1
0

Fixed elaboration errors in LockingArbiter and BoradcastHub. Fixed ioDecoupled direction error in XactTracker

This commit is contained in:
Henry Cook 2012-03-01 18:23:46 -08:00
parent 9d7707a0a2
commit da39810bb2
2 changed files with 85 additions and 59 deletions

View File

@ -163,13 +163,12 @@ trait FourStateCoherence extends CoherencePolicy {
def needsSecondaryXact (cmd: Bits, outstanding: TransactionInit): Bool def needsSecondaryXact (cmd: Bits, outstanding: TransactionInit): Bool
def getMetaUpdateOnProbe (incoming: ProbeRequest): Bits = { def newStateOnProbe (incoming: ProbeRequest, state: UFix): Bits = {
val state = UFix(0) MuxLookup(incoming.p_type, state, Array(
switch(incoming.p_type) { probeInvalidate -> tileInvalid,
is(probeInvalidate) { state := tileInvalid } probeDowngrade -> tileShared,
is(probeDowngrade) { state := tileShared } probeCopy -> state
} ))
state.toBits
} }
def replyTypeHasData (reply: TransactionReply): Bool = { def replyTypeHasData (reply: TransactionReply): Bool = {
@ -187,9 +186,10 @@ class XactTracker(id: Int) extends Component with CoherencePolicy {
val p_req_cnt_inc = Bits(NTILES, INPUT) val p_req_cnt_inc = Bits(NTILES, INPUT)
val p_rep_data = (new ioDecoupled) { new ProbeReplyData() } val p_rep_data = (new ioDecoupled) { new ProbeReplyData() }
val x_init_data = (new ioDecoupled) { new TransactionInitData() } val x_init_data = (new ioDecoupled) { new TransactionInitData() }
val sent_x_rep_ack = Bool(INPUT)
val mem_req_cmd = (new ioDecoupled) { new MemReqCmd() } val mem_req_cmd = (new ioDecoupled) { new MemReqCmd() }.flip
val mem_req_data = (new ioDecoupled) { new MemData() } val mem_req_data = (new ioDecoupled) { new MemData() }.flip
val mem_req_lock = Bool(OUTPUT) val mem_req_lock = Bool(OUTPUT)
val probe_req = (new ioDecoupled) { new ProbeRequest() }.flip val probe_req = (new ioDecoupled) { new ProbeRequest() }.flip
val busy = Bool(OUTPUT) val busy = Bool(OUTPUT)
@ -222,7 +222,7 @@ class XactTracker(id: Int) extends Component with CoherencePolicy {
(t_type === X_WRITE_UNCACHED) (t_type === X_WRITE_UNCACHED)
} }
val s_idle :: s_mem :: s_probe :: s_busy :: Nil = Enum(4){ UFix() } val s_idle :: s_ack :: s_mem :: s_probe :: s_busy :: Nil = Enum(5){ UFix() }
val state = Reg(resetVal = s_idle) val state = Reg(resetVal = s_idle)
val addr_ = Reg{ UFix() } val addr_ = Reg{ UFix() }
val t_type_ = Reg{ Bits() } val t_type_ = Reg{ Bits() }
@ -241,7 +241,10 @@ class XactTracker(id: Int) extends Component with CoherencePolicy {
def doMemReqWrite(req_cmd: ioDecoupled[MemReqCmd], req_data: ioDecoupled[MemData], lock: Bool, data: ioDecoupled[MemData], trigger: Bool, pop: Bool) { def doMemReqWrite(req_cmd: ioDecoupled[MemReqCmd], req_data: ioDecoupled[MemData], lock: Bool, data: ioDecoupled[MemData], trigger: Bool, pop: Bool) {
req_cmd.valid := mem_cmd_sent req_cmd.valid := mem_cmd_sent
req_cmd.bits.rw := Bool(true) req_cmd.bits.rw := Bool(true)
req_data <> data //TODO: why does req_data <> data segfault?
