fix broadcast hub and TL -> NASTI converter to support subblock operations

uncore/src/main/scala/broadcast.scala

@@ -30,11 +30,17 @@ class L2BroadcastHub extends ManagerCoherenceAgent
   val internalDataBits = new DataQueueLocation().getWidth
   val inStoreQueue :: inVolWBQueue :: inClientReleaseQueue :: Nil = Enum(UInt(), nDataQueueLocations)
 
+  val trackerTLParams = params.alterPartial({
+    case TLDataBits => internalDataBits
+    case TLWriteMaskBits => innerWriteMaskBits
+  })
+
   // Create SHRs for outstanding transactions
-  val trackerList = (0 until nReleaseTransactors).map(id =>
+  val trackerList =
-    Module(new BroadcastVoluntaryReleaseTracker(id), {case TLDataBits => internalDataBits})) ++
+    (0 until nReleaseTransactors).map(id =>
-      (nReleaseTransactors until nTransactors).map(id =>
+      Module(new BroadcastVoluntaryReleaseTracker(id))(trackerTLParams)) ++
-        Module(new BroadcastAcquireTracker(id), {case TLDataBits => internalDataBits}))
+    (nReleaseTransactors until nTransactors).map(id =>
+      Module(new BroadcastAcquireTracker(id))(trackerTLParams))
 
   // Propagate incoherence flags
   trackerList.map(_.io.incoherent := io.incoherent)
@@ -100,7 +106,8 @@ class L2BroadcastHub extends ManagerCoherenceAgent
   // Create an arbiter for the one memory port
   val outer_arb = Module(new ClientUncachedTileLinkIOArbiter(trackerList.size),
                          { case TLId => params(OuterTLId)
-                           case TLDataBits => internalDataBits })
+                           case TLDataBits => internalDataBits
+                           case TLWriteMaskBits => innerWriteMaskBits })
   outer_arb.io.in <> trackerList.map(_.io.outer)
   // Get the pending data out of the store data queue
   val outer_data_ptr = new DataQueueLocation().fromBits(outer_arb.io.out.acquire.bits.data)
@@ -112,8 +119,6 @@ class L2BroadcastHub extends ManagerCoherenceAgent
   io.outer.acquire.bits.data := MuxLookup(outer_data_ptr.loc, io.irel().data, Array(
                                           inStoreQueue -> sdq(outer_data_ptr.idx),
                                           inVolWBQueue -> vwbdq(outer_data_ptr.idx)))
-  io.outer.acquire.bits.union := Cat(Fill(io.outer.acquire.bits.tlWriteMaskBits, outer_arb.io.out.acquire.bits.union(1)),
-                                   outer_arb.io.out.acquire.bits.union(0))
 
   // Update SDQ valid bits
   when (io.outer.acquire.valid || sdq_enq) {
@@ -209,14 +214,17 @@ class BroadcastAcquireTracker(trackerId: Int) extends BroadcastXactTracker {
   val s_idle :: s_probe :: s_mem_read :: s_mem_write :: s_make_grant :: s_mem_resp :: s_ack :: Nil = Enum(UInt(), 7)
   val state = Reg(init=s_idle)
 
-  val xact = Reg(Bundle(new AcquireFromSrc, { case TLId => params(InnerTLId); case TLDataBits => 0 }))
+  val xact = Reg(Bundle(new AcquireFromSrc, {
+    case TLId => params(InnerTLId)
+    case TLDataBits => 0
+    case TLWriteMaskBits => innerWriteMaskBits
+  }))
   val data_buffer = Reg(Vec(io.iacq().data, innerDataBeats))
   val coh = ManagerMetadata.onReset
 
   assert(!(state != s_idle && xact.isBuiltInType() && 
-      Vec(Acquire.getType, Acquire.putType, Acquire.putAtomicType,
+      Vec(Acquire.putAtomicType, Acquire.prefetchType).contains(xact.a_type)),
-        Acquire.prefetchType).contains(xact.a_type)),
+    "Broadcast Hub does not support PutAtomics or prefetches") // TODO
-    "Broadcast Hub does not support PutAtomics, subblock Gets/Puts, or prefetches") // TODO
 
   val release_count = Reg(init=UInt(0, width = log2Up(io.inner.tlNCachingClients+1)))
   val pending_probes = Reg(init=Bits(0, width = io.inner.tlNCachingClients))
@@ -236,6 +244,7 @@ class BroadcastAcquireTracker(trackerId: Int) extends BroadcastXactTracker {
   val pending_outer_write_ = io.iacq().hasData()
   val pending_outer_read = io.ignt().hasData()
   val pending_outer_read_ = coh.makeGrant(io.iacq(), UInt(trackerId)).hasData()
+  val subblock_type = xact.isSubBlockType()
 
   io.has_acquire_conflict := xact.conflicts(io.iacq()) && 
                               (state != s_idle) &&
@@ -246,22 +255,32 @@ class BroadcastAcquireTracker(trackerId: Int) extends BroadcastXactTracker {
                             !io.irel().isVoluntary() &&
                             (state === s_probe)
 
