diff --git a/rocket/src/main/scala/arbiter.scala b/rocket/src/main/scala/arbiter.scala
index 4a60eb10..da5495e6 100644
--- a/rocket/src/main/scala/arbiter.scala
+++ b/rocket/src/main/scala/arbiter.scala
@@ -24,6 +24,7 @@ class ioMemArbiter extends Bundle() {
 //   val icache = new ioICache();
   val icache = new ioIPrefetcherMem().flip();
   val vicache = new ioICache();
+  val htif = new ioDCache();
 }
 
 class rocketMemArbiter extends Component {
@@ -34,26 +35,31 @@ class rocketMemArbiter extends Component {
   // *****************************
 
   // Memory request is valid if either icache or dcache have a valid request
-  io.mem.req_val := (io.icache.req_val || io.vicache.req_val || io.dcache.req_val);
+  io.mem.req_val := io.icache.req_val || io.vicache.req_val || io.dcache.req_val || io.htif.req_val
 
-  // Set read/write bit.  ICache always reads
-  io.mem.req_rw := Mux(io.dcache.req_val, io.dcache.req_rw, Bool(false));
+  // Set read/write bit.  I$ always reads
+  io.mem.req_rw :=
+    Mux(io.dcache.req_val, io.dcache.req_rw,
+    Mux(io.icache.req_val, Bool(false),
+    Mux(io.vicache.req_val, Bool(false),
+    io.htif.req_rw)))
 
-  // Give priority to ICache
+  // Give priority to D$
   io.mem.req_addr :=
     Mux(io.dcache.req_val, io.dcache.req_addr,
     Mux(io.icache.req_val, io.icache.req_addr,
-        io.vicache.req_addr))
+    Mux(io.vicache.req_val, io.vicache.req_addr,
+    io.htif.req_addr)))
+
+  io.mem.req_wdata := Mux(io.dcache.req_val, io.dcache.req_wdata, io.htif.req_wdata)
 
   // low bit of tag to indicate D$, I$, and VI$
-  val t_dcache :: t_icache :: t_vicache :: Nil = Enum(3){ UFix() }
+  val t_dcache :: t_icache :: t_vicache :: t_htif :: Nil = Enum(4){ UFix() }
   io.mem.req_tag :=
     Mux(io.dcache.req_val, Cat(io.dcache.req_tag, t_dcache),
     Mux(io.icache.req_val, Cat(io.icache.req_tag, t_icache),
-        Cat(Bits(0, MEM_TAG_BITS-2), t_vicache)))
-
-  // Just pass through write data (only D$ will write)
-  io.mem.req_wdata := io.dcache.req_wdata;
+    Mux(io.vicache.req_val, t_vicache,
+    t_htif)))
 
   // *****************************
   // Interface to caches
@@ -64,16 +70,19 @@ class rocketMemArbiter extends Component {
   io.dcache.req_rdy := io.mem.req_rdy;
   io.icache.req_rdy := io.mem.req_rdy && !io.dcache.req_val;
   io.vicache.req_rdy := io.mem.req_rdy && !io.dcache.req_val && !io.icache.req_val
+  io.htif.req_rdy := io.mem.req_rdy && !io.dcache.req_val && !io.icache.req_val && !io.vicache.req_val
 
   // Response will only be valid for D$ or I$ not both because of tag bits
   io.dcache.resp_val := io.mem.resp_val && (io.mem.resp_tag(1,0) === t_dcache)
   io.icache.resp_val := io.mem.resp_val && (io.mem.resp_tag(1,0) === t_icache)
   io.vicache.resp_val := io.mem.resp_val && (io.mem.resp_tag(1,0) === t_vicache)
+  io.htif.resp_val := io.mem.resp_val && (io.mem.resp_tag(1,0) === t_htif)
 