req_data.valid := data.valid
req_data.bits.data := data.bits.data
data.ready := req_data.ready
lock := Bool(true) lock := Bool(true)
when(req_cmd.ready && req_cmd.valid) { when(req_cmd.ready && req_cmd.valid) {
mem_cmd_sent := Bool(false) mem_cmd_sent := Bool(false)
@ -287,6 +290,8 @@ class XactTracker(id: Int) extends Component with CoherencePolicy {
io.pop_x_init := Bool(false) io.pop_x_init := Bool(false)
io.pop_x_init_data := Bool(false) io.pop_x_init_data := Bool(false)
io.send_x_rep_ack := Bool(false) io.send_x_rep_ack := Bool(false)
io.x_init_data.ready := Bool(false) // don't care
io.p_rep_data.ready := Bool(false) // don't care
switch (state) { switch (state) {
is(s_idle) { is(s_idle) {
@ -336,20 +341,19 @@ class XactTracker(id: Int) extends Component with CoherencePolicy {
} . elsewhen (x_needs_read) { } . elsewhen (x_needs_read) {
doMemReqRead(io.mem_req_cmd, x_needs_read) doMemReqRead(io.mem_req_cmd, x_needs_read)
} . otherwise { } . otherwise {
io.send_x_rep_ack := needsAckRep(t_type_, UFix(0)) state := Mux(needsAckRep(t_type_, UFix(0)), s_ack, s_busy)
state := s_busy
} }
} }
is(s_ack) {
io.send_x_rep_ack := Bool(true)
when(io.sent_x_rep_ack) { state := s_busy }
}
is(s_busy) { // Nothing left to do but wait for transaction to complete is(s_busy) { // Nothing left to do but wait for transaction to complete
when (io.xact_finish) { when (io.xact_finish) {
state := s_idle state := s_idle
} }
} }
} }
//TODO: Decrement the probe count when final data piece is written
// Connent io.mem.ready sig to correct pop* outputs
// P_rep and x_init must be popped on same cycle of receipt
} }
abstract class CoherenceHub extends Component with CoherencePolicy { abstract class CoherenceHub extends Component with CoherencePolicy {
@ -385,19 +389,16 @@ class CoherenceHubBroadcast extends CoherenceHub {
addr1(PADDR_BITS-1, OFFSET_BITS) === addr2(PADDR_BITS-1, OFFSET_BITS) addr1(PADDR_BITS-1, OFFSET_BITS) === addr2(PADDR_BITS-1, OFFSET_BITS)
} }
def getTransactionReplyType(t_type: UFix, count: UFix): Bits = { def getTransactionReplyType(t_type: UFix, count: UFix): Bits = {
val ret = Wire() { Bits(width = TTYPE_BITS) } MuxLookup(t_type, X_READ_UNCACHED, Array(
switch (t_type) { X_READ_SHARED -> Mux(count > UFix(0), X_READ_SHARED, X_READ_EXCLUSIVE),
is(X_READ_SHARED) { ret := Mux(count > UFix(0), X_READ_SHARED, X_READ_EXCLUSIVE) } X_READ_EXCLUSIVE -> X_READ_EXCLUSIVE,
is(X_READ_EXCLUSIVE) { ret := X_READ_EXCLUSIVE } X_READ_UNCACHED -> X_READ_UNCACHED,
is(X_READ_UNCACHED) { ret := X_READ_UNCACHED } X_WRITE_UNCACHED -> X_WRITE_UNCACHED
is(X_WRITE_UNCACHED) { ret := X_WRITE_UNCACHED } ))
}
ret
} }
val trackerList = (0 until NGLOBAL_XACTS).map(new XactTracker(_)) val trackerList = (0 until NGLOBAL_XACTS).map(new XactTracker(_))
/*
val busy_arr = Vec(NGLOBAL_XACTS){ Wire(){Bool()} } val busy_arr = Vec(NGLOBAL_XACTS){ Wire(){Bool()} }
val addr_arr = Vec(NGLOBAL_XACTS){ Wire(){Bits(width=PADDR_BITS)} } val addr_arr = Vec(NGLOBAL_XACTS){ Wire(){Bits(width=PADDR_BITS)} }
val init_tile_id_arr = Vec(NGLOBAL_XACTS){ Wire(){Bits(width=TILE_ID_BITS)} } val init_tile_id_arr = Vec(NGLOBAL_XACTS){ Wire(){Bits(width=TILE_ID_BITS)} }
