first steps in uncore constant/tilelink data refactor

.gitignore

@@ -1,2 +1,3 @@
 target/
 project/target
+*.swp

Makefrag

@@ -109,6 +109,7 @@ asm_p_tests = \
 	rv64uf-p-fcmp \
 	rv64uf-p-fcvt \
 	rv64uf-p-fcvt_w \
+	rv64uf-p-fclass \
 	rv64uf-p-fadd \
 	rv64uf-p-fmin \
 	rv64uf-p-fmadd \
@@ -205,6 +206,7 @@ asm_v_tests = \
 	rv64uf-v-fcmp \
 	rv64uf-v-fcvt \
 	rv64uf-v-fcvt_w \
+	rv64uf-v-fclass \
 	rv64uf-v-fadd \
 	rv64uf-v-fmin \
 	rv64uf-v-fmadd \

csrc/emulator.cc

@@ -139,7 +139,7 @@ int main(int argc, char** argv)
 
   if (htif->exit_code())
   {
-    fprintf(stderr, "*** FAILED *** (code = %d) after %lld cycles\n", htif->exit_code(), (long long)trace_count);
+    fprintf(stderr, "*** FAILED *** (code = %d, seed %d) after %lld cycles\n", htif->exit_code(), random_seed, (long long)trace_count);
     ret = htif->exit_code();
   }
   else if (trace_count == max_cycles)

csrc/vcs_main.cc

@@ -8,12 +8,41 @@
 #include <sstream>
 #include <iterator>
 
-static htif_emulator_t* htif = NULL;
-static unsigned htif_bytes;
-static mm_t* mm = NULL;
-
 extern "C" {
 
+extern int vcs_main(int argc, char** argv);
+
+static htif_emulator_t* htif;
+static unsigned htif_bytes;
+static mm_t* mm;
+static const char* loadmem;
+
+void htif_fini(vc_handle failure)
+{
+  delete htif;
+  htif = NULL;
+  exit(vc_getScalar(failure));
+}
+
+int main(int argc, char** argv)
+{
+  bool dramsim = false;
+
+  for (int i = 1; i < argc; i++)
+  {
+    if (!strcmp(argv[i], "+dramsim"))
+      dramsim = true;
+    else if (!strncmp(argv[i], "+loadmem=", 9))
+      loadmem = argv[i]+9;
+  }
+
+  mm = dramsim ? (mm_t*)(new mm_dramsim2_t) : (mm_t*)(new mm_magic_t);
+  htif = new htif_emulator_t(std::vector<std::string>(argv + 1, argv + argc));
+
+  vcs_main(argc, argv);
+  abort(); // should never get here
+}
+
 void memory_tick(
   vc_handle mem_req_val,
   vc_handle mem_req_rdy,
@@ -62,55 +91,18 @@ void memory_tick(
   );
 }
 
-void htif_init
+void htif_init(vc_handle htif_width, vc_handle mem_width)
-(
-  vc_handle htif_width,
-  vc_handle mem_width,
-  vc_handle argv,
-  vc_handle loadmem,
-  vc_handle dramsim
-)
 {
   int mw = vc_4stVectorRef(mem_width)->d;
-  mm = vc_getScalar(dramsim) ? (mm_t*)(new mm_dramsim2_t) : (mm_t*)(new mm_magic_t);
   assert(mw && (mw & (mw-1)) == 0);
   mm->init(MEM_SIZE, mw/8, LINE_SIZE);
 
+  if (loadmem)
+    load_mem(mm->get_data(), loadmem);
+
   vec32* w = vc_4stVectorRef(htif_width);
   assert(w->d <= 32 && w->d % 8 == 0); // htif_tick assumes data fits in a vec32
   htif_bytes = w->d/8;
-
-  char loadmem_str[1024];
-  vc_VectorToString(loadmem, loadmem_str);
-  if (*loadmem_str)
-    load_mem(mm->get_data(), loadmem_str);
-
-  char argv_str[1024];
-  vc_VectorToString(argv, argv_str);
-  if (!*argv_str)
-  {
-    if (*loadmem_str)
-      strcpy(argv_str, "none");
-    else
-    {
-      fprintf(stderr, "Usage: ./simv [host options] +argv=\"<target program> [target args]\"\n");
-      exit(-1);
-    }
-  }
-
-  std::vector<std::string> args;
-  std::stringstream ss(argv_str);
-  std::istream_iterator<std::string> begin(ss), end;
-  std::copy(begin, end, std::back_inserter<std::vector<std::string>>(args));
-
-  htif = new htif_emulator_t(args);
-}
-
-void htif_fini(vc_handle failure)
-{
-  delete htif;
-  htif = NULL;
-  exit(vc_getScalar(failure));
 }
 
 void htif_tick

emulator/Makefile

@@ -15,7 +15,7 @@ LDFLAGS := $(LDFLAGS) -L$(RISCV)/lib -Wl,-rpath,$(RISCV)/lib -L. -ldramsim -lfes
 OBJS := $(addsuffix .o,$(CXXSRCS) $(MODEL))
 DEBUG_OBJS := $(addsuffix -debug.o,$(CXXSRCS) $(MODEL))
 