   // Feed through data to both 
   io.dcache.resp_data := io.mem.resp_data;
   io.icache.resp_data := io.mem.resp_data;
   io.vicache.resp_data := io.mem.resp_data
+  io.htif.resp_data := io.mem.resp_data
   
   io.dcache.resp_tag := io.mem.resp_tag >> UFix(2)
   io.icache.resp_tag := io.mem.resp_tag >> UFix(2)
diff --git a/rocket/src/main/scala/consts.scala b/rocket/src/main/scala/consts.scala
index 34d3a70c..e00e51cb 100644
--- a/rocket/src/main/scala/consts.scala
+++ b/rocket/src/main/scala/consts.scala
@@ -138,7 +138,6 @@ object Constants
   val PCR_TOHOST   = UFix(16, 5);
   val PCR_FROMHOST = UFix(17, 5);
   val PCR_VECBANK  = UFix(18, 5);
-  val PCR_CONSOLE  = UFix(19, 5);
 
   // definition of bits in PCR status reg
   val SR_ET   = 0;  // enable traps
diff --git a/rocket/src/main/scala/cpu.scala b/rocket/src/main/scala/cpu.scala
index 96b33035..c8ac9643 100644
--- a/rocket/src/main/scala/cpu.scala
+++ b/rocket/src/main/scala/cpu.scala
@@ -10,25 +10,10 @@ class ioDebug(view: List[String] = null) extends Bundle(view)
   val error_mode  = Bool(OUTPUT);
 }
 
-class ioHost(view: List[String] = null) extends Bundle(view)
-{
-  val from_wen   = Bool(INPUT);
-  val from       = Bits(64, INPUT);
-  val to         = Bits(64, OUTPUT);
-}
-
-class ioConsole(view: List[String] = null) extends Bundle(view)
-{
-  val rdy   = Bool(INPUT);
-  val valid = Bool(OUTPUT);
-  val bits  = Bits(8, OUTPUT);
-}
-
 class ioRocket extends Bundle()
 {
   val debug   = new ioDebug();
-  val console = new ioConsole();
-  val host    = new ioHost();
+  val host    = new ioHTIF();
   val imem    = new ioImem().flip();
   val vimem   = new ioImem().flip();
   val dmem    = new ioDmem().flip();
@@ -47,6 +32,7 @@ class rocketProc extends Component
   val ptw   = new rocketPTW();
   val arb   = new rocketDmemArbiter();
 
+  ctrl.io.htif_reset := io.host.reset
   ctrl.io.dpath             <> dpath.io.ctrl;
   dpath.io.host             <> io.host;
   dpath.io.debug            <> io.debug;
@@ -112,10 +98,6 @@ class rocketProc extends Component
   dpath.io.dmem.resp_data := arb.io.cpu.resp_data;  
   dpath.io.dmem.resp_data_subword := io.dmem.resp_data_subword;
 
-  io.console.bits     := dpath.io.console.bits;
-  io.console.valid    := dpath.io.console.valid;
-  ctrl.io.console.rdy := io.console.rdy;
-
   if (HAVE_FPU)
   {
     val fpu = new rocketFPU(4,6)
diff --git a/rocket/src/main/scala/ctrl.scala b/rocket/src/main/scala/ctrl.scala
index 12228da9..4bf8f4cc 100644
--- a/rocket/src/main/scala/ctrl.scala
+++ b/rocket/src/main/scala/ctrl.scala
@@ -77,8 +77,8 @@ class ioCtrlDpath extends Bundle()
 
 class ioCtrlAll extends Bundle()
 {
+  val htif_reset = Bool(INPUT)
   val dpath   = new ioCtrlDpath();
-  val console = new ioConsole(List("rdy"));
   val imem    = new ioImem(List("req_val", "resp_val")).flip();
   val dmem    = new ioDmem(List("req_val", "req_kill", "req_rdy", "req_cmd", "req_type", "resp_miss", "resp_nack")).flip();
   val ext_mem = new ioDmem(List("req_val", "req_cmd", "req_type", "resp_nack"))
@@ -298,8 +298,6 @@ class rocketCtrl extends Component
   val id_raddr1 = io.dpath.inst(26,22);
   val id_waddr  = Mux(id_sel_wa === WA_RA, RA, io.dpath.inst(31,27));
 
