fixed the recoding logic

Author: vikramjain236

src/main/scala/gemmini/Arithmetic.scala

@@ -331,7 +331,7 @@ object Arithmetic {
     override implicit def cast(self: Float): ArithmeticOps[Float] = new ArithmeticOps(self) {
       override def *(t: Float): Float = {
         val t_rec = if (t.isRecoded) t.bits else recFNFromFN(t.expWidth, t.sigWidth, t.bits)
-        val self_rec = if (t.isRecoded) t.bits else recFNFromFN(self.expWidth, self.sigWidth, self.bits)
+        val self_rec = if (self.isRecoded) self.bits else recFNFromFN(self.expWidth, self.sigWidth, self.bits)
 
         val t_resizer =  Module(new RecFNToRecFN(t.expWidth, t.sigWidth, self.expWidth, self.sigWidth))
         t_resizer.io.in := t_rec
@@ -488,7 +488,7 @@ object Arithmetic {
         resizer.io.roundingMode := consts.round_near_even // consts.round_near_maxMag
         resizer.io.detectTininess := consts.tininess_afterRounding
 
-        val result = Wire(Float(t.expWidth, t.sigWidth, self.isRecoded))
+        val result = Wire(Float(t.expWidth, t.sigWidth, t.isRecoded))
         result.bits := (if (result.isRecoded) resizer.io.out else fNFromRecFN(t.expWidth, t.sigWidth, resizer.io.out))
         result
       }
@@ -502,7 +502,7 @@ object Arithmetic {
         resizer.io.roundingMode := consts.round_near_even // consts.round_near_maxMag
         resizer.io.detectTininess := consts.tininess_afterRounding
 
-        val result = Wire(Float(t.expWidth, t.sigWidth, self.isRecoded))
+        val result = Wire(Float(t.expWidth, t.sigWidth, t.isRecoded))
         result.bits := (if (result.isRecoded) resizer.io.out else fNFromRecFN(t.expWidth, t.sigWidth, resizer.io.out))
         result
       }

src/main/scala/gemmini/ExecuteController.scala

@@ -833,9 +833,9 @@ class ExecuteController[T <: Data, U <: Data, V <: Data](xLen: Int, tagWidth: In
   val dataB_unpadded = MuxCase(readData(cntl.b_bank), Seq(cntl.accumulate_zeros -> 0.U, cntl.b_read_from_acc -> accReadData(cntl.b_bank_acc)))
   val dataD_unpadded = MuxCase(readData(cntl.d_bank), Seq(cntl.preload_zeros -> 0.U, cntl.d_read_from_acc -> accReadData(cntl.d_bank_acc)))
 
-  val dataA = VecInit(dataA_unpadded.asTypeOf(Vec(block_size, inputType)).zipWithIndex.map { case (d, i) => Mux(i.U < cntl.a_unpadded_cols, d, inputType.zero)})
-  val dataB = VecInit(dataB_unpadded.asTypeOf(Vec(block_size, inputType)).zipWithIndex.map { case (d, i) => Mux(i.U < cntl.b_unpadded_cols, d, inputType.zero)})
-  val dataD = VecInit(dataD_unpadded.asTypeOf(Vec(block_size, inputType)).zipWithIndex.map { case (d, i) => Mux(i.U < cntl.d_unpadded_cols, d, inputType.zero)})
+  val dataA = VecInit(dataA_unpadded.asTypeOf(Vec(block_size, inputType)).zipWithIndex.map { case (d, i) => Mux(i.U < cntl.a_unpadded_cols, d, inputType.zero)}.map(d => d.asTypeOf(inputType).withWidthOf(spatialArrayInputType)))
+  val dataB = VecInit(dataB_unpadded.asTypeOf(Vec(block_size, inputType)).zipWithIndex.map { case (d, i) => Mux(i.U < cntl.b_unpadded_cols, d, inputType.zero)}.map(d => d.asTypeOf(inputType).withWidthOf(spatialArrayWeightType)))
+  val dataD = VecInit(dataD_unpadded.asTypeOf(Vec(block_size, inputType)).zipWithIndex.map { case (d, i) => Mux(i.U < cntl.d_unpadded_cols, d, inputType.zero)}.map(d => d.asTypeOf(inputType).withWidthOf(spatialArrayWeightType)))
 
   // Pop responses off the scratchpad io ports
   when (mesh_cntl_signals_q.io.deq.fire) {
@@ -876,9 +876,9 @@ class ExecuteController[T <: Data, U <: Data, V <: Data](xLen: Int, tagWidth: In
     mesh.io.b.valid := cntl.b_fire && dataB_valid
     mesh.io.d.valid := cntl.d_fire && dataD_valid
 
-    mesh.io.a.bits := dataA.asTypeOf(Vec(meshRows, Vec(tileRows, inputType)))
-    mesh.io.b.bits := dataB.asTypeOf(Vec(meshColumns, Vec(tileColumns, inputType)))
-    mesh.io.d.bits := dataD.asTypeOf(Vec(meshColumns, Vec(tileColumns, inputType)))
+    mesh.io.a.bits := dataA.asTypeOf(Vec(meshRows, Vec(tileRows, spatialArrayInputType)))
+    mesh.io.b.bits := dataB.asTypeOf(Vec(meshColumns, Vec(tileColumns, spatialArrayWeightType)))
+    mesh.io.d.bits := dataD.asTypeOf(Vec(meshColumns, Vec(tileColumns, spatialArrayWeightType)))
 
     mesh.io.req.valid := mesh_cntl_signals_q.io.deq.fire && (cntl.a_fire || cntl.b_fire || cntl.d_fire)
 
@@ -888,13 +888,13 @@ class ExecuteController[T <: Data, U <: Data, V <: Data](xLen: Int, tagWidth: In
   }
 
   when (cntl_valid && cntl.perform_single_preload) {
-    mesh.io.a.bits := Mux(a_should_be_fed_into_transposer, dataA.asUInt, 0.U).asTypeOf(Vec(meshRows, Vec(tileRows, inputType)))
-    mesh.io.b.bits := Mux(b_should_be_fed_into_transposer, dataB.asUInt, 0.U).asTypeOf(Vec(meshColumns, Vec(tileColumns, inputType)))
+    mesh.io.a.bits := Mux(a_should_be_fed_into_transposer, dataA.asUInt, 0.U).asTypeOf(Vec(meshRows, Vec(tileRows, spatialArrayInputType)))
+    mesh.io.b.bits := Mux(b_should_be_fed_into_transposer, dataB.asUInt, 0.U).asTypeOf(Vec(meshColumns, Vec(tileColumns, spatialArrayWeightType)))
   }
 
   when (cntl_valid && cntl.perform_single_mul) {
-    mesh.io.a.bits := Mux(a_should_be_fed_into_transposer, 0.U, dataA.asUInt).asTypeOf(Vec(meshRows, Vec(tileRows, inputType)))
-    mesh.io.b.bits := Mux(b_should_be_fed_into_transposer, 0.U, dataB.asUInt).asTypeOf(Vec(meshColumns, Vec(tileColumns, inputType)))
+    mesh.io.a.bits := Mux(a_should_be_fed_into_transposer, 0.U, dataA.asUInt).asTypeOf(Vec(meshRows, Vec(tileRows, spatialArrayInputType)))
+    mesh.io.b.bits := Mux(b_should_be_fed_into_transposer, 0.U, dataB.asUInt).asTypeOf(Vec(meshColumns, Vec(tileColumns, spatialArrayWeightType)))
     mesh.io.req.bits.tag.addr.make_this_garbage()
   }