-CHISEL_ARGS := $(MODEL) --noIoDebug --backend c --targetDir emulator/generated-src
+CHISEL_ARGS := $(MODEL) --noIoDebug --backend Chisel.Fame1CppBackend --targetDir emulator/generated-src
 CHISEL_ARGS_DEBUG := $(CHISEL_ARGS)-debug --debug --vcd --ioDebug
 
 generated-src/$(MODEL).h: $(base_dir)/rocket/$(src_path)/*.scala $(base_dir)/hwacha/$(src_path)/*.scala $(base_dir)/uncore/$(src_path)/*.scala $(base_dir)/$(src_path)/*.scala

project/build.scala

@@ -31,7 +31,8 @@ object BuildSettings extends Build {
   lazy val uncore = Project("uncore", file("uncore"), settings = buildSettings) dependsOn(hardfloat)
   lazy val rocket = Project("rocket", file("rocket"), settings = buildSettings) dependsOn(uncore)
   lazy val hwacha = Project("hwacha", file("hwacha"), settings = buildSettings) dependsOn(uncore, rocket)
-  lazy val referencechip = Project("referencechip", file("."), settings = buildSettings ++ chipSettings) dependsOn(rocket, hwacha)
+  lazy val rekall = Project("rekall", file("rekall"), settings = buildSettings) dependsOn(chisel)
+  lazy val referencechip = Project("referencechip", file("."), settings = buildSettings ++ chipSettings) dependsOn(rocket, hwacha, rekall)
 
   val elaborateTask = InputKey[Unit]("elaborate", "convert chisel components into backend source code")
   val makeTask = InputKey[Unit]("make", "trigger backend-specific makefile command")

src/main/scala/RocketChip.scala

@@ -5,11 +5,11 @@ import uncore._
 import rocket._
 import rocket.Util._
 import ReferenceChipBackend._
-import scala.collection.mutable.ArrayBuffer
 import scala.collection.mutable.HashMap
+import DRAMModel._
 
 object DummyTopLevelConstants {
-  val NTILES = 1
+  val NTILES = 2
   val NBANKS = 1
   val HTIF_WIDTH = 16
   val ENABLE_SHARING = true
@@ -18,6 +18,10 @@ object DummyTopLevelConstants {
   val NL2_REL_XACTS = 1
   val NL2_ACQ_XACTS = 7
   val NMSHRS = 2
+  val MEM_TAG_BITS = 5
+  val MEM_DATA_BITS = 128
+  val MEM_ADDR_BITS = PADDR_BITS - OFFSET_BITS
+  val MEM_DATA_BEATS = 4
 }
 
 import DummyTopLevelConstants._
@@ -83,22 +87,25 @@ class ReferenceChipBackend extends VerilogBackend
   transforms += ((c: Module) => collectNodesIntoComp(initializeDFS))
 }
 
+class Fame1ReferenceChipBackend extends ReferenceChipBackend with Fame1Transform
+
 class OuterMemorySystem(htif_width: Int)(implicit conf: UncoreConfiguration) extends Module
 {
-  implicit val (tl, ln, l2) = (conf.tl, conf.tl.ln, conf.l2)
+  implicit val (tl, ln, l2, mif) = (conf.tl, conf.tl.ln, conf.l2, conf.mif)
   val io = new Bundle {
     val tiles = Vec.fill(conf.nTiles){new TileLinkIO}.flip
     val htif = (new TileLinkIO).flip
     val incoherent = Vec.fill(ln.nClients){Bool()}.asInput
-    val mem = new ioMem
+    val mem = new MemIO
-    val mem_backup = new ioMemSerialized(htif_width)
+    val mem_backup = new MemSerializedIO(htif_width)
     val mem_backup_en = Bool(INPUT)
   }
 
+  val refill_cycles = tl.dataBits/mif.dataBits
   val llc_tag_leaf = Mem(Bits(width = 152), 512, seqRead = true)
   val llc_data_leaf = Mem(Bits(width = 64), 4096, seqRead = true)
-  val llc = Module(new DRAMSideLLC(sets=512, ways=8, outstanding=16, tagLeaf=llc_tag_leaf, dataLeaf=llc_data_leaf))
+  val llc = Module(new DRAMSideLLC(sets=512, ways=8, outstanding=16, refill_cycles=refill_cycles, tagLeaf=llc_tag_leaf, dataLeaf=llc_data_leaf))
-  //val llc = Module(new DRAMSideLLCNull(NL2_REL_XACTS+NL2_ACQ_XACTS, REFILL_CYCLES))
+  //val llc = Module(new DRAMSideLLCNull(NL2_REL_XACTS+NL2_ACQ_XACTS, refill_cycles))
   val mem_serdes = Module(new MemSerdes(htif_width))
 
   val masterEndpoints = (0 until ln.nMasters).map(i => Module(new L2CoherenceAgent(i)))
@@ -116,7 +123,7 @@ class OuterMemorySystem(htif_width: Int)(implicit conf: UncoreConfiguration) ext
     conv.io.uncached <> masterEndpoints.head.io.master
   }
   llc.io.cpu.req_cmd <> Queue(conv.io.mem.req_cmd)
-  llc.io.cpu.req_data <> Queue(conv.io.mem.req_data, REFILL_CYCLES)
+  llc.io.cpu.req_data <> Queue(conv.io.mem.req_data, refill_cycles)
   conv.io.mem.resp <> llc.io.cpu.resp
 
   // mux between main and backup memory ports
@@ -128,7 +135,7 @@ class OuterMemorySystem(htif_width: Int)(implicit conf: UncoreConfiguration) ext
   mem_serdes.io.wide.req_cmd.valid := mem_cmdq.io.deq.valid && io.mem_backup_en
   mem_serdes.io.wide.req_cmd.bits := mem_cmdq.io.deq.bits
 
-  val mem_dataq = Module(new Queue(new MemData, REFILL_CYCLES))
+  val mem_dataq = Module(new Queue(new MemData, refill_cycles))
   mem_dataq.io.enq <> llc.io.mem.req_data
   mem_dataq.io.deq.ready := Mux(io.mem_backup_en, mem_serdes.io.wide.req_data.ready, io.mem.req_data.ready)
   io.mem.req_data.valid := mem_dataq.io.deq.valid && !io.mem_backup_en
@@ -143,18 +150,18 @@ class OuterMemorySystem(htif_width: Int)(implicit conf: UncoreConfiguration) ext
   io.mem_backup <> mem_serdes.io.narrow
 }
 