-  val id_console_out_val  = id_wen_pcr.toBool && (id_raddr2 === PCR_CONSOLE);
-
   val wb_reg_div_mul_val = Reg(resetVal = Bool(false))
   val wb_reg_dcache_miss = Reg(io.dmem.resp_miss, resetVal = Bool(false));
 
@@ -699,7 +697,7 @@ class rocketCtrl extends Component
   io.dpath.wen_btb := !ex_reg_btb_hit && br_taken
   io.dpath.clr_btb := ex_reg_btb_hit && !br_taken || id_reg_icmiss;
   
-  io.imem.req_val  := take_pc_wb || !mem_reg_replay && !ex_reg_replay && (take_pc_ex || !id_reg_replay)
+  io.imem.req_val  := !io.htif_reset && (take_pc_wb || !mem_reg_replay && !ex_reg_replay && (take_pc_ex || !id_reg_replay))
 
   // stall for RAW/WAW hazards on loads, AMOs, and mul/div in execute stage.
   val data_hazard_ex = ex_reg_wen &&
@@ -753,8 +751,7 @@ class rocketCtrl extends Component
       id_mem_val.toBool && !(io.dmem.req_rdy && io.dtlb_rdy) ||
       id_vec_val.toBool && !(io.vec_iface.vcmdq_ready && io.vec_iface.vximm1q_ready && io.vec_iface.vximm2q_ready) || // being conservative
       ((id_sync === SYNC_D) || (id_sync === SYNC_I)) && !io.dmem.req_rdy ||
-      vec_cpfence ||
-      id_console_out_val && !io.console.rdy
+      vec_cpfence
     );
   val ctrl_stallf = ctrl_stalld;
     
diff --git a/rocket/src/main/scala/dpath.scala b/rocket/src/main/scala/dpath.scala
index 2d5a7c85..9b167c5e 100644
--- a/rocket/src/main/scala/dpath.scala
+++ b/rocket/src/main/scala/dpath.scala
@@ -29,9 +29,8 @@ class ioDpathImem extends Bundle()
 
 class ioDpathAll extends Bundle()
 {
-  val host  = new ioHost();
+  val host  = new ioHTIF();
   val ctrl  = new ioCtrlDpath().flip();
-  val console = new ioConsole(List("valid","bits"));
   val debug = new ioDebug();
   val dmem  = new ioDpathDmem();
   val ext_mem = new ioDmem(List("req_val", "req_idx", "req_ppn", "req_data", "req_tag", "resp_val", "resp_data", "resp_tag"))
@@ -313,9 +312,7 @@ class rocketDpath extends Component
   	Mux(ex_reg_ctrl_eret, PCR_EPC, 
   		ex_reg_raddr2);
 
-  pcr.io.host.from_wen <> io.host.from_wen;
-  pcr.io.host.from     <> io.host.from;
-  pcr.io.host.to       <> io.host.to;
+  pcr.io.host <> io.host
 
   io.ctrl.irq_timer    := pcr.io.irq_timer;
   io.ctrl.irq_ipi      := pcr.io.irq_ipi;  
@@ -483,8 +480,6 @@ class rocketDpath extends Component
   pcr.io.cause 			  := io.ctrl.cause;
   pcr.io.pc					  := wb_reg_pc;
   pcr.io.badvaddr_wen := io.ctrl.badvaddr_wen;
-  io.console.bits := pcr.io.console_data;
-  io.console.valid := pcr.io.console_val;
 }
 