@ -409,39 +410,60 @@ class CoherenceHubBroadcast extends CoherenceHub {
val do_free_arr = Vec(NGLOBAL_XACTS){ Wire(){Bool()} } val do_free_arr = Vec(NGLOBAL_XACTS){ Wire(){Bool()} }
val p_rep_cnt_dec_arr = Vec(NGLOBAL_XACTS){ Wire(){Bits(width=NTILES)} } val p_rep_cnt_dec_arr = Vec(NGLOBAL_XACTS){ Wire(){Bits(width=NTILES)} }
val p_req_cnt_inc_arr = Vec(NGLOBAL_XACTS){ Wire(){Bits(width=NTILES)} } val p_req_cnt_inc_arr = Vec(NGLOBAL_XACTS){ Wire(){Bits(width=NTILES)} }
val sent_x_rep_ack_arr = Vec(NGLOBAL_XACTS){ Wire(){Bool()} }
for( i <- 0 until NGLOBAL_XACTS) { for( i <- 0 until NGLOBAL_XACTS) {
busy_arr.write( UFix(i), trackerList(i).io.busy) val t = trackerList(i).io
addr_arr.write( UFix(i), trackerList(i).io.addr) busy_arr(i) := t.busy
init_tile_id_arr.write( UFix(i), trackerList(i).io.init_tile_id) addr_arr(i) := t.addr
tile_xact_id_arr.write( UFix(i), trackerList(i).io.tile_xact_id) init_tile_id_arr(i) := t.init_tile_id
t_type_arr.write( UFix(i), trackerList(i).io.t_type) tile_xact_id_arr(i) := t.tile_xact_id
sh_count_arr.write( UFix(i), trackerList(i).io.sharer_count) t_type_arr(i) := t.t_type
send_x_rep_ack_arr.write(UFix(i), trackerList(i).io.send_x_rep_ack) sh_count_arr(i) := t.sharer_count
trackerList(i).io.xact_finish := do_free_arr.read(UFix(i)) send_x_rep_ack_arr(i) := t.send_x_rep_ack
trackerList(i).io.p_rep_cnt_dec := p_rep_cnt_dec_arr.read(UFix(i)) do_free_arr(i) := Bool(false)
trackerList(i).io.p_req_cnt_inc := p_req_cnt_inc_arr.read(UFix(i)) p_rep_cnt_dec_arr(i) := Bits(0)
p_req_cnt_inc_arr(i) := Bits(0)
sent_x_rep_ack_arr(i) := Bool(false)
t.xact_finish := do_free_arr(i)
t.p_rep_cnt_dec := p_rep_cnt_dec_arr(i)
t.p_req_cnt_inc := p_req_cnt_inc_arr(i)
t.sent_x_rep_ack := sent_x_rep_ack_arr(i)
} }
// Free finished transactions // Free finished transactions
for( j <- 0 until NTILES ) { for( j <- 0 until NTILES ) {
val finish = io.tiles(j).xact_finish val finish = io.tiles(j).xact_finish
do_free_arr.write(finish.bits.global_xact_id, finish.valid) do_free_arr(finish.bits.global_xact_id) := finish.valid
finish.ready := Bool(true) finish.ready := Bool(true)
} }
// Reply to initial requestor // Reply to initial requestor
// Forward memory responses from mem to tile // Forward memory responses from mem to tile or arbitrate to ack
val idx = io.mem.resp.bits.tag val mem_idx = io.mem.resp.bits.tag
val ack_idx = PriorityEncoder(send_x_rep_ack_arr.toBits, NGLOBAL_XACTS)
for( j <- 0 until NTILES ) { for( j <- 0 until NTILES ) {
io.tiles(j).xact_rep.bits.t_type := getTransactionReplyType(t_type_arr.read(idx), sh_count_arr.read(idx)) val rep = io.tiles(j).xact_rep
io.tiles(j).xact_rep.bits.tile_xact_id := tile_xact_id_arr.read(idx) rep.bits.t_type := UFix(0)
io.tiles(j).xact_rep.bits.global_xact_id := idx rep.bits.tile_xact_id := UFix(0)
io.tiles(j).xact_rep.bits.data := io.mem.resp.bits.data rep.bits.global_xact_id := UFix(0)
io.tiles(j).xact_rep.valid := (UFix(j) === init_tile_id_arr.read(idx)) && (io.mem.resp.valid || send_x_rep_ack_arr.read(idx)) rep.valid := Bool(false)