-case class UncoreConfiguration(l2: L2CoherenceAgentConfiguration, tl: TileLinkConfiguration, nTiles: Int, nBanks: Int, bankIdLsb: Int, nSCR: Int)
+case class UncoreConfiguration(l2: L2CoherenceAgentConfiguration, tl: TileLinkConfiguration, mif: MemoryIFConfiguration, nTiles: Int, nBanks: Int, bankIdLsb: Int, nSCR: Int)
 
 class Uncore(htif_width: Int)(implicit conf: UncoreConfiguration) extends Module
 {
-  implicit val tl = conf.tl
+  implicit val (tl, mif) = (conf.tl, conf.mif)
   val io = new Bundle {
     val host = new HostIO(htif_width)
-    val mem = new ioMem
+    val mem = new MemIO
     val tiles = Vec.fill(conf.nTiles){new TileLinkIO}.flip
     val htif = Vec.fill(conf.nTiles){new HTIFIO(conf.nTiles)}.flip
     val incoherent = Vec.fill(conf.nTiles){Bool()}.asInput
-    val mem_backup = new ioMemSerialized(htif_width)
+    val mem_backup = new MemSerializedIO(htif_width)
     val mem_backup_en = Bool(INPUT)
   }
   val htif = Module(new HTIF(htif_width, CSRs.reset, conf.nSCR))
@@ -167,21 +174,15 @@ class Uncore(htif_width: Int)(implicit conf: UncoreConfiguration) extends Module
 
   // Add networking headers and endpoint queues
   def convertAddrToBank(addr: Bits): UInt = {
-    require(conf.bankIdLsb + log2Up(conf.nBanks) < MEM_ADDR_BITS, {println("Invalid bits for bank multiplexing.")})
+    require(conf.bankIdLsb + log2Up(conf.nBanks) < conf.mif.addrBits, {println("Invalid bits for bank multiplexing.")})
     addr(conf.bankIdLsb + log2Up(conf.nBanks) - 1, conf.bankIdLsb)
   }
 
   (outmemsys.io.tiles :+ outmemsys.io.htif).zip(io.tiles :+ htif.io.mem).zipWithIndex.map { 
     case ((outer, client), i) => 
-      outer.acquire <> TileLinkHeaderAppender(client.acquire, i, conf.nBanks, convertAddrToBank _)
+      outer.acquire <> Queue(TileLinkHeaderOverwriter(client.acquire, i, conf.nBanks, convertAddrToBank _))
-      outer.release <> TileLinkHeaderAppender(client.release, i, conf.nBanks, convertAddrToBank _)
+      outer.release <> Queue(TileLinkHeaderOverwriter(client.release, i, conf.nBanks, convertAddrToBank _))
-
+      outer.grant_ack <> Queue(TileLinkHeaderOverwriter(client.grant_ack, i))
-      val grant_ack_q = Queue(client.grant_ack)
-      outer.grant_ack.valid := grant_ack_q.valid
-      outer.grant_ack.bits := grant_ack_q.bits
-      outer.grant_ack.bits.header.src := UInt(i)
-      grant_ack_q.ready := outer.grant_ack.ready
-
       client.grant <> Queue(outer.grant, 1, pipe = true)
       client.probe <> Queue(outer.probe)
   }
@@ -216,12 +217,12 @@ class Uncore(htif_width: Int)(implicit conf: UncoreConfiguration) extends Module
   io.host.debug_stats_pcr := htif.io.host.debug_stats_pcr
 }
 
-class TopIO(htifWidth: Int) extends Bundle  {
+class TopIO(htifWidth: Int)(implicit conf: MemoryIFConfiguration) extends Bundle  {
   val host    = new HostIO(htifWidth)
-  val mem     = new ioMem
+  val mem     = new MemIO
 }
 
-class VLSITopIO(htifWidth: Int) extends TopIO(htifWidth) {
+class VLSITopIO(htifWidth: Int)(implicit conf: MemoryIFConfiguration) extends TopIO(htifWidth)(conf) {
   val mem_backup_en = Bool(INPUT)
   val in_mem_ready = Bool(OUTPUT)
   val in_mem_valid = Bool(INPUT)
@@ -231,14 +232,16 @@ class VLSITopIO(htifWidth: Int) extends TopIO(htifWidth) {
 
 
 class MemDessert extends Module {
+  implicit val mif = MemoryIFConfiguration(MEM_ADDR_BITS, MEM_DATA_BITS, MEM_TAG_BITS, MEM_DATA_BEATS)
   val io = new MemDesserIO(HTIF_WIDTH)
   val x = Module(new MemDesser(HTIF_WIDTH))
   io.narrow <> x.io.narrow
   io.wide <> x.io.wide
 }
 
+
 class Top extends Module {
-  val co =  if(ENABLE_SHARING) {
+  val co = if(ENABLE_SHARING) {
               if(ENABLE_CLEAN_EXCLUSIVE) new MESICoherence
               else new MSICoherence
             } else {
@@ -247,14 +250,15 @@ class Top extends Module {
             }
 
   implicit val ln = LogicalNetworkConfiguration(log2Up(NTILES)+1, NBANKS, NTILES+1)
-  implicit val tl = TileLinkConfiguration(co, ln, log2Up(NL2_REL_XACTS+NL2_ACQ_XACTS), 2*log2Up(NMSHRS*NTILES+1), MEM_DATA_BITS)
+  implicit val tl = TileLinkConfiguration(co, ln, log2Up(NL2_REL_XACTS+NL2_ACQ_XACTS), 2*log2Up(NMSHRS*NTILES+1), CACHE_DATA_SIZE_IN_BYTES*8)
   implicit val l2 = L2CoherenceAgentConfiguration(tl, NL2_REL_XACTS, NL2_ACQ_XACTS)
-  implicit val uc = UncoreConfiguration(l2, tl, NTILES, NBANKS, bankIdLsb = 5, nSCR = 64)
+  implicit val mif = MemoryIFConfiguration(MEM_ADDR_BITS, MEM_DATA_BITS, MEM_TAG_BITS, MEM_DATA_BEATS)
+  implicit val uc = UncoreConfiguration(l2, tl, mif, NTILES, NBANKS, bankIdLsb = 5, nSCR = 64)
 