-  val outer_write_acq = Bundle(PutBlock(
+  val oacq_type = MuxLookup(state, Acquire.getBlockType, Seq(
-                                client_xact_id = UInt(trackerId),
+    (s_probe, Acquire.putBlockType),
-                                addr_block = xact.addr_block,
+    (s_mem_write, Mux(subblock_type, Acquire.putType, Acquire.putBlockType)),
-                                addr_beat = oacq_data_cnt,
+    (s_mem_read, Mux(subblock_type, Acquire.getType, Acquire.getBlockType))))
-                                data = data_buffer(oacq_data_cnt)))(outerTLParams)
+  val oacq_beat = MuxLookup(state, UInt(0), Seq(
-  val outer_write_rel = Bundle(PutBlock(
+    (s_probe, io.irel().addr_beat),
-                                client_xact_id = UInt(trackerId),
+    (s_mem_write, Mux(subblock_type, xact.addr_beat, oacq_data_cnt)),
-                                addr_block = xact.addr_block,
+    (s_mem_read, Mux(subblock_type, xact.addr_beat, UInt(0)))))
-                                addr_beat = io.irel().addr_beat,
+  val oacq_data = MuxLookup(state, Bits(0), Seq(
-                                data = io.irel().data))(outerTLParams)
+    (s_probe, io.irel().data),
-  val outer_read = Bundle(GetBlock(
+    (s_mem_write, Mux(subblock_type,
-                            client_xact_id = UInt(trackerId),
+      data_buffer(0), data_buffer(oacq_data_cnt)))))
-                            addr_block = xact.addr_block))(outerTLParams)
+  val oacq_union = MuxLookup(state, Bits(0), Seq(
+    (s_probe, Acquire.fullWriteMask),
+    (s_mem_write, xact.wmask()),
+    (s_mem_read, Cat(xact.addr_byte(), xact.op_size(), M_XRD))))
 
   io.outer.acquire.valid := Bool(false)
-  io.outer.acquire.bits := outer_read //default
+  io.outer.acquire.bits := Bundle(Acquire(
+    is_builtin_type = Bool(true),
+    a_type = oacq_type,
+    client_xact_id = UInt(trackerId),
+    addr_block = xact.addr_block,
+    addr_beat = oacq_beat,
+    data = oacq_data,
+    union = Cat(oacq_union, Bool(true))))(outerTLParams)
   io.outer.grant.ready := Bool(false)
 
   io.inner.probe.valid := Bool(false)
@@ -331,7 +350,6 @@ class BroadcastAcquireTracker(trackerId: Int) extends BroadcastXactTracker {
       when(io.inner.release.valid) {
         when(io.irel().hasData()) {
           io.outer.acquire.valid := Bool(true)
-          io.outer.acquire.bits := outer_write_rel
           when(io.outer.acquire.ready) {
             when(oacq_data_done) {
               pending_ognt_ack := Bool(true)
@@ -353,7 +371,6 @@ class BroadcastAcquireTracker(trackerId: Int) extends BroadcastXactTracker {
     }
     is(s_mem_write) { // Write data to outer memory
       io.outer.acquire.valid := !pending_ognt_ack || !collect_iacq_data || iacq_data_valid(oacq_data_cnt)
-      io.outer.acquire.bits := outer_write_acq
       when(oacq_data_done) {
         pending_ognt_ack := Bool(true)
         state := Mux(pending_outer_read, s_mem_read, s_mem_resp)
@@ -361,7 +378,6 @@ class BroadcastAcquireTracker(trackerId: Int) extends BroadcastXactTracker {
     }
     is(s_mem_read) { // Read data from outer memory (possibly what was just written)
       io.outer.acquire.valid := !pending_ognt_ack
-      io.outer.acquire.bits := outer_read
       when(io.outer.acquire.fire()) { state := s_mem_resp }
     }
     is(s_mem_resp) { // Wait to forward grants from outer memory

uncore/src/main/scala/tilelink.scala

@@ -34,6 +34,8 @@ case object TLDataBits extends Field[Int]
 case object TLDataBeats extends Field[Int]
 /** Whether the underlying physical network preserved point-to-point ordering of messages */
 case object TLNetworkIsOrderedP2P extends Field[Boolean]
+/** Number of bits in write mask (usually one per byte in beat) */
+case object TLWriteMaskBits extends Field[Int]
 
