diff --git a/rocket/src/main/scala/coherence.scala b/rocket/src/main/scala/coherence.scala index 1bd19bbd..e0f5e5c8 100644 --- a/rocket/src/main/scala/coherence.scala +++ b/rocket/src/main/scala/coherence.scala @@ -50,7 +50,7 @@ object cpuCmdToRW { } } -trait CoherencePolicy { +abstract class CoherencePolicy { def isHit (cmd: Bits, state: UFix): Bool def isValid (state: UFix): Bool @@ -72,9 +72,10 @@ trait CoherencePolicy { def newProbeReply (incoming: ProbeRequest, state: UFix): ProbeReply - def hasData (reply: ProbeReply): Bool - def hasData (init: TransactionInit): Bool - def hasData (reply: TransactionReply): Bool + def messageHasData (reply: ProbeReply): Bool + def messageHasData (init: TransactionInit): Bool + def messageHasData (reply: TransactionReply): Bool + def messageUpdatesDataArray (reply: TransactionReply): Bool def isCoherenceConflict(addr1: Bits, addr2: Bits): Bool def getTransactionReplyType(x_type: UFix, count: UFix): Bits @@ -84,7 +85,14 @@ trait CoherencePolicy { def needsAckReply(x_type: UFix, global_state: UFix): Bool } -trait IncoherentPolicy extends CoherencePolicy { +trait UncachedTransactions { + def getTransactionInitTypeOnUncachedRead(): UFix + def getTransactionInitTypeOnUncachedWrite(): UFix +} + +abstract class CoherencePolicyWithUncached extends CoherencePolicy with UncachedTransactions + +abstract class IncoherentPolicy extends CoherencePolicy { // UNIMPLEMENTED def newStateOnProbeRequest(incoming: ProbeRequest, state: UFix): Bits = state def newProbeReply (incoming: ProbeRequest, state: UFix): ProbeReply = { @@ -93,7 +101,7 @@ trait IncoherentPolicy extends CoherencePolicy { reply.global_xact_id := UFix(0) reply } - def hasData (reply: ProbeReply) = Bool(false) + def messageHasData (reply: ProbeReply) = Bool(false) def isCoherenceConflict(addr1: Bits, addr2: Bits): Bool = Bool(false) def getTransactionReplyType(x_type: UFix, count: UFix): Bits = Bits(0) def getProbeRequestType(x_type: UFix, global_state: UFix): UFix = UFix(0) @@ -102,7 +110,7 @@ trait IncoherentPolicy extends CoherencePolicy { def needsAckReply(x_type: UFix, global_state: UFix): Bool = Bool(false) } -trait ThreeStateIncoherence extends IncoherentPolicy { +class ThreeStateIncoherence extends IncoherentPolicy { val tileInvalid :: tileClean :: tileDirty :: Nil = Enum(3){ UFix() } val xactInitReadClean :: xactInitReadDirty :: xactInitWriteback :: Nil = Enum(3){ UFix() } val xactReplyData :: xactReplyAck :: Nil = Enum(2){ UFix() } @@ -141,11 +149,256 @@ trait ThreeStateIncoherence extends IncoherentPolicy { def getTransactionInitTypeOnCacheControl(cmd: Bits): Bits = xactInitWriteback //TODO def getTransactionInitTypeOnWriteback(): Bits = xactInitWriteback - def hasData (init: TransactionInit): Bool = (init.x_type === xactInitWriteback) - def hasData (reply: TransactionReply) = (reply.x_type === xactReplyData) + def messageHasData (init: TransactionInit): Bool = (init.x_type === xactInitWriteback) + def messageHasData (reply: TransactionReply) = (reply.x_type === xactReplyData) + def messageUpdatesDataArray (reply: TransactionReply) = (reply.x_type === xactReplyData) } -trait FourStateCoherence extends CoherencePolicy { +class TwoStateCoherence