-  val ic = ICacheConfig(128, 2, ntlb = 8, nbtb = 38)
+  val ic = ICacheConfig(128, 2, ntlb = 8, nbtb = 38, tl = tl)
   val dc = DCacheConfig(128, 4, ntlb = 8, 
-                        nmshr = NMSHRS, nrpq = 16, nsdq = 17, states = co.nClientStates)
+                        nmshr = NMSHRS, nrpq = 16, nsdq = 17, tl = tl)
-  val vic = ICacheConfig(128, 1)
+  val vic = ICacheConfig(128, 1, tl = tl)
   val hc = hwacha.HwachaConfiguration(vic, dc, 8, 256, ndtlb = 8, nptlb = 2)
   val fpu = if (HAS_FPU) Some(FPUConfig(sfmaLatency = 2, dfmaLatency = 3)) else None
   val rc = RocketConfiguration(tl, ic, dc, fpu
@@ -295,3 +299,5 @@ class Top extends Module {
   io.mem_backup_en <> uncore.io.mem_backup_en
   io.mem <> uncore.io.mem
 }
+
+

src/main/scala/fpga.scala

@@ -4,15 +4,17 @@ import Chisel._
 import Node._
 import uncore._
 import rocket._
+import DRAMModel._
+import DRAMModel.MemModelConstants._
 
 class FPGAOuterMemorySystem(htif_width: Int)(implicit conf: UncoreConfiguration) extends Module
 {
-  implicit val (tl, ln, l2) = (conf.tl, conf.tl.ln, conf.l2)
+  implicit val (tl, ln, l2, mif) = (conf.tl, conf.tl.ln, conf.l2, conf.mif)
   val io = new Bundle {
     val tiles = Vec.fill(conf.nTiles){new TileLinkIO}.flip
     val htif = (new TileLinkIO).flip
     val incoherent = Vec.fill(ln.nClients){Bool()}.asInput
-    val mem = new ioMem
+    val mem = new MemIO
   }
 
   val masterEndpoints = (0 until ln.nMasters).map(i => Module(new L2CoherenceAgent(i)))
@@ -31,16 +33,16 @@ class FPGAOuterMemorySystem(htif_width: Int)(implicit conf: UncoreConfiguration)
     conv.io.uncached <> masterEndpoints.head.io.master
   }
   io.mem.req_cmd <> Queue(conv.io.mem.req_cmd)
-  io.mem.req_data <> Queue(conv.io.mem.req_data, REFILL_CYCLES)
+  io.mem.req_data <> Queue(conv.io.mem.req_data, tl.dataBits/mif.dataBits)
   conv.io.mem.resp <> Queue(io.mem.resp)
 }
 
 class FPGAUncore(htif_width: Int)(implicit conf: UncoreConfiguration) extends Module
 {
-  implicit val (tl, ln) = (conf.tl, conf.tl.ln)
+  implicit val (tl, ln, mif) = (conf.tl, conf.tl.ln, conf.mif)
   val io = new Bundle {
     val host = new HostIO(htif_width)
-    val mem = new ioMem
+    val mem = new MemIO
     val tiles = Vec.fill(conf.nTiles){new TileLinkIO}.flip
     val htif = Vec.fill(conf.nTiles){new HTIFIO(conf.nTiles)}.flip
     val incoherent = Vec.fill(conf.nTiles){Bool()}.asInput
@@ -54,21 +56,15 @@ class FPGAUncore(htif_width: Int)(implicit conf: UncoreConfiguration) extends Mo
 
   // Add networking headers and endpoint queues
   def convertAddrToBank(addr: Bits): UInt = {
-    require(conf.bankIdLsb + log2Up(conf.nBanks) < MEM_ADDR_BITS, {println("Invalid bits for bank multiplexing.")})
+    require(conf.bankIdLsb + log2Up(conf.nBanks) < conf.mif.addrBits, {println("Invalid bits for bank multiplexing.")})
     addr(conf.bankIdLsb + log2Up(conf.nBanks) - 1, conf.bankIdLsb)
   }
 
   (outmemsys.io.tiles :+ outmemsys.io.htif).zip(io.tiles :+ htif.io.mem).zipWithIndex.map { 
     case ((outer, client), i) => 
-      outer.acquire <> TileLinkHeaderAppender(client.acquire, i, conf.nBanks, convertAddrToBank _)
+      outer.acquire <> Queue(TileLinkHeaderOverwriter(client.acquire, i, conf.nBanks, convertAddrToBank _))
-      outer.release <> TileLinkHeaderAppender(client.release, i, conf.nBanks, convertAddrToBank _)
+      outer.release <> Queue(TileLinkHeaderOverwriter(client.release, i, conf.nBanks, convertAddrToBank _))
-
+      outer.grant_ack <> Queue(TileLinkHeaderOverwriter(client.release, i))
-      val grant_ack_q = Queue(client.grant_ack)
-      outer.grant_ack.valid := grant_ack_q.valid
-      outer.grant_ack.bits := grant_ack_q.bits
-      outer.grant_ack.bits.header.src := UInt(i)
-      grant_ack_q.ready := outer.grant_ack.ready
-
       client.grant <> Queue(outer.grant, 1, pipe = true)
       client.probe <> Queue(outer.probe)
   }
@@ -77,26 +73,32 @@ class FPGAUncore(htif_width: Int)(implicit conf: UncoreConfiguration) extends Mo
   htif.io.host.in <> io.host.in
 }
 