 }
diff --git a/rocket/src/main/scala/dpath_util.scala b/rocket/src/main/scala/dpath_util.scala
index cc89def5..62aab094 100644
--- a/rocket/src/main/scala/dpath_util.scala
+++ b/rocket/src/main/scala/dpath_util.scala
@@ -60,7 +60,7 @@ class rocketDpathBTB(entries: Int) extends Component
 
 class ioDpathPCR extends Bundle()
 {
-  val host  = new ioHost(List("from", "from_wen", "to"));
+  val host  = new ioHTIF()
   val debug = new ioDebug(List("error_mode", "log_control"));
   val r     = new ioReadPort();
   val w     = new ioWritePort();
@@ -78,8 +78,6 @@ class ioDpathPCR extends Bundle()
   val ptbr_wen  = Bool(OUTPUT);
   val irq_timer = Bool(OUTPUT);
   val irq_ipi   = Bool(OUTPUT);
-  val console_data = Bits(8, OUTPUT);
-  val console_val  = Bool(OUTPUT);
   val vecbank   = Bits(8, OUTPUT)
   val vecbankcnt = UFix(4, OUTPUT)
 }
@@ -119,11 +117,19 @@ class rocketDpathPCR extends Component
   val reg_status = Cat(reg_status_sx, reg_status_ux, reg_status_s, reg_status_ps, reg_status_ec, reg_status_ev, reg_status_ef, reg_status_et);
   val rdata = Wire() { Bits() };
 
-  io.ptbr_wen           := reg_status_vm.toBool && io.w.en && (io.w.addr === PCR_PTBR);
+  val ren = io.r.en || io.host.pcr_ren
+  val raddr = Mux(io.r.en, io.r.addr, io.host.pcr_addr)
+  io.host.pcr_rdata := rdata
+
+  val wen = io.w.en || io.host.pcr_wen
+  val waddr = Mux(io.w.en, io.w.addr, io.host.pcr_addr)
+  val wdata = Mux(io.w.en, io.w.data, io.host.pcr_wdata)
+  io.host.pcr_rdy := Mux(io.host.pcr_wen, !io.w.en, !io.r.en)
+
+  io.ptbr_wen           := reg_status_vm.toBool && wen && (waddr === PCR_PTBR);
   io.status              := Cat(reg_status_vm, reg_status_im, reg_status);
   io.evec               := reg_ebase;
   io.ptbr               := reg_ptbr;
-  io.host.to             := Mux(io.host.from_wen, Bits(0), reg_tohost);
   io.debug.error_mode    := reg_error_mode;
   io.r.data := rdata;
 
@@ -133,19 +139,6 @@ class rocketDpathPCR extends Component
     cnt = cnt + reg_vecbank(i)
   io.vecbankcnt := cnt(3,0)
 
-  val console_wen = !io.exception && io.w.en && (io.w.addr === PCR_CONSOLE);
-  io.console_data := Mux(console_wen, io.w.data(7,0), Bits(0,8));
-  io.console_val := console_wen;
-
-  when (io.host.from_wen) {
-    reg_tohost   := Bits(0);
-    reg_fromhost := io.host.from;
-  }
-  .elsewhen (io.w.en && (io.w.addr === PCR_TOHOST)) {
-    reg_tohost   := io.w.data;
-    reg_fromhost := Bits(0);
-  }
-
   val badvaddr_sign = Mux(io.w.data(VADDR_BITS-1), ~io.w.data(63,VADDR_BITS) === UFix(0), io.w.data(63,VADDR_BITS) != UFix(0))
   when (io.badvaddr_wen) {
     reg_badvaddr     := Cat(badvaddr_sign, io.w.data(VADDR_BITS-1,0)).toUFix;
@@ -185,37 +178,38 @@ class rocketDpathPCR extends Component
   io.irq_timer := r_irq_timer;
   io.irq_ipi := r_irq_ipi;
 