when(io.mem.resp.valid) {
rep.bits.t_type := getTransactionReplyType(t_type_arr(mem_idx), sh_count_arr(mem_idx))
rep.bits.tile_xact_id := tile_xact_id_arr(mem_idx)
rep.bits.global_xact_id := mem_idx
rep.valid := (UFix(j) === init_tile_id_arr(mem_idx))
} . otherwise {
rep.bits.t_type := getTransactionReplyType(t_type_arr(ack_idx), sh_count_arr(ack_idx))
rep.bits.tile_xact_id := tile_xact_id_arr(ack_idx)
rep.bits.global_xact_id := ack_idx
rep.valid := (UFix(j) === init_tile_id_arr(ack_idx)) && send_x_rep_ack_arr(ack_idx)
} }
io.tiles(j).xact_rep.bits.data := io.mem.resp.bits.data
}
sent_x_rep_ack_arr(ack_idx) := !io.mem.resp.valid && send_x_rep_ack_arr(ack_idx)
// If there were a ready signal due to e.g. intervening network use: // If there were a ready signal due to e.g. intervening network use:
//io.mem.resp.ready := io.tiles(init_tile_id_arr.read(idx)).xact_rep.ready //io.mem.resp.ready := io.tiles(init_tile_id_arr.read(mem_idx)).xact_rep.ready
// Create an arbiter for the one memory port // Create an arbiter for the one memory port
// We have to arbitrate between the different trackers' memory requests // We have to arbitrate between the different trackers' memory requests
@ -466,7 +488,8 @@ class CoherenceHubBroadcast extends CoherenceHub {
p_rep_data.ready := foldR(trackerList.map(_.io.pop_p_rep_data(j)))(_ || _) p_rep_data.ready := foldR(trackerList.map(_.io.pop_p_rep_data(j)))(_ || _)
} }
for( i <- 0 until NGLOBAL_XACTS ) { for( i <- 0 until NGLOBAL_XACTS ) {
trackerList(i).io.p_rep_data <> io.tiles(trackerList(i).io.p_rep_tile_id).probe_rep_data trackerList(i).io.p_rep_data.valid := io.tiles(trackerList(i).io.p_rep_tile_id).probe_rep_data.valid
trackerList(i).io.p_rep_data.bits := io.tiles(trackerList(i).io.p_rep_tile_id).probe_rep_data.bits
for( j <- 0 until NTILES) { for( j <- 0 until NTILES) {
val p_rep = io.tiles(j).probe_rep val p_rep = io.tiles(j).probe_rep
val dec = p_rep.valid && (p_rep.bits.global_xact_id === UFix(i)) val dec = p_rep.valid && (p_rep.bits.global_xact_id === UFix(i))
@ -475,14 +498,14 @@ class CoherenceHubBroadcast extends CoherenceHub {
} }
// Nack conflicting transaction init attempts // Nack conflicting transaction init attempts
val aborting = Wire() { Bits(width = NTILES) } val aborting = Bits(0, width = NTILES)
for( j <- 0 until NTILES ) { for( j <- 0 until NTILES ) {
val x_init = io.tiles(j).xact_init val x_init = io.tiles(j).xact_init
val x_abort = io.tiles(j).xact_abort val x_abort = io.tiles(j).xact_abort
val conflicts = Bits(width = NGLOBAL_XACTS) val conflicts = Bits(width = NGLOBAL_XACTS)
for( i <- 0 until NGLOBAL_XACTS) { for( i <- 0 until NGLOBAL_XACTS) {
val t = trackerList(i).io val t = trackerList(i).io
conflicts(UFix(i), t.busy(i) && coherenceConflict(t.addr, x_init.bits.address) && conflicts(UFix(i), t.busy && coherenceConflict(t.addr, x_init.bits.address) &&
!(x_init.bits.has_data && (UFix(j) === t.init_tile_id))) !(x_init.bits.has_data && (UFix(j) === t.init_tile_id)))
// Don't abort writebacks stalled on mem. // Don't abort writebacks stalled on mem.