-class FPGATopIO(htifWidth: Int) extends TopIO(htifWidth)
+class ReferenceChip(htif_width: Int)(implicit mif: MemoryIFConfiguration) extends Module {
-
+  val io = new Bundle {
-class FPGATop extends Module {
+    val host_in = new DecoupledIO(new HostPacket(htif_width)).flip()
-  val htif_width = 16
+    val host_out = new DecoupledIO(new HostPacket(htif_width))
+    val host_clk = Bool(OUTPUT)
+    val host_clk_edge = Bool(OUTPUT)
+    val host_debug_stats_pcr = Bool(OUTPUT)
+    val mem_req_cmd = new DecoupledIO(new MemReqCmd())
+    val mem_req_data = new DecoupledIO(new MemData())
+    val mem_resp = (new DecoupledIO(new MemResp())).flip()
+  }
+  
   val co = new MESICoherence
   val ntiles = 1
   val nbanks = 1
   val nmshrs = 2
   implicit val ln = LogicalNetworkConfiguration(log2Up(ntiles)+1, nbanks, ntiles+1)
-  implicit val tl = TileLinkConfiguration(co, ln, log2Up(1+8), 2*log2Up(nmshrs*ntiles+1), MEM_DATA_BITS)
+  implicit val tl = TileLinkConfiguration(co, ln, log2Up(1+8), 2*log2Up(nmshrs*ntiles+1), CACHE_DATA_SIZE_IN_BYTES*8)
   implicit val l2 = L2CoherenceAgentConfiguration(tl, 1, 8)
-  implicit val uc = UncoreConfiguration(l2, tl, ntiles, nbanks, bankIdLsb = 5, nSCR = 64)
+  implicit val uc = UncoreConfiguration(l2, tl, mif, ntiles, nbanks, bankIdLsb = 5, nSCR = 64)
 
-  val ic = ICacheConfig(64, 1, ntlb = 4, nbtb = 4)
+  val ic = ICacheConfig(64, 1, ntlb = 4, nbtb = 4, tl = tl)
-  val dc = DCacheConfig(64, 1, ntlb = 4, nmshr = 2, nrpq = 16, nsdq = 17, states = co.nClientStates)
+  val dc = DCacheConfig(64, 1, ntlb = 4, nmshr = 2, nrpq = 16, nsdq = 17, tl = tl)
   val rc = RocketConfiguration(tl, ic, dc, fpu = None,
                                fastMulDiv = false)
 
-  val io = new FPGATopIO(htif_width)
-
   val resetSigs = Vec.fill(uc.nTiles){Bool()}
   val tileList = (0 until uc.nTiles).map(r => Module(new Tile(resetSignal = resetSigs(r))(rc)))
   val uncore = Module(new FPGAUncore(htif_width))
@@ -118,9 +120,93 @@ class FPGATop extends Module {
     hl.ipi_req <> Queue(tile.io.host.ipi_req)
     tile.io.host.ipi_rep <> Queue(hl.ipi_rep)
   }
+  
+  io.host_in.ready := uncore.io.host.in.ready
+  uncore.io.host.in.bits := io.host_in.bits.data
+  uncore.io.host.in.valid := io.host_in.valid
+  
+  uncore.io.host.out.ready := io.host_out.ready
+  io.host_out.bits.data := uncore.io.host.out.bits
+  io.host_out.valid := uncore.io.host.out.valid
+  
+  io.host_clk := uncore.io.host.clk
+  io.host_clk_edge := uncore.io.host.clk_edge
+  io.host_debug_stats_pcr := uncore.io.host.debug_stats_pcr
 