-  when (io.w.en) {
-    when (io.w.addr === PCR_STATUS) {
-      reg_status_vm := io.w.data(SR_VM).toBool;
-      reg_status_im := io.w.data(15,8);
-      reg_status_sx := io.w.data(SR_SX).toBool;
-      reg_status_ux := io.w.data(SR_UX).toBool;
-      reg_status_s  := io.w.data(SR_S).toBool;
-      reg_status_ps := io.w.data(SR_PS).toBool;
-      reg_status_ev := Bool(HAVE_VEC) && io.w.data(SR_EV).toBool;
-      reg_status_ef := Bool(HAVE_FPU) && io.w.data(SR_EF).toBool;
-      reg_status_ec := Bool(HAVE_RVC) && io.w.data(SR_EC).toBool;
-      reg_status_et := io.w.data(SR_ET).toBool;
+  when (wen) {
+    when (waddr === PCR_STATUS) {
+      reg_status_vm := wdata(SR_VM).toBool;
+      reg_status_im := wdata(15,8);
+      reg_status_sx := wdata(SR_SX).toBool;
+      reg_status_ux := wdata(SR_UX).toBool;
+      reg_status_s  := wdata(SR_S).toBool;
+      reg_status_ps := wdata(SR_PS).toBool;
+      reg_status_ev := Bool(HAVE_VEC) && wdata(SR_EV).toBool;
+      reg_status_ef := Bool(HAVE_FPU) && wdata(SR_EF).toBool;
+      reg_status_ec := Bool(HAVE_RVC) && wdata(SR_EC).toBool;
+      reg_status_et := wdata(SR_ET).toBool;
     }
-    when (io.w.addr === PCR_EPC)      { reg_epc          := io.w.data(VADDR_BITS,0).toUFix; }
-    when (io.w.addr === PCR_BADVADDR) { reg_badvaddr     := io.w.data(VADDR_BITS,0).toUFix; }
-    when (io.w.addr === PCR_EVEC)     { reg_ebase       := io.w.data(VADDR_BITS-1,0).toUFix; }
-    when (io.w.addr === PCR_COUNT)    { reg_count       := io.w.data(31,0).toUFix; }
-    when (io.w.addr === PCR_COMPARE)  { reg_compare     := io.w.data(31,0).toUFix; r_irq_timer := Bool(false); }
-    when (io.w.addr === PCR_CAUSE)    { reg_cause       := io.w.data(4,0); }
-    when (io.w.addr === PCR_FROMHOST) { reg_fromhost     := io.w.data; }
-    when (io.w.addr === PCR_SEND_IPI) { r_irq_ipi        := Bool(true); }
-    when (io.w.addr === PCR_CLR_IPI)  { r_irq_ipi        := Bool(false); }
-    when (io.w.addr === PCR_K0)       { reg_k0          := io.w.data; }
-    when (io.w.addr === PCR_K1)       { reg_k1          := io.w.data; }
-    when (io.w.addr === PCR_PTBR)     { reg_ptbr        := Cat(io.w.data(PADDR_BITS-1, PGIDX_BITS), Bits(0, PGIDX_BITS)).toUFix; }
-    when (io.w.addr === PCR_VECBANK)  { reg_vecbank     := io.w.data(7,0) }
+    when (waddr === PCR_EPC)      { reg_epc          := wdata(VADDR_BITS,0).toUFix; }
+    when (waddr === PCR_BADVADDR) { reg_badvaddr     := wdata(VADDR_BITS,0).toUFix; }
+    when (waddr === PCR_EVEC)     { reg_ebase       := wdata(VADDR_BITS-1,0).toUFix; }
+    when (waddr === PCR_COUNT)    { reg_count       := wdata(31,0).toUFix; }
+    when (waddr === PCR_COMPARE)  { reg_compare     := wdata(31,0).toUFix; r_irq_timer := Bool(false); }
+    when (waddr === PCR_CAUSE)    { reg_cause       := wdata(4,0); }
+    when (waddr === PCR_TOHOST)   { reg_tohost       := wdata; reg_fromhost := Bits(0) }
+    when (waddr === PCR_FROMHOST) { reg_fromhost     := wdata; reg_tohost := Bits(0) }
+    when (waddr === PCR_SEND_IPI) { r_irq_ipi        := Bool(true); }
+    when (waddr === PCR_CLR_IPI)  { r_irq_ipi        := Bool(false); }
+    when (waddr === PCR_K0)       { reg_k0          := wdata; }
+    when (waddr === PCR_K1)       { reg_k1          := wdata; }
+    when (waddr === PCR_PTBR)     { reg_ptbr        := Cat(wdata(PADDR_BITS-1, PGIDX_BITS), Bits(0, PGIDX_BITS)).toUFix; }
+    when (waddr === PCR_VECBANK)  { reg_vecbank     := wdata(7,0) }
   }
 
   rdata := Bits(0, 64)
-  when (io.r.en) {
-    switch (io.r.addr) {
+  when (ren) {
+    switch (raddr) {
       is (PCR_STATUS)   { rdata := Cat(Bits(0,47), reg_status_vm, reg_status_im, reg_status); }
       is (PCR_EPC)      { rdata := Cat(Fill(64-VADDR_BITS-1, reg_epc(VADDR_BITS)), reg_epc); }
       is (PCR_BADVADDR) { rdata := Cat(Fill(64-VADDR_BITS-1, reg_badvaddr(VADDR_BITS)), reg_badvaddr); }
diff --git a/rocket/src/main/scala/htif.scala b/rocket/src/main/scala/htif.scala
new file mode 100644
index 00000000..39cbdc90
--- /dev/null
+++ b/rocket/src/main/scala/htif.scala
@@ -0,0 +1,168 @@
+package Top
+
+import Chisel._
+import Node._;
+import Constants._;
+
+class ioHost(w: Int, view: List[String] = null) extends Bundle(view)
+{
+  val in = new ioDecoupled()(Bits(width = w))
+  val out = new ioDecoupled()(Bits(width = w)).flip()
+}
+
+class ioHTIF extends Bundle
+{
+  val reset = Bool(INPUT)
+  val pcr_wen = Bool(INPUT)
+  val pcr_ren = Bool(INPUT)
+  val pcr_rdy = Bool(OUTPUT)
+  val pcr_addr = Bits(5, INPUT)
+  val pcr_wdata = Bits(64, INPUT)
+  val pcr_rdata = Bits(64, OUTPUT)
+}
+
+class rocketHTIF(w: Int, ncores: Int) extends Component
+{
+  val io = new Bundle {
+    val host = new ioHost(w)
+    val cpu = Vec(ncores) { new ioHTIF().flip() }
+    val mem = new ioDCache().flip()
+  }
+
+  val short_request_bits = 64
+  val long_request_bits = 576
+  require(short_request_bits % w == 0)
+
+  val rx_count_w = 13 + log2up(8) - log2up(w) // data size field is 12 bits
+  val rx_count = Reg(resetVal = UFix(0,rx_count_w))
+  val rx_shifter = Reg() { Bits(width = short_request_bits) }
+  val header = Reg() { Bits() }
+  val rx_shifter_in = Cat(io.host.in.bits, rx_shifter(short_request_bits-1,w))
+  when (io.host.in.valid && io.host.in.ready) {
+    rx_shifter := rx_shifter_in
+    rx_count := rx_count + UFix(1)
+    when (rx_count === UFix(short_request_bits/w-1)) {
+      header := rx_shifter_in
+    }
+  }
+
+  val rx_count_words = rx_count >> UFix(log2up(short_request_bits/w))
+  val packet_ram_wen = rx_count(log2up(short_request_bits/w)-1,0).andR &&
+                       io.host.in.valid && io.host.in.ready
+  val packet_ram = Mem(long_request_bits/short_request_bits-1,
+                       packet_ram_wen, rx_count_words - UFix(1), rx_shifter_in)
+
+  val cmd_readmem :: cmd_writemem :: cmd_readcr :: cmd_writecr :: cmd_ack :: cmd_nack :: Nil = Enum(6) { UFix() }
+  val cmd = header(3,0)
+  val size = header(15,4)
+  val seqno = header(23,16)
+  val addr = header(63,24).toUFix
+
+  val pcr_addr = addr(19,0)
+  val pcr_coreid = addr(39,20)
+  val pcr_wdata = packet_ram(UFix(0))
+