extends CoherencePolicyWithUncached { + + val tileInvalid :: tileValid :: Nil = Enum(2){ UFix() } + val globalInvalid :: globalValid :: Nil = Enum(2){ UFix() } + + val xactInitReadExclusive :: xactInitReadUncached :: xactInitWriteUncached :: Nil = Enum(3){ UFix() } + val xactReplyReadExclusive :: xactReplyReadUncached :: xactReplyWriteUncached :: Nil = Enum(3){ UFix() } + val probeReqInvalidate :: probeReqCopy :: Nil = Enum(2){ UFix() } + val probeRepInvalidateData :: probeRepCopyData :: probeRepInvalidateAck :: probeRepCopyAck :: Nil = Enum(4){ UFix() } + + def isHit (cmd: Bits, state: UFix): Bool = state != tileInvalid + def isValid (state: UFix): Bool = state != tileInvalid + + def needsTransactionOnSecondaryMiss(cmd: Bits, outstanding: TransactionInit): Bool = (outstanding.x_type != xactInitReadExclusive) + def needsTransactionOnCacheControl(cmd: Bits, state: UFix): Bool = { + MuxLookup(cmd, (state === tileValid), Array( + M_INV -> (state === tileValid), + M_CLN -> (state === tileValid) + )) + } + def needsWriteback (state: UFix): Bool = { + needsTransactionOnCacheControl(M_INV, state) + } + + def newStateOnHit(cmd: Bits, state: UFix): UFix = state + def newStateOnCacheControl(cmd: Bits) = { + MuxLookup(cmd, tileInvalid, Array( + M_INV -> tileInvalid, + M_CLN -> tileValid + )) + } + def newStateOnWriteback() = newStateOnCacheControl(M_INV) + def newStateOnFlush() = newStateOnCacheControl(M_INV) + def newStateOnTransactionReply(incoming: TransactionReply, outstanding: TransactionInit): UFix = { + MuxLookup(incoming.x_type, tileInvalid, Array( + xactReplyReadExclusive -> tileValid, + xactReplyReadUncached -> tileInvalid, + xactReplyWriteUncached -> tileInvalid + )) + } + def newStateOnProbeRequest(incoming: ProbeRequest, state: UFix): Bits = { + MuxLookup(incoming.p_type, state, Array( + probeReqInvalidate -> tileInvalid, + probeReqCopy -> state + )) + } + + def getTransactionInitTypeOnUncachedRead() = xactInitReadUncached + def getTransactionInitTypeOnUncachedWrite() = xactInitWriteUncached + def getTransactionInitTypeOnPrimaryMiss(cmd: Bits, state: UFix): UFix = xactInitReadExclusive + def getTransactionInitTypeOnSecondaryMiss(cmd: Bits, state: UFix, outstanding: TransactionInit): UFix = xactInitReadExclusive + def getTransactionInitTypeOnCacheControl(cmd: Bits): Bits = xactInitWriteUncached + def getTransactionInitTypeOnWriteback(): Bits = getTransactionInitTypeOnCacheControl(M_INV) + + def newProbeReply (incoming: ProbeRequest, state: UFix): ProbeReply = { + val reply = Wire() { new ProbeReply() } + val with_data = MuxLookup(incoming.p_type, probeRepInvalidateData, Array( + probeReqInvalidate -> probeRepInvalidateData, + probeReqCopy -> probeRepCopyData + )) + val without_data = MuxLookup(incoming.p_type, probeRepInvalidateAck, Array( + probeReqInvalidate -> probeRepInvalidateAck, + probeReqCopy -> probeRepCopyAck + )) + reply.p_type := Mux(needsWriteback(state), with_data, without_data) + reply.global_xact_id := incoming.global_xact_id + reply + } + + def messageHasData (reply: ProbeReply): Bool = { + (reply.p_type === probeRepInvalidateData || + reply.p_type === probeRepCopyData) + } + def messageHasData (init: TransactionInit): Bool = { + (init.x_type === xactInitWriteUncached) + } + def messageHasData (reply: TransactionReply): Bool = { + (reply.x_type != xactReplyWriteUncached) + } + def messageUpdatesDataArray (reply: TransactionReply): Bool = { + (reply.x_type === xactReplyReadExclusive) + } + + def isCoherenceConflict(addr1: Bits, addr2: Bits): Bool = (addr1 === addr2) + + def getTransactionReplyType(x_type: UFix, count: UFix): Bits = { + MuxLookup(x_type, xactReplyReadUncached, Array( + xactInitReadExclusive -> xactReplyReadExclusive, + xactInitReadUncached -> xactReplyReadUncached, + xactInitWriteUncached -> xactReplyWriteUncached + )) + } + + def getProbeRequestType(x_type: UFix, global_state: UFix): UFix = { + MuxLookup(x_type, probeReqCopy, Array( + xactInitReadExclusive -> probeReqInvalidate, + xactInitReadUncached -> probeReqCopy, + xactInitWriteUncached -> probeReqInvalidate + )) + } + + def needsMemRead(x_type: UFix, global_state: UFix): Bool = { + (x_type != xactInitWriteUncached) + } + def needsMemWrite(x_type: UFix, global_state: UFix): Bool = { + (x_type === xactInitWriteUncached) + } + def needsAckReply(x_type: UFix, global_state: UFix): Bool = { + (x_type === xactInitWriteUncached) + } +} + +class ThreeStateCoherence extends CoherencePolicyWithUncached { //MEI + + val tileInvalid :: tileExclusiveClean :: tileExclusiveDirty :: Nil = Enum(3){ UFix() } + val globalInvalid :: globalExclusiveClean :: Nil = Enum(2){ UFix() } + + val xactInitReadExclusiveClean :: xactInitReadExclusiveDirty :: xactInitReadUncached :: xactInitWriteUncached :: Nil = Enum(4){ UFix() } + val xactReplyReadExclusive :: xactReplyReadUncached :: xactReplyWriteUncached :: xactReplyReadExclusiveAck :: Nil = Enum(4){ UFix() } + val probeReqInvalidate :: probeReqDowngrade :: probeReqCopy :: Nil = Enum(3){ UFix() } + val probeRepInvalidateData :: probeRepDowngradeData :: probeRepCopyData :: probeRepInvalidateAck :: probeRepDowngradeAck :: probeRepCopyAck :: Nil = Enum(6){ UFix() } + + def isHit (cmd: Bits, state: UFix): Bool = state != tileInvalid + def isValid (state: UFix): Bool = state != tileInvalid + + def needsTransactionOnSecondaryMiss(cmd: Bits, outstanding: TransactionInit): Bool = { + val (read, write) = cpuCmdToRW(cmd) + (read && (outstanding.x_type === xactInitReadUncached || outstanding.x_type === xactInitWriteUncached)) || + (write && (outstanding.x_type != xactInitReadExclusiveDirty)) + } + def needsTransactionOnCacheControl(cmd: Bits, state: UFix): Bool = { + MuxLookup(cmd, (state === tileExclusiveDirty), Array( + M_INV -> (state === tileExclusiveDirty), + M_CLN -> (state === tileExclusiveDirty) + )) + } + def needsWriteback (state: UFix): Bool = { + needsTransactionOnCacheControl(M_INV, state) + } + + def newStateOnHit(cmd: Bits, state: UFix): UFix = { + val (read, write) = cpuCmdToRW(cmd) + Mux(write, tileExclusiveDirty, state) + } + def newStateOnCacheControl(cmd: Bits) = { + MuxLookup(cmd, tileInvalid, Array( + M_INV -> tileInvalid, + M_CLN -> tileExclusiveClean + )) + } + def newStateOnWriteback() = newStateOnCacheControl(M_INV) + def newStateOnFlush() = newStateOnCacheControl(M_INV) + def newStateOnTransactionReply(incoming: TransactionReply, outstanding: TransactionInit): UFix = { + MuxLookup(incoming.x_type, tileInvalid, Array( + xactReplyReadExclusive -> Mux(outstanding.x_type === xactInitReadExclusiveDirty, tileExclusiveDirty, tileExclusiveClean), + xactReplyReadExclusiveAck -> tileExclusiveDirty, + xactReplyReadUncached -> tileInvalid, + xactReplyWriteUncached -> tileInvalid + )) + } + def newStateOnProbeRequest(incoming: ProbeRequest, state: UFix): Bits = { + MuxLookup(incoming.p_type, state, Array( + probeReqInvalidate -> tileInvalid, + probeReqDowngrade -> tileExclusiveClean, + probeReqCopy -> state + )) + } + + def getTransactionInitTypeOnUncachedRead() = xactInitReadUncached + def getTransactionInitTypeOnUncachedWrite() = xactInitWriteUncached + def getTransactionInitTypeOnPrimaryMiss(cmd: Bits, state: UFix): UFix = { + val (read, write) = cpuCmdToRW(cmd) + Mux(write, xactInitReadExclusiveDirty, xactInitReadExclusiveClean) + } + def getTransactionInitTypeOnSecondaryMiss(cmd: Bits, state: UFix, outstanding: TransactionInit): UFix = { + val (read, write) = cpuCmdToRW(cmd) + Mux(write, xactInitReadExclusiveDirty, outstanding.x_type) + } + def getTransactionInitTypeOnCacheControl(cmd: Bits): Bits = xactInitWriteUncached + def getTransactionInitTypeOnWriteback(): Bits = getTransactionInitTypeOnCacheControl(M_INV) + + def newProbeReply (incoming: ProbeRequest, state: UFix): ProbeReply = { + val reply = Wire() { new ProbeReply() } + val with_data = MuxLookup(incoming.p_type, probeRepInvalidateData, Array( + probeReqInvalidate -> probeRepInvalidateData, + probeReqDowngrade -> probeRepDowngradeData, + probeReqCopy -> probeRepCopyData + )) + val without_data = MuxLookup(incoming.p_type, probeRepInvalidateAck, Array( + probeReqInvalidate -> probeRepInvalidateAck, + probeReqDowngrade -> probeRepDowngradeAck, + probeReqCopy -> probeRepCopyAck + )) + reply.p_type := Mux(needsWriteback(state), with_data, without_data) + reply.global_xact_id := incoming.global_xact_id + reply + } + + def messageHasData (reply: ProbeReply): Bool = { + (reply.p_type === probeRepInvalidateData || + reply.p_type === probeRepDowngradeData || + reply.p_type === probeRepCopyData) + } + def messageHasData (init: TransactionInit): Bool = { + (init.x_type === xactInitWriteUncached) + } + def messageHasData (reply: TransactionReply): Bool = { + (reply.x_type != xactReplyWriteUncached && reply.x_type != xactReplyReadExclusiveAck) + } + def messageUpdatesDataArray (reply: TransactionReply): Bool = { + (reply.x_type === xactReplyReadExclusive) + } + + def isCoherenceConflict(addr1: Bits, addr2: Bits): Bool = (addr1 === addr2) + + def getTransactionReplyType(x_type: UFix, count: UFix): Bits = { + MuxLookup(x_type, xactReplyReadUncached, Array( + xactInitReadExclusiveClean -> xactReplyReadExclusive, + xactInitReadExclusiveDirty -> xactReplyReadExclusive, + xactInitReadUncached -> xactReplyReadUncached, + xactInitWriteUncached -> xactReplyWriteUncached + )) + } + + def getProbeRequestType(x_type: UFix, global_state: UFix): UFix = { + MuxLookup(x_type, probeReqCopy, Array( + xactInitReadExclusiveClean -> probeReqInvalidate, + xactInitReadExclusiveDirty -> probeReqInvalidate, + xactInitReadUncached -> probeReqCopy, + xactInitWriteUncached -> probeReqInvalidate + )) + } + + def needsMemRead(x_type: UFix, global_state: UFix): Bool = { + (x_type != xactInitWriteUncached) + } + def needsMemWrite(x_type: UFix, global_state: UFix): Bool = { + (x_type === xactInitWriteUncached) + } + def needsAckReply(x_type: UFix, global_state: UFix): Bool = { + (x_type === xactInitWriteUncached) + } +} + +class FourStateCoherence extends CoherencePolicyWithUncached { val tileInvalid :: tileShared :: tileExclusiveClean :: tileExclusiveDirty :: Nil = Enum(4){ UFix() } val globalInvalid :: globalShared :: globalExclusiveClean :: Nil = Enum(3){ UFix() } @@ -208,6 +461,8 @@ trait FourStateCoherence extends CoherencePolicy { )) } + def getTransactionInitTypeOnUncachedRead() = xactInitReadUncached + def getTransactionInitTypeOnUncachedWrite() = xactInitWriteUncached def getTransactionInitTypeOnPrimaryMiss(cmd: Bits, state: UFix): UFix = { val (read, write) = cpuCmdToRW(cmd) Mux(write || cmd === M_PFW, xactInitReadExclusive, xactInitReadShared) @@ -236,17 +491,20 @@ trait FourStateCoherence extends CoherencePolicy { reply } - def hasData (reply: ProbeReply): Bool = { + def messageHasData (reply: ProbeReply): Bool = { (reply.p_type === probeRepInvalidateData || reply.p_type === probeRepDowngradeData || reply.p_type === probeRepCopyData) } - def hasData (init: TransactionInit): Bool = { + def messageHasData (init: TransactionInit): Bool = { (init.x_type === xactInitWriteUncached) } - def hasData (reply: TransactionReply): Bool = { + def messageHasData (reply: TransactionReply): Bool = { (reply.x_type != xactReplyWriteUncached && reply.x_type != xactReplyReadExclusiveAck) } + def messageUpdatesDataArray (reply: TransactionReply): Bool = { + (reply.x_type === xactReplyReadShared || reply.x_type === xactReplyReadExclusive) + } def isCoherenceConflict(addr1: Bits, addr2: Bits): Bool = (addr1 === addr2) diff --git a/rocket/src/main/scala/htif.scala b/rocket/src/main/scala/htif.scala index 0b312136..a8bf52cc 100644 --- a/rocket/src/main/scala/htif.scala +++ b/rocket/src/main/scala/htif.scala @@ -30,7 +30,7 @@ class ioHTIF extends Bundle val pcr_rep = (new ioPipe) { Bits(width = 64) } } -class rocketHTIF(w: Int, ncores: Int) extends Component with FourStateCoherence +class rocketHTIF(w: Int, ncores: Int, co: CoherencePolicyWithUncached) extends Component { val io = new Bundle { val host = new ioHost(w) @@ -165,7 +165,7 @@ class rocketHTIF(w: Int, ncores: Int) extends Component with FourStateCoherence mem_req_data = Cat(packet_ram(idx), mem_req_data) } io.mem.xact_init.valid := state === state_mem_req - io.mem.xact_init.bits.x_type := Mux(cmd === cmd_writemem, xactInitWriteUncached, xactInitReadUncached) + io.mem.xact_init.bits.x_type := Mux(cmd === cmd_writemem, co.getTransactionInitTypeOnUncachedWrite, co.getTransactionInitTypeOnUncachedRead) io.mem.xact_init.bits.address := addr >> UFix(OFFSET_BITS-3) io.mem.xact_init_data.valid:= state === state_mem_wdata io.mem.xact_init_data.bits.data := mem_req_data @@ -175,7 +175,7 @@ class rocketHTIF(w: Int, ncores: Int) extends Component with FourStateCoherence val probe_q = (new queue(1)) { new ProbeReply } probe_q.io.enq.valid := io.mem.probe_req.valid