-  io.host <> uncore.io.host
+  io.mem_req_cmd <> uncore.io.mem.req_cmd
-  io.mem <> uncore.io.mem
+  io.mem_req_data <> uncore.io.mem.req_data
+  io.mem_resp <> uncore.io.mem.resp
+}
+
+
+class FPGATopIO(htifWidth: Int)(implicit conf: MemoryIFConfiguration) extends TopIO(htifWidth)(conf)
+
+class FPGATop extends Module {
+  val htif_width = 16
+  
+  implicit val mif = MemoryIFConfiguration(PADDR_BITS - OFFSET_BITS, 128, 5, 4)
+  val deviceWidth = ROW_WIDTH/mif.dataBits
+  implicit val mc = MemoryControllerConfiguration(deviceWidth, (if(deviceWidth == 4) 0 else log2Up(deviceWidth/4)), mif)
+
+  val io = new FPGATopIO(htif_width)
+  
+  val referenceChip = Module(new Fame1Wrapper(new ReferenceChip(htif_width)))
+  val dramModel = Module(new DRAMSystemWrapper())
+  //dram model parameters setup
+  dramModel.io.params.tRAS := UInt(4)
+  dramModel.io.params.tRCD := UInt(4)
+  dramModel.io.params.tRP := UInt(4)
+  dramModel.io.params.tCCD := UInt(4)
+  dramModel.io.params.tRTP := UInt(4)
+  dramModel.io.params.tWTR := UInt(4)
+  dramModel.io.params.tWR := UInt(4)
+  dramModel.io.params.tRRD := UInt(4)
+  
+  //host to reference chip connections
+  referenceChip.DecoupledIOs("host_in").host_valid := Bool(true)
+  referenceChip.DecoupledIOs("host_in").target.bits := io.host.in.bits
+  referenceChip.DecoupledIOs("host_in").target.valid := io.host.in.valid
+  io.host.in.ready := referenceChip.DecoupledIOs("host_in").host_ready && referenceChip.DecoupledIOs("host_in").target.ready
+
+  io.host.out.valid := referenceChip.DecoupledIOs("host_out").host_valid && referenceChip.DecoupledIOs("host_out").target.valid
+  io.host.out.bits := referenceChip.DecoupledIOs("host_out").target.bits
+  referenceChip.DecoupledIOs("host_out").target.ready := io.host.out.ready
+  referenceChip.DecoupledIOs("host_out").host_ready := Bool(true)
+  
+  io.host.clk := referenceChip.DebugIOs("host_clk")
+  io.host.clk_edge := referenceChip.DebugIOs("host_clk_edge")
+  io.host.debug_stats_pcr := referenceChip.DebugIOs("host_debug_stats_pcr")
+
+  //reference chip to dram model connections
+  val mem_req_cmd_queue = Module(new FameQueue(8)(new MemReqCmd()))
+  val mem_req_data_queue = Module(new FameQueue(8)(new MemData()))
+  val mem_resp_queue = Module(new FameQueue(8)(new MemResp()))
+  
+  //cmd queue
+  FameDecoupledIO.connect(referenceChip.DecoupledIOs("mem_req_cmd"),  mem_req_cmd_queue.io.enq, new MemReqCmd)
+  mem_req_cmd_queue.io.deq <> dramModel.io.memReqCmd
+  
+  //data queue
+  FameDecoupledIO.connect(referenceChip.DecoupledIOs("mem_req_data"), mem_req_data_queue.io.enq, new MemData)
+  mem_req_data_queue.io.deq <> dramModel.io.memReqData
+  
+  //resp queue
+  mem_resp_queue.io.enq <> dramModel.io.memResp
+  FameDecoupledIO.connect(referenceChip.DecoupledIOs("mem_resp"), mem_resp_queue.io.deq, new MemResp)
+  
+  //dram model to outside memory connections
+  val host_mem_cmd_queue = Module(new Queue(new MemReqCmd, 2))
+  val host_mem_data_queue = Module(new Queue(new MemData, mif.dataBeats))
+  val host_mem_resp_queue = Module(new Queue(new MemResp, mif.dataBeats))
+  
+  host_mem_cmd_queue.io.enq <> dramModel.io.mem.req_cmd
+  host_mem_cmd_queue.io.deq <> io.mem.req_cmd
+  
+  host_mem_data_queue.io.enq <> dramModel.io.mem.req_data
+  host_mem_data_queue.io.deq <> io.mem.req_data
+  
+  host_mem_resp_queue.io.enq <> io.mem.resp
+  host_mem_resp_queue.io.deq <> dramModel.io.mem.resp
 }
 
 abstract class AXISlave extends Module {
@@ -174,7 +260,7 @@ class Slave extends AXISlave
 
   // write cr1 -> mem.resp (nonblocking)
   val in_count = Reg(init=UInt(0, log2Up(memw/dw)))
-  val rf_count = Reg(init=UInt(0, log2Up(REFILL_CYCLES)))
+  val rf_count = Reg(init=UInt(0, log2Up(CACHE_DATA_SIZE_IN_BYTES*8/memw)))
   require(memw % dw == 0 && isPow2(memw/dw))
   val in_reg = Reg(top.io.mem.resp.bits.data)
   top.io.mem.resp.bits.data := Cat(io.in.bits, in_reg(in_reg.getWidth-1,dw))

src/main/scala/network.scala

@@ -5,115 +5,42 @@ import uncore._
 import scala.reflect._
 import scala.reflect.runtime.universe._
 
-object TileLinkHeaderAppender {
+object TileLinkHeaderOverwriter {
-  def apply[T <: ClientSourcedMessage with HasPhysicalAddress, U <: ClientSourcedMessage with HasTileLinkData](in: PairedDataIO[LogicalNetworkIO[T],LogicalNetworkIO[U]], clientId: Int, nBanks: Int, addrConvert: Bits => UInt)(implicit conf: TileLinkConfiguration) = {
+  def apply[T <: ClientSourcedMessage](in: DecoupledIO[LogicalNetworkIO[T]], clientId: Int)(implicit conf: TileLinkConfiguration): DecoupledIO[LogicalNetworkIO[T]] = {
-    val shim = Module(new TileLinkHeaderAppender(in.meta.bits.payload, in.data.bits.payload, clientId, nBanks, addrConvert))
+    val out = in.clone.asDirectionless
-    shim.io.in <> in
+    out.bits.payload := in.bits.payload
-    shim.io.out
+    out.bits.header.src := UInt(clientId)
+    out.bits.header.dst := in.bits.header.dst
+    out.valid := in.valid
+    in.ready := out.ready
+    out
   }
-  def apply[T <: ClientSourcedMessage with HasPhysicalAddress](in: DecoupledIO[LogicalNetworkIO[T]], clientId: Int, nBanks: Int, addrConvert: Bits => UInt)(implicit conf: TileLinkConfiguration) = {
+  def apply[T <: ClientSourcedMessage with HasPhysicalAddress](in: DecoupledIO[LogicalNetworkIO[T]], clientId: Int, nBanks: Int, addrConvert: UInt => UInt)(implicit conf: TileLinkConfiguration): DecoupledIO[LogicalNetworkIO[T]] = {
-    val shim = Module(new TileLinkHeaderAppender(in.bits.payload.clone, new AcquireData, clientId, nBanks, addrConvert))
+    val out: DecoupledIO[LogicalNetworkIO[T]] = apply(in, clientId)
-    shim.io.in.meta <> in
+    out.bits.header.dst := (if(nBanks > 1) addrConvert(in.bits.payload.addr) else UInt(0))
-    shim.io.out.meta
+    out
   }
 }
 
