Save pre-TMEM-bank Radiance changes

src/main/scala/radiance/tile/RadianceTile.scala

@@ -1197,18 +1197,6 @@ class VortexTLAdapter(
     val outResp = chiselTypeOf(outTL._1.d)
   })
   val (bundle, edge) = outTL
-  val sourceGen = Module(
-    new SourceGenerator(
-      newSourceWidth,
-      Some(inReqT.source),
-      ignoreInUse = false
-    )
-  )
-  sourceGen.io.gen := io.outReq.fire // use up a source ID only when request is created
-  sourceGen.io.reclaim.valid := io.outResp.fire
-  sourceGen.io.reclaim.bits := io.outResp.bits.source
-  sourceGen.io.meta := io.inReq.bits.source
-
   // io passthrough logic
   // TLBundleA <> VortexBundleA
   io.outReq.valid := io.inReq.valid
@@ -1217,29 +1205,70 @@ class VortexTLAdapter(
   io.outReq.bits.size := io.inReq.bits.size
   io.outReq.bits.source := io.inReq.bits.source
   io.outReq.bits.address := io.inReq.bits.address
-  // Get requires contiguous mask; only copy core's potentially-partial mask
-  // when writing
+  val outMaskWidth = io.outReq.bits.mask.getWidth
+  val inMaskWidth = io.inReq.bits.mask.getWidth
+  val outDataWidth = io.outReq.bits.data.getWidth
+  val inDataWidth = io.inReq.bits.data.getWidth
+  val byteOffset = io.inReq.bits.address(log2Ceil(outMaskWidth) - 1, 0)
+  val responseOffsetWidth = log2Ceil(outMaskWidth)
+  val responseSourceWidth = inReqT.source.getWidth
+  val sourceGen = Module(
+    new SourceGenerator(
+      newSourceWidth,
+      Some(UInt((responseSourceWidth + responseOffsetWidth).W)),
+      ignoreInUse = false
+    )
+  )
+  sourceGen.io.gen := io.outReq.fire // use up a source ID only when request is created
+  sourceGen.io.reclaim.valid := io.outResp.fire
+  sourceGen.io.reclaim.bits := io.outResp.bits.source
+  sourceGen.io.meta := Cat(byteOffset, io.inReq.bits.source)
+
+  val alignedMask = Wire(UInt(outMaskWidth.W))
+  val alignedData = Wire(UInt(outDataWidth.W))
+  if (outMaskWidth == inMaskWidth) {
+    alignedMask := io.inReq.bits.mask
+  } else {
+    val paddedMask = Wire(UInt(outMaskWidth.W))
+    paddedMask := io.inReq.bits.mask
+    alignedMask := (paddedMask << byteOffset)(outMaskWidth - 1, 0)
+  }
+  if (outDataWidth == inDataWidth) {
+    alignedData := io.inReq.bits.data
+  } else {
+    val paddedData = Wire(UInt(outDataWidth.W))
+    paddedData := io.inReq.bits.data
+    alignedData := (paddedData << (byteOffset << 3))(outDataWidth - 1, 0)
+  }
+
+  // PutFull requires the TL-canonical full mask for address+size; PutPartial
+  // can carry the core-provided byte mask.
   io.outReq.bits.mask := Mux(
-    edge.hasData(io.outReq.bits),
-    io.inReq.bits.mask,
-    // generate TL-correct mask
+    io.outReq.bits.opcode === TLMessages.PutPartialData,
+    alignedMask,
     edge.mask(io.inReq.bits.address, io.inReq.bits.size)
   )
-  io.outReq.bits.data := io.inReq.bits.data
+  io.outReq.bits.data := alignedData
   io.outReq.bits.corrupt := 0.U
   io.inReq.ready := io.outReq.ready
   // VortexBundleD <> TLBundleD
   io.inResp.valid := io.outResp.valid
   io.inResp.bits.opcode := io.outResp.bits.opcode
   io.inResp.bits.size := io.outResp.bits.size
-  io.inResp.bits.source := io.outResp.bits.source
-  io.inResp.bits.data := io.outResp.bits.data
+  val responseMeta = sourceGen.io.peek.asUInt
+  val responseSource = responseMeta(responseSourceWidth - 1, 0)
+  val responseByteOffset =
+    responseMeta(responseSourceWidth + responseOffsetWidth - 1, responseSourceWidth)
+  io.inResp.bits.source := responseSource
+  if (outDataWidth == inDataWidth) {
+    io.inResp.bits.data := io.outResp.bits.data
+  } else {
+    io.inResp.bits.data := (io.outResp.bits.data >> (responseByteOffset << 3))(inDataWidth - 1, 0)
+  }
   io.outResp.ready := io.inResp.ready
 
   // "man-in-the-middle"
   io.inReq.ready := io.outReq.ready && sourceGen.io.id.valid
   io.outReq.valid := io.inReq.valid && sourceGen.io.id.valid
   io.outReq.bits.source := sourceGen.io.id.bits
-  // translate upstream response back to its old sourceId
-  io.inResp.bits.source := sourceGen.io.peek
 }