+  val nack = Mux(cmd === cmd_readmem || cmd === cmd_writemem, size != UFix((1 << OFFSET_BITS)/8),
+             Mux(cmd === cmd_readcr || cmd === cmd_writecr, size != UFix(1),
+             Bool(true)))
+
+  val tx_count = Reg(resetVal = UFix(0, log2up(long_request_bits/w+1)))
+  val packet_ram_raddr = (tx_count >> UFix(log2up(short_request_bits/w)))
+  when (io.host.out.valid && io.host.out.ready) {
+    tx_count := tx_count + UFix(1)
+  }
+
+  val rx_size = Mux(cmd === cmd_writemem || cmd === cmd_writecr, size, UFix(0))
+  val rx_done = rx_count >= UFix(short_request_bits/w) && rx_count_words-UFix(1) === rx_size
+  val tx_size = Mux(!nack && cmd === cmd_readmem, UFix((1 << OFFSET_BITS)/8),
+                Mux(!nack && cmd === cmd_readcr, UFix(1), UFix(0)))
+  val tx_done = packet_ram_raddr - UFix(1) === tx_size
+
+  val state_rx :: state_pcr :: state_mem_req :: state_mem_resp :: state_tx :: Nil = Enum(5) { UFix() }
+  val state = Reg(resetVal = state_rx)
+
+  when (state === state_rx && rx_done) {
+    state := Mux(cmd === cmd_readmem || cmd === cmd_writemem, state_mem_req,
+             Mux(cmd === cmd_readcr || cmd === cmd_writecr, state_pcr,
+             state_tx))
+  }
+
+  val pcr_done = Reg() { Bool() }
+  when (state === state_pcr && pcr_done) {
+    state := state_tx
+  }
+
+  val mem_cnt = Reg(resetVal = UFix(0, log2up(REFILL_CYCLES)))
+  when (state === state_mem_req && io.mem.req_rdy) {
+    when (cmd === cmd_writemem) {
+      when (mem_cnt.andR)  {
+        state := state_tx
+      }
+      mem_cnt := mem_cnt + UFix(1)
+    }
+    .otherwise {
+      state := state_mem_resp
+    }
+  }
+  when (state === state_mem_resp && io.mem.resp_val) {
+    when (mem_cnt.andR) {
+      state := state_tx
+    }
+    mem_cnt := mem_cnt + UFix(1)
+  }
+  when (state === state_tx && tx_done) {
+    rx_count := UFix(0)
+    tx_count := UFix(0)
+    state := state_rx
+  }
+
+  var mem_req_data: Bits = null
+  for (i <- 0 until MEM_DATA_BITS/short_request_bits) {
+    val idx = Cat(mem_cnt, UFix(i, log2up(MEM_DATA_BITS/short_request_bits)))
+    packet_ram.write(idx, io.mem.resp_data((i+1)*short_request_bits-1, i*short_request_bits),
+                     state === state_mem_resp && io.mem.resp_val)
+    mem_req_data = Cat(packet_ram.read(idx), mem_req_data)
+  }
+  io.mem.req_val := state === state_mem_req
+  io.mem.req_rw := cmd === cmd_writemem
+  io.mem.req_addr := addr >> UFix(3)
+  io.mem.req_wdata := mem_req_data
+
+  pcr_done := Bool(false)
+  val pcr_mux = (new Mux1H(ncores)) { Bits(width = 64) }
+  for (i <- 0 until ncores) {
+    val me = pcr_coreid === UFix(i)
+    io.cpu(i).pcr_wen := Reg(state === state_pcr && cmd === cmd_writecr && me, resetVal = Bool(false))