// TODO: This assumes overlapped writeback init reqs to // TODO: This assumes overlapped writeback init reqs to
@ -505,9 +528,9 @@ class CoherenceHubBroadcast extends CoherenceHub {
trackerList(i).io.alloc_req.bits <> init_arb.io.out.bits trackerList(i).io.alloc_req.bits <> init_arb.io.out.bits
trackerList(i).io.alloc_req.valid := init_arb.io.out.valid trackerList(i).io.alloc_req.valid := init_arb.io.out.valid
trackerList(i).io.x_init_data <> io.tiles(trackerList(i).io.init_tile_id).xact_init_data trackerList(i).io.x_init_data.bits := io.tiles(trackerList(i).io.init_tile_id).xact_init_data.bits
trackerList(i).io.x_init_data.valid := io.tiles(trackerList(i).io.init_tile_id).xact_init_data.valid
} }
for( j <- 0 until NTILES ) { for( j <- 0 until NTILES ) {
val x_init = io.tiles(j).xact_init val x_init = io.tiles(j).xact_init
val x_init_data = io.tiles(j).xact_init_data val x_init_data = io.tiles(j).xact_init_data
@ -522,7 +545,6 @@ class CoherenceHubBroadcast extends CoherenceHub {
alloc_arb.io.out.ready := init_arb.io.out.valid && !busy_arr.toBits.andR && alloc_arb.io.out.ready := init_arb.io.out.valid && !busy_arr.toBits.andR &&
!foldR(trackerList.map(t => t.io.busy && coherenceConflict(t.io.addr, init_arb.io.out.bits.xact_init.address)))(_||_) !foldR(trackerList.map(t => t.io.busy && coherenceConflict(t.io.addr, init_arb.io.out.bits.xact_init.address)))(_||_)
// Handle probe request generation // Handle probe request generation
// Must arbitrate for each request port // Must arbitrate for each request port
val p_req_arb_arr = List.fill(NTILES)((new Arbiter(NGLOBAL_XACTS)) { new ProbeRequest() }) val p_req_arb_arr = List.fill(NTILES)((new Arbiter(NGLOBAL_XACTS)) { new ProbeRequest() })
@ -535,5 +557,5 @@ class CoherenceHubBroadcast extends CoherenceHub {
} }
p_req_arb_arr(j).io.out <> io.tiles(j).probe_req p_req_arb_arr(j).io.out <> io.tiles(j).probe_req
} }
*/
} }

View File

@ -206,17 +206,21 @@ class Arbiter[T <: Data](n: Int)(data: => T) extends Component {
class ioLockingArbiter[T <: Data](n: Int)(data: => T) extends Bundle { class ioLockingArbiter[T <: Data](n: Int)(data: => T) extends Bundle {
val in = Vec(n) { (new ioDecoupled()) { data } } val in = Vec(n) { (new ioDecoupled()) { data } }
val lock = Vec(n) { Bool() } val lock = Vec(n) { Bool() }.asInput
val out = (new ioDecoupled()) { data }.flip() val out = (new ioDecoupled()) { data }.flip()
} }
class LockingArbiter[T <: Data](n: Int)(data: => T) extends Component { class LockingArbiter[T <: Data](n: Int)(data: => T) extends Component {
val io = new ioLockingArbiter(n)(data) val io = new ioLockingArbiter(n)(data)
val locked = Reg(resetVal = Bits(0, width = n)) val locked = Vec(n) { Reg(resetVal = Bool(false)) }
var dout = io.in(0).bits var dout = io.in(0).bits
var vout = Bool(false) var vout = Bool(false)
val any_lock_held = (locked & io.lock.toBits).orR for (i <- 0 until n) {
io.in(i).ready := io.out.ready
}
val any_lock_held = (locked.toBits & io.lock.toBits).orR
when(any_lock_held) { when(any_lock_held) {
vout = io.in(0).valid && locked(0) vout = io.in(0).valid && locked(0)
for (i <- 0 until n) { for (i <- 0 until n) {
@ -226,10 +230,10 @@ class LockingArbiter[T <: Data](n: Int)(data: => T) extends Component {
} }
} .otherwise { } .otherwise {
io.in(0).ready := io.out.ready io.in(0).ready := io.out.ready
locked.bitSet(UFix(0), io.out.ready && io.lock(0)) locked(0) := io.out.ready && io.lock(0)
for (i <- 1 until n) { for (i <- 1 until n) {
io.in(i).ready := !io.in(i-1).valid && io.in(i-1).ready io.in(i).ready := !io.in(i-1).valid && io.in(i-1).ready
locked.bitSet(UFix(i), !io.in(i-1).valid && io.in(i-1).ready && io.lock(i)) locked(i) := !io.in(i-1).valid && io.in(i-1).ready && io.lock(i)
} }
dout = io.in(n-1).bits dout = io.in(n-1).bits