io.mem.probe_req.ready := probe_q.io.enq.ready - probe_q.io.enq.bits := newProbeReply(io.mem.probe_req.bits, newStateOnFlush()) + probe_q.io.enq.bits := co.newProbeReply(io.mem.probe_req.bits, co.newStateOnFlush()) io.mem.probe_rep <> probe_q.io.deq io.mem.probe_rep_data.valid := Bool(false) diff --git a/rocket/src/main/scala/icache.scala b/rocket/src/main/scala/icache.scala index 08d44ae7..0b716563 100644 --- a/rocket/src/main/scala/icache.scala +++ b/rocket/src/main/scala/icache.scala @@ -27,7 +27,7 @@ class ioRocketICache extends Bundle() // 32 bit wide cpu port, 128 bit wide memory port, 64 byte cachelines // parameters : // lines = # cache lines -class rocketICache(sets: Int, assoc: Int) extends Component with FourStateCoherence +class rocketICache(sets: Int, assoc: Int, co: CoherencePolicyWithUncached) extends Component { val io = new ioRocketICache(); @@ -136,7 +136,7 @@ class rocketICache(sets: Int, assoc: Int) extends Component with FourStateCohere rdy := !io.cpu.itlb_miss && (state === s_ready) && (!r_cpu_req_val || tag_hit); io.cpu.resp_data := data_mux.io.out io.mem.xact_init.valid := (state === s_request) && finish_q.io.enq.ready - io.mem.xact_init.bits.x_type := xactInitReadUncached + io.mem.xact_init.bits.x_type := co.getTransactionInitTypeOnUncachedRead io.mem.xact_init.bits.address := r_cpu_miss_addr(tagmsb,indexlsb).toUFix io.mem.xact_finish <> finish_q.io.deq diff --git a/rocket/src/main/scala/icache_prefetch.scala b/rocket/src/main/scala/icache_prefetch.scala index 9e2b52f5..956dd08f 100644 --- a/rocket/src/main/scala/icache_prefetch.scala +++ b/rocket/src/main/scala/icache_prefetch.scala @@ -11,7 +11,7 @@ class ioIPrefetcher extends Bundle() { val invalidate = Bool(INPUT) } -class rocketIPrefetcher extends Component with FourStateCoherence +class rocketIPrefetcher(co: CoherencePolicyWithUncached) extends Component { val io = new ioIPrefetcher(); val pdq = (new queue(REFILL_CYCLES, flushable = true)) { Bits(width = MEM_DATA_BITS) }; @@ -34,7 +34,7 @@ class rocketIPrefetcher extends Component with FourStateCoherence val finish_q = (new queue(1)) { new TransactionFinish } io.mem.xact_abort.ready := Bool(true) io.mem.xact_init.valid := prefetch_miss || (state === s_req_wait) && finish_q.io.enq.ready - io.mem.xact_init.bits.x_type := xactInitReadUncached + io.mem.xact_init.bits.x_type := co.getTransactionInitTypeOnUncachedRead io.mem.xact_init.bits.tile_xact_id := Mux(prefetch_miss, UFix(0), UFix(1)) io.mem.xact_init.bits.address := Mux(prefetch_miss, io.icache.xact_init.bits.address, prefetch_addr); diff --git a/rocket/src/main/scala/nbdcache.scala b/rocket/src/main/scala/nbdcache.scala index a26f6634..cc91f1de 100644 --- a/rocket/src/main/scala/nbdcache.scala +++ b/rocket/src/main/scala/nbdcache.scala @@ -158,7 +158,7 @@ class MetaArrayReq extends Bundle { val data = new MetaData() } -class MSHR(id: Int) extends Component with FourStateCoherence { +class MSHR(id: Int, co: CoherencePolicy) extends Component { val io = new Bundle { val req_pri_val = Bool(INPUT) val req_pri_rdy = Bool(OUTPUT) @@ -196,7 +196,7 @@ class MSHR(id: Int) extends Component with FourStateCoherence { val req_cmd = io.req_bits.cmd val req_use_rpq = (req_cmd != M_PFR) && (req_cmd != M_PFW) && (req_cmd != M_FLA) val idx_match = req.idx === io.req_bits.idx - val sec_rdy = idx_match && !flush && (state === s_wb_req || state === s_wb_resp || (state === s_refill_req || state === s_refill_resp) && !needsTransactionOnSecondaryMiss(req_cmd, io.mem_req.bits)) + val sec_rdy = idx_match && !flush && (state === s_wb_req || state === s_wb_resp || (state === s_refill_req || state === s_refill_resp) && !co.needsTransactionOnSecondaryMiss(req_cmd, io.mem_req.bits)) val rpq = (new queue(NRPQ)) { new RPQEntry } rpq.io.enq.valid := (io.req_pri_val && io.req_pri_rdy || io.req_sec_val && sec_rdy) && req_use_rpq @@ -220,7 +220,7 @@ class MSHR(id: Int) extends Component with FourStateCoherence { when (refill_done) { state := s_drain_rpq } when (reply) { refill_count := refill_count + UFix(1) - line_state := newStateOnTransactionReply(io.mem_rep.bits, io.mem_req.bits) + line_state := co.newStateOnTransactionReply(io.mem_rep.bits, io.mem_req.bits) } when (abort) { state := s_refill_req } } @@ -239,13 +239,13 @@ class MSHR(id: Int) extends Component with FourStateCoherence { } when (io.req_sec_val && io.req_sec_rdy) { // s_wb_req, s_wb_resp, s_refill_req - xacx_type := getTransactionInitTypeOnSecondaryMiss(req_cmd, newStateOnFlush(), io.mem_req.bits) + xacx_type := co.getTransactionInitTypeOnSecondaryMiss(req_cmd, co.newStateOnFlush(), io.mem_req.bits) } when ((state === s_invalid) && io.req_pri_val) { flush := req_cmd === M_FLA - line_state := newStateOnFlush() + line_state := co.newStateOnFlush() refill_count := UFix(0) - xacx_type := getTransactionInitTypeOnPrimaryMiss(req_cmd, newStateOnFlush()) + xacx_type := co.getTransactionInitTypeOnPrimaryMiss(req_cmd, co.newStateOnFlush()) req := io.req_bits when (io.req_bits.tag_miss) { @@ -289,7 +289,7 @@ class MSHR(id: Int) extends Component with FourStateCoherence { io.replay.bits.way_oh := req.way_oh } -class MSHRFile extends Component { +class MSHRFile(co: CoherencePolicy) extends Component { val io = new Bundle { val req = (new ioDecoupled) { new MSHRReq }.flip val secondary_miss = Bool(OUTPUT) @@ -343,7 +343,7 @@ class MSHRFile extends Component { var refill_probe_rdy = Bool(true) for (i <- 0 to NMSHR-1) { - val mshr = new MSHR(i) + val mshr = new MSHR(i, co) tag_mux.io.sel(i) := mshr.io.idx_match tag_mux.io.in(i) := mshr.io.tag @@ -409,7 +409,8 @@ class MSHRFile extends Component { io.cpu_resp_tag := Reg(replay.bits.cpu_tag) } -class WritebackUnit extends Component with FourStateCoherence{ + +class WritebackUnit(co: CoherencePolicy) extends Component { val io = new Bundle { val req = (new ioDecoupled) { new WritebackReq() }.flip val probe = (new ioDecoupled) { new WritebackReq() }.flip @@ -469,7 +470,7 @@ class WritebackUnit extends Component with FourStateCoherence{ io.data_req.bits.data := Bits(0) io.mem_req.valid := valid && !cmd_sent - io.mem_req.bits.x_type := getTransactionInitTypeOnWriteback() + io.mem_req.bits.x_type := co.getTransactionInitTypeOnWriteback() io.mem_req.bits.address := Cat(req.tag, req.idx).toUFix io.mem_req.bits.tile_xact_id := req.tile_xact_id io.mem_req_data.valid := data_req_fired && !is_probe @@ -478,7 +479,7 @@ class WritebackUnit