 /** Utility trait for building Modules and Bundles that use TileLink parameters */
 trait TileLinkParameters extends UsesParameters {
@@ -53,7 +55,7 @@ trait TileLinkParameters extends UsesParameters {
   val tlDataBits = params(TLDataBits)
   val tlDataBytes = tlDataBits/8
   val tlDataBeats = params(TLDataBeats)
-  val tlWriteMaskBits = if(tlDataBits/8 < 1) 1 else tlDataBits/8
+  val tlWriteMaskBits = params(TLWriteMaskBits)
   val tlBeatAddrBits = log2Up(tlDataBeats)
   val tlByteAddrBits = log2Up(tlWriteMaskBits)
   val tlMemoryOpcodeBits = M_SZ
@@ -1274,11 +1276,38 @@ class NASTIMasterIOTileLinkIOConverter extends TLModule with NASTIParameters {
   val addr_out = Reg(UInt(width = nastiXAddrBits))
   val has_data = Reg(init=Bool(false))
   val data_from_rel = Reg(init=Bool(false))
+  val is_subblock = io.tl.acquire.bits.isSubBlockType()
   val (tl_cnt_out, tl_wrap_out) =
-    Counter((io.tl.acquire.fire() && acq_has_data) ||
+    Counter((io.tl.acquire.fire() && io.tl.acquire.bits.hasMultibeatData()) ||
               (io.tl.release.fire() && rel_has_data), tlDataBeats)
   val tl_done_out = Reg(init=Bool(false))
 
+  val roq_size = 4
+  val roq_data  = Reg(Vec(UInt(width = tlByteAddrBits), roq_size))
+  val roq_tags = Reg(Vec(UInt(width = nastiRIdBits), roq_size))
+  val roq_free = Reg(init = Fill(roq_size, Bits(1, 1)))
+  val roq_full = !roq_free.orR
+
+  val roq_enq_addr = PriorityEncoder(roq_free)
+  val roq_enq_valid = io.tl.acquire.fire() && !acq_has_data && is_subblock
+  val roq_enq_data = io.tl.acquire.bits.addr_byte()
+  val roq_enq_tag = io.nasti.ar.bits.id
+
+  val roq_deq_tag = io.nasti.r.bits.id
+  val roq_deq_addr = PriorityEncoder(roq_tags.map(_ === roq_deq_tag))
+  val roq_deq_data = roq_data(roq_deq_addr)
+  val roq_deq_valid = io.nasti.r.fire() && !io.nasti.r.bits.id(0)
+
+  when (roq_enq_valid) {
+    roq_data(roq_enq_addr) := roq_enq_data
+    roq_tags(roq_enq_addr) := roq_enq_tag
+    roq_free(roq_enq_addr) := Bool(false)
+  }
+
+  when (roq_deq_valid) {
+    roq_free(roq_deq_addr) := Bool(true)
+  }
+
   io.nasti.ar.bits.id := tag_out
   io.nasti.ar.bits.addr := addr_out
   io.nasti.ar.bits.len := Mux(has_data, UInt(tlDataBeats-1), UInt(0)) 
@@ -1293,7 +1322,7 @@ class NASTIMasterIOTileLinkIOConverter extends TLModule with NASTIParameters {
   io.nasti.aw.bits := io.nasti.ar.bits
   io.nasti.w.bits.strb := Mux(data_from_rel, SInt(-1), io.tl.acquire.bits.wmask())
   io.nasti.w.bits.data := Mux(data_from_rel, io.tl.release.bits.data, io.tl.acquire.bits.data)
-  io.nasti.w.bits.last := tl_wrap_out
+  io.nasti.w.bits.last := tl_wrap_out || (io.tl.acquire.fire() && is_subblock)
 
