH(writeBackCounter)
@@ -1240,8 +1241,9 @@ class T1(val parameter: T1Parameter) extends Module with SerializableModule[T1Pa
           decodeResultReg(Decoder.maskDestination) ||
           decodeResultReg(Decoder.ffo)
       // How many data path(32 bit) will used by maskDestination instruction.
+      val maskDestinationByteSize: Bits = csrRegForMaskUnit.vl(log2Ceil(parameter.dLen) - 1, 0) << csrRegForMaskUnit.vSew
       val maskDestinationUseDataPathSize =
-        (csrRegForMaskUnit.vl(log2Ceil(parameter.dLen) - 1, 0) << csrRegForMaskUnit.vSew >> 2).asUInt
+        (maskDestinationByteSize >> 2).asUInt + maskDestinationByteSize(1, 0).orR
       val lastGroupCountForThisGroup: UInt = maskDestinationUseDataPathSize(log2Ceil(parameter.laneNumber) - 1, 0)
       val counterForMaskDestination: UInt = if(parameter.laneNumber > 1) {
         (lastGroupCountForThisGroup - 1.U) |

t1/src/laneStage/LaneExecutionBridge.scala

@@ -541,5 +541,8 @@ class LaneExecutionBridge(parameter: LaneParameter, isLastSlot: Boolean, slotInd
         recordNotExecute)) || reduceLastResponse
   assert(!queue.io.enq.valid || queue.io.enq.ready)
   dequeue <> queue.io.deq
-  updateMaskResult.foreach(_ := !recordQueue.io.deq.bits.sSendResponse.get && queue.io.enq.fire)
+  updateMaskResult.foreach(_ :=
+    (!recordQueue.io.deq.bits.sSendResponse.get && queue.io.enq.fire) ||
+      (enqueue.fire && enqueue.bits.groupCounter === 0.U)
+  )
 }