-class TileLinkHeaderAppender[T <: ClientSourcedMessage with HasPhysicalAddress, U <: ClientSourcedMessage with HasTileLinkData](mType: T, dType: U, clientId: Int, nBanks: Int, addrConvert: Bits => UInt)(implicit conf: TileLinkConfiguration) extends Module {
-  implicit val ln = conf.ln
-  val io = new Bundle {
-    val in = new PairedDataIO(new LogicalNetworkIO(mType), new LogicalNetworkIO(dType)).flip
-    val out = new PairedDataIO(new LogicalNetworkIO(mType), new LogicalNetworkIO(dType))
-  }
-
-  val meta_q = Queue(io.in.meta)
-  val data_q = Queue(io.in.data)
-  if(nBanks == 1) {
-    io.out.meta.bits.payload := meta_q.bits.payload
-    io.out.meta.bits.header.src := UInt(clientId)
-    io.out.meta.bits.header.dst := UInt(0)
-    io.out.meta.valid := meta_q.valid
-    meta_q.ready := io.out.meta.ready
-    io.out.data.bits.payload := data_q.bits.payload
-    io.out.data.bits.header.src := UInt(clientId)
-    io.out.data.bits.header.dst := UInt(0)
-    io.out.data.valid := data_q.valid
-    data_q.ready := io.out.data.ready
-  } else {
-    val meta_has_data = conf.co.messageHasData(meta_q.bits.payload)
-    val addr_q = Module(new Queue(io.in.meta.bits.payload.addr.clone, 2, pipe = true, flow = true))
-    val data_cnt = Reg(init=UInt(0, width = log2Up(REFILL_CYCLES)))
-    val data_cnt_up = data_cnt + UInt(1)
-
-    io.out.meta.bits.payload := meta_q.bits.payload
-    io.out.meta.bits.header.src := UInt(clientId)
-    io.out.meta.bits.header.dst := addrConvert(meta_q.bits.payload.addr)
-    io.out.data.bits.payload := meta_q.bits.payload
-    io.out.data.bits.header.src := UInt(clientId)
-    io.out.data.bits.header.dst := addrConvert(addr_q.io.deq.bits)
-    addr_q.io.enq.bits := meta_q.bits.payload.addr
-
-    io.out.meta.valid := meta_q.valid && addr_q.io.enq.ready
-    meta_q.ready := io.out.meta.ready && addr_q.io.enq.ready
-    io.out.data.valid := data_q.valid && addr_q.io.deq.valid
-    data_q.ready := io.out.data.ready && addr_q.io.deq.valid
-    addr_q.io.enq.valid := meta_q.valid && io.out.meta.ready && meta_has_data
-    addr_q.io.deq.ready := Bool(false)
-
-    when(data_q.valid && data_q.ready) {
-      data_cnt := data_cnt_up
-      when(data_cnt_up === UInt(0)) {
-        addr_q.io.deq.ready := Bool(true)
-      }
-    }
-  }
-}
-
-//Adapter betweewn an UncachedTileLinkIO and a mem controller MemIO
-class MemIOUncachedTileLinkIOConverter(qDepth: Int)(implicit conf: TileLinkConfiguration) extends Module {
-  val io = new Bundle {
-    val uncached = new UncachedTileLinkIO().flip
-    val mem = new ioMem
-  }
-  val mem_cmd_q = Module(new Queue(new MemReqCmd, qDepth))
-  val mem_data_q = Module(new Queue(new MemData, qDepth))
-  mem_cmd_q.io.enq.valid := io.uncached.acquire.meta.valid
-  io.uncached.acquire.meta.ready := mem_cmd_q.io.enq.ready 
-  mem_cmd_q.io.enq.bits.rw := conf.co.needsOuterWrite(io.uncached.acquire.meta.bits.payload.a_type, UInt(0))
-  mem_cmd_q.io.enq.bits.tag := io.uncached.acquire.meta.bits.payload.client_xact_id
-  mem_cmd_q.io.enq.bits.addr := io.uncached.acquire.meta.bits.payload.addr
-  mem_data_q.io.enq.valid := io.uncached.acquire.data.valid
-  io.uncached.acquire.data.ready := mem_data_q.io.enq.ready
-  mem_data_q.io.enq.bits.data := io.uncached.acquire.data.bits.payload.data 
-  io.uncached.grant.valid := io.mem.resp.valid
-  io.mem.resp.ready := io.uncached.grant.ready
-  io.uncached.grant.bits.payload.data := io.mem.resp.bits.data
-  io.uncached.grant.bits.payload.client_xact_id := io.mem.resp.bits.tag
-  io.uncached.grant.bits.payload.master_xact_id := UInt(0) // DNC
-  io.uncached.grant.bits.payload.g_type := UInt(0) // DNC
-  io.mem.req_cmd <> mem_cmd_q.io.deq
-  io.mem.req_data <> mem_data_q.io.deq
-}
-
 class ReferenceChipCrossbarNetwork(implicit conf: UncoreConfiguration) extends LogicalNetwork[TileLinkIO]()(conf.tl.ln) {
   implicit val (tl, ln, co) = (conf.tl, conf.tl.ln, conf.tl.co)
   val io = new Bundle {
     val clients = Vec.fill(ln.nClients){(new TileLinkIO).flip}
     val masters = Vec.fill(ln.nMasters){new TileLinkIO}
   }
+
   implicit val pconf = new PhysicalNetworkConfiguration(ln.nEndpoints, ln.idBits) // Same config for all networks
 
   // Actually instantiate the particular networks required for TileLink
-  val acqNet = Module(new PairedCrossbar(new Acquire, new AcquireData, REFILL_CYCLES, (acq: PhysicalNetworkIO[Acquire]) => co.messageHasData(acq.payload)))
+  val acqNet = Module(new BasicCrossbar(new Acquire))
-  val relNet = Module(new PairedCrossbar(new Release, new ReleaseData, REFILL_CYCLES, (rel: PhysicalNetworkIO[Release]) => co.messageHasData(rel.payload)))
+  val relNet = Module(new BasicCrossbar(new Release))
-  val probeNet = Module(new BasicCrossbar(new Probe))
+  val prbNet = Module(new BasicCrossbar(new Probe))
-  val grantNet = Module(new BasicCrossbar(new Grant))
+  val gntNet = Module(new BasicCrossbar(new Grant))
   val ackNet = Module(new BasicCrossbar(new GrantAck))
 