+    io.cpu(i).pcr_addr := Reg(pcr_addr)
+    io.cpu(i).pcr_wdata := Reg(pcr_wdata)
+
+    val my_reset = Reg(resetVal = Bool(true))
+    when (io.cpu(i).pcr_wen && io.cpu(i).pcr_rdy) {
+      when (io.cpu(i).pcr_addr === UFix(15)) { my_reset := io.cpu(i).pcr_wdata(0) }
+      pcr_done := Bool(true)
+    }
+    io.cpu(i).reset := my_reset
+
+    io.cpu(i).pcr_ren := Reg(state === state_pcr && cmd === cmd_readcr && me, resetVal = Bool(false))
+    val rdata = Reg() { Bits() }
+    when (io.cpu(i).pcr_ren && io.cpu(i).pcr_rdy) {
+      rdata := io.cpu(i).pcr_rdata
+      when (io.cpu(i).pcr_addr === UFix(15)) { rdata := my_reset }
+      pcr_done := Bool(true)
+    }
+    pcr_mux.io.sel(i) := Reg(me)
+    pcr_mux.io.in(i) := rdata
+  }
+
+  val tx_cmd = Mux(nack, cmd_nack, cmd_ack)
+  val tx_cmd_ext = Cat(Bits(0, 4-tx_cmd.getWidth), tx_cmd)
+  val tx_size_ext = Cat(Bits(0, 12-tx_size.getWidth), tx_size)
+  val tx_header = Cat(addr, seqno, tx_size_ext, tx_cmd_ext)
+  val tx_data = Mux(packet_ram_raddr === UFix(0), tx_header,
+                Mux(packet_ram_raddr === UFix(1) && cmd === cmd_readcr, pcr_mux.io.out,
+                packet_ram(packet_ram_raddr - UFix(1))))
+
+  io.host.in.ready := state === state_rx && !rx_done
+  io.host.out.valid := state === state_tx && !tx_done
+  io.host.out.bits := tx_data >> Cat(tx_count(log2up(short_request_bits/w)-1,0), Bits(0, log2up(w)))
+}
diff --git a/rocket/src/main/scala/top.scala b/rocket/src/main/scala/top.scala
index a3b1f5c5..293fc9f2 100644
--- a/rocket/src/main/scala/top.scala
+++ b/rocket/src/main/scala/top.scala
@@ -4,15 +4,17 @@ import Chisel._
 import Node._;
 import Constants._;
 
-class ioTop extends Bundle  {
+class ioTop(htif_width: Int) extends Bundle  {
   val debug   = new ioDebug();
-  val console = new ioConsole();
-  val host    = new ioHost();
+  val host    = new ioHost(htif_width);
   val mem     = new ioMem();
 }
 
 class Top() extends Component {
-  val io        = new ioTop();
+
+  val htif_width = 16
+  val io = new ioTop(htif_width);
+  val htif = new rocketHTIF(htif_width, 1)
   
   val cpu       = new rocketProc();
   val icache    = new rocketICache(128, 2); // 128 sets x 2 ways
@@ -25,10 +27,11 @@ class Top() extends Component {
   arbiter.io.dcache <> dcache.io.mem;
   arbiter.io.icache <> icache_pf.io.mem;
   arbiter.io.vicache <> vicache.io.mem
+  arbiter.io.htif <> htif.io.mem
 
-  cpu.io.host       <> io.host;
+  htif.io.host <> io.host
+  cpu.io.host       <> htif.io.cpu(0);
   cpu.io.debug      <> io.debug;
-  cpu.io.console    <> io.console;
 
   icache_pf.io.invalidate := cpu.io.imem.invalidate
   icache.io.mem     <> icache_pf.io.icache;