   when(!active_out){
     io.tl.release.ready := io.nasti.w.ready
@@ -1307,7 +1336,6 @@ class NASTIMasterIOTileLinkIOConverter extends TLModule with NASTIParameters {
       io.nasti.aw.valid := is_write
       io.nasti.ar.valid := !is_write
       cmd_sent_out := (!is_write && io.nasti.ar.ready) || (is_write && io.nasti.aw.ready)
-      tl_done_out := tl_wrap_out
       when(io.tl.release.valid) {
         data_from_rel := Bool(true)
         io.nasti.w.bits.data := io.tl.release.bits.data
@@ -1319,34 +1347,35 @@ class NASTIMasterIOTileLinkIOConverter extends TLModule with NASTIParameters {
         io.nasti.aw.bits.id := tag
         io.nasti.aw.bits.addr := addr
         io.nasti.aw.bits.len := UInt(tlDataBeats-1)
-        io.nasti.aw.bits.size := MT_Q
         tag_out := tag
         addr_out := addr
         has_data := rel_has_data
+        tl_done_out := tl_wrap_out
       } .elsewhen(io.tl.acquire.valid) {
         data_from_rel := Bool(false)
         io.nasti.w.bits.data := io.tl.acquire.bits.data
         io.nasti.w.bits.strb := io.tl.acquire.bits.wmask()
+        // The last bit indicates to the Grant logic what g_type to send back
+        // For read, true = getDataBlockType, false = getDataBeatType
+        // For write, it should always be false, so that putAckType is sent
         val tag = Cat(io.tl.acquire.bits.client_id,
                       io.tl.acquire.bits.client_xact_id,
-                      io.tl.acquire.bits.isBuiltInType())
+                      !is_write && !is_subblock)
         val addr = io.tl.acquire.bits.full_addr()
         when(is_write) {
           io.nasti.aw.bits.id := tag
           io.nasti.aw.bits.addr := addr
-          io.nasti.aw.bits.len := Mux(io.tl.acquire.bits.isBuiltInType(Acquire.putBlockType),
+          io.nasti.aw.bits.len := Mux(!is_subblock, UInt(tlDataBeats-1), UInt(0))
-                                    UInt(tlDataBeats-1), UInt(0)) 
-          io.nasti.aw.bits.size := bytesToXSize(PopCount(io.tl.acquire.bits.wmask()))
         } .otherwise {
           io.nasti.ar.bits.id := tag
           io.nasti.ar.bits.addr := addr
-          io.nasti.ar.bits.len := Mux(io.tl.acquire.bits.isBuiltInType(Acquire.getBlockType),
+          io.nasti.ar.bits.len := Mux(!is_subblock, UInt(tlDataBeats-1), UInt(0))
-                                    UInt(tlDataBeats-1), UInt(0)) 
           io.nasti.ar.bits.size := io.tl.acquire.bits.op_size()
         }
         tag_out := tag
         addr_out := addr
         has_data := acq_has_data
+        tl_done_out := tl_wrap_out || is_subblock
       }
     }
   }
@@ -1364,24 +1393,36 @@ class NASTIMasterIOTileLinkIOConverter extends TLModule with NASTIParameters {
       }
     }
     when(tl_wrap_out) { tl_done_out := Bool(true) }
-    when(cmd_sent_out && (!has_data || tl_done_out)) { active_out := Bool(false) }
+    when(cmd_sent_out && !roq_full && (!has_data || tl_done_out)) {
+      active_out := Bool(false)
+    }
   }
 
+  assert (!io.nasti.r.valid || !io.nasti.r.bits.resp(1),
+    "NASTI read response error")
+  assert (!io.nasti.b.valid || !io.nasti.b.bits.resp(1),
+    "NASTI write response error")
+
   // Aggregate incoming NASTI responses into TL Grants
   val (tl_cnt_in, tl_wrap_in) = Counter(io.tl.grant.fire() && io.tl.grant.bits.hasMultibeatData(), tlDataBeats)
   val gnt_arb = Module(new Arbiter(new GrantToDst, 2))
   io.tl.grant <> gnt_arb.io.out
 
+  val r_aligned_data = Mux(io.nasti.r.bits.id(0),
+    io.nasti.r.bits.data,
+    io.nasti.r.bits.data << Cat(roq_deq_data, UInt(0, 3)))
+
   gnt_arb.io.in(0).valid := io.nasti.r.valid
   io.nasti.r.ready := gnt_arb.io.in(0).ready
   gnt_arb.io.in(0).bits := Grant(
     dst = (if(dstIdBits > 0) io.nasti.r.bits.id(dst_off, tlClientXactIdBits + 1) else UInt(0)),
-    is_builtin_type = io.nasti.r.bits.id(0),
+    is_builtin_type = Bool(true),
-    g_type = Mux(io.nasti.r.bits.id(0), Grant.getDataBlockType, UInt(0)), // TODO: Assumes MI or MEI protocol
+    g_type = Mux(io.nasti.r.bits.id(0),
+      Grant.getDataBlockType, Grant.getDataBeatType), // TODO: Assumes MI or MEI protocol
     client_xact_id = io.nasti.r.bits.id >> 1,
     manager_xact_id = UInt(0),
     addr_beat = tl_cnt_in,
-    data = io.nasti.r.bits.data)
+    data = r_aligned_data)
 
   gnt_arb.io.in(1).valid := io.nasti.b.valid
   io.nasti.b.ready := gnt_arb.io.in(1).ready

uncore/src/main/scala/uncore.scala

@@ -20,6 +20,7 @@ trait CoherenceAgentParameters extends UsesParameters {
   val innerTLParams = params.alterPartial({case TLId => params(InnerTLId)})
   val innerDataBeats = innerTLParams(TLDataBeats)
   val innerDataBits = innerTLParams(TLDataBits)
+  val innerWriteMaskBits = innerTLParams(TLWriteMaskBits)
   val innerBeatAddrBits = log2Up(innerDataBeats)
   val innerByteAddrBits = log2Up(innerDataBits/8)
   require(outerDataBeats == innerDataBeats) //TODO: must fix all xact_data Vecs to remove this requirement