   // Aliases for the various network IO bundle types
   type FBCIO[T <: Data] = DecoupledIO[PhysicalNetworkIO[T]]
   type FLNIO[T <: Data] = DecoupledIO[LogicalNetworkIO[T]]
-  type PBCIO[M <: Data, D <: Data] = PairedDataIO[PhysicalNetworkIO[M], PhysicalNetworkIO[D]]
-  type PLNIO[M <: Data, D <: Data] = PairedDataIO[LogicalNetworkIO[M], LogicalNetworkIO[D]]
   type FromCrossbar[T <: Data] = FBCIO[T] => FLNIO[T]
   type ToCrossbar[T <: Data] = FLNIO[T] => FBCIO[T]
 
@@ -176,10 +103,10 @@ class ReferenceChipCrossbarNetwork(implicit conf: UncoreConfiguration) extends L
 
   def doFIFOHookup[T <: Data](isEndpointSourceOfMessage: Boolean, physIn: FBCIO[T], physOut: FBCIO[T], logIO: FLNIO[T], inShim: ToCrossbar[T], outShim: FromCrossbar[T]) = {
     if(isEndpointSourceOfMessage) doFIFOInputHookup(physIn, physOut, logIO, inShim)
-    else                 doFIFOOutputHookup(physIn, physOut, logIO, outShim)
+    else doFIFOOutputHookup(physIn, physOut, logIO, outShim)
   }
     
-  //Hookup all instances of a particular subbundle of 
+  //Hookup all instances of a particular subbundle of TileLink
   def doFIFOHookups[T <: Data: TypeTag](physIO: BasicCrossbarIO[T], getLogIO: TileLinkIO => FLNIO[T]) = {
     typeTag[T].tpe match{ 
       case t if t <:< typeTag[ClientSourcedMessage].tpe => {
@@ -194,33 +121,9 @@ class ReferenceChipCrossbarNetwork(implicit conf: UncoreConfiguration) extends L
     }
   }
 
-  def doPairedDataHookup[T <: Data, R <: Data](isEndpointSourceOfMessage: Boolean, physIn: PBCIO[T,R], physOut: PBCIO[T,R], logIO: PLNIO[T,R], inShim: ToCrossbar[T], outShim: FromCrossbar[T], inShimD: ToCrossbar[R], outShimD: FromCrossbar[R]) = {
+  doFIFOHookups(acqNet.io, (tl: TileLinkIO) => tl.acquire)
-    if(isEndpointSourceOfMessage) {
+  doFIFOHookups(relNet.io, (tl: TileLinkIO) => tl.release)
-      doFIFOInputHookup[T](physIn.meta, physOut.meta, logIO.meta, inShim)
+  doFIFOHookups(prbNet.io, (tl: TileLinkIO) => tl.probe)
-      doFIFOInputHookup[R](physIn.data, physOut.data, logIO.data, inShimD)
+  doFIFOHookups(gntNet.io, (tl: TileLinkIO) => tl.grant)
-    } else {
-      doFIFOOutputHookup[T](physIn.meta, physOut.meta, logIO.meta, outShim)
-      doFIFOOutputHookup[R](physIn.data, physOut.data, logIO.data, outShimD)
-    }
-  }
-
-  def doPairedDataHookups[T <: Data: TypeTag, R <: Data](physIO: PairedCrossbarIO[T,R], getLogIO: TileLinkIO => PLNIO[T,R]) = {
-    typeTag[T].tpe match{ 
-      case t if t <:< typeTag[ClientSourcedMessage].tpe => {
-        io.masters.zipWithIndex.map{ case (i, id) => doPairedDataHookup[T,R](false, physIO.in(id), physIO.out(id), getLogIO(i), ClientToCrossbarShim, CrossbarToMasterShim, ClientToCrossbarShim, CrossbarToMasterShim) }
-        io.clients.zipWithIndex.map{ case (i, id) => doPairedDataHookup[T,R](true, physIO.in(id+ln.nMasters), physIO.out(id+ln.nMasters), getLogIO(i), ClientToCrossbarShim, CrossbarToMasterShim, ClientToCrossbarShim, CrossbarToMasterShim) }
-      }
-      case t if t <:< typeTag[MasterSourcedMessage].tpe => {
-        io.masters.zipWithIndex.map{ case (i, id) => doPairedDataHookup[T,R](true, physIO.in(id), physIO.out(id), getLogIO(i), MasterToCrossbarShim, CrossbarToClientShim, MasterToCrossbarShim, CrossbarToClientShim) }
-        io.clients.zipWithIndex.map{ case (i, id) => doPairedDataHookup[T,R](false, physIO.in(id+ln.nMasters), physIO.out(id+ln.nMasters), getLogIO(i), MasterToCrossbarShim, CrossbarToClientShim, MasterToCrossbarShim, CrossbarToClientShim) }
-      }
-      case _ => require(false, "Unknown message sourcing.")
-    }
-  }
-
-  doPairedDataHookups(acqNet.io, (tl: TileLinkIO) => tl.acquire)
-  doPairedDataHookups(relNet.io, (tl: TileLinkIO) => tl.release)
-  doFIFOHookups(probeNet.io, (tl: TileLinkIO) => tl.probe)
-  doFIFOHookups(grantNet.io, (tl: TileLinkIO) => tl.grant)
   doFIFOHookups(ackNet.io, (tl: TileLinkIO) => tl.grant_ack)
 }