extends Component with FourStateCoherence{ io.probe_rep_data.bits.data := io.data_resp } -class ProbeUnit extends Component with FourStateCoherence { +class ProbeUnit(co: CoherencePolicy) extends Component { val io = new Bundle { val req = (new ioDecoupled) { new ProbeRequest }.flip val rep = (new ioDecoupled) { new ProbeReply } @@ -504,7 +505,7 @@ class ProbeUnit extends Component with FourStateCoherence { state := s_writeback_resp } when ((state === s_probe_rep) && io.meta_req.ready && io.rep.ready) { - state := Mux(hit && needsWriteback(line_state), s_writeback_req, s_invalid) + state := Mux(hit && co.needsWriteback(line_state), s_writeback_req, s_invalid) } when (state === s_meta_resp) { way_oh := io.tag_match_way_oh @@ -521,13 +522,13 @@ class ProbeUnit extends Component with FourStateCoherence { io.req.ready := state === s_invalid io.rep.valid := state === s_probe_rep && io.meta_req.ready - io.rep.bits := newProbeReply(req, Mux(hit, line_state, newStateOnFlush())) + io.rep.bits := co.newProbeReply(req, Mux(hit, line_state, co.newStateOnFlush())) io.meta_req.valid := state === s_meta_req || state === s_meta_resp || state === s_probe_rep && hit io.meta_req.bits.way_en := Mux(state === s_probe_rep, way_oh, ~UFix(0, NWAYS)) io.meta_req.bits.rw := state === s_probe_rep io.meta_req.bits.idx := req.address - io.meta_req.bits.data.state := newStateOnProbeRequest(req, line_state) + io.meta_req.bits.data.state := co.newStateOnProbeRequest(req, line_state) io.meta_req.bits.data.tag := req.address >> UFix(IDX_BITS) io.mshr_req.valid := state === s_meta_resp io.address := req.address @@ -538,7 +539,7 @@ class ProbeUnit extends Component with FourStateCoherence { io.wb_req.bits.tag := req.address >> UFix(IDX_BITS) } -class FlushUnit(lines: Int) extends Component with FourStateCoherence{ +class FlushUnit(lines: Int, co: CoherencePolicy) extends Component { val io = new Bundle { val req = (new ioDecoupled) { Bool() }.flip val meta_req = (new ioDecoupled) { new MetaArrayReq() } @@ -583,7 +584,7 @@ class FlushUnit(lines: Int) extends Component with FourStateCoherence{ io.meta_req.bits.way_en := UFixToOH(way_cnt, NWAYS) io.meta_req.bits.idx := idx_cnt io.meta_req.bits.rw := (state === s_reset) - io.meta_req.bits.data.state := newStateOnFlush() + io.meta_req.bits.data.state := co.newStateOnFlush() io.meta_req.bits.data.tag := UFix(0) } @@ -732,16 +733,7 @@ class ioDmem(view: List[String] = null) extends Bundle(view) { val resp_tag = Bits(DCACHE_TAG_BITS, OUTPUT); } -abstract class HellaCache extends Component { - def isHit ( cmd: Bits, state: UFix): Bool - def isValid (state: UFix): Bool - def needsWriteback (state: UFix): Bool - def newStateOnWriteback(): UFix - def newStateOnFlush(): UFix - def newStateOnHit(cmd: Bits, state: UFix): UFix -} - -class HellaCacheUniproc extends HellaCache with FourStateCoherence { +class HellaCache(co: CoherencePolicy) extends Component { val io = new Bundle { val cpu = new ioDmem() val mem = new ioTileLink @@ -794,10 +786,10 @@ class HellaCacheUniproc extends HellaCache with FourStateCoherence { val r_req_readwrite = r_req_read || r_req_write || r_req_prefetch val nack_hit = Wire() { Bool() } - val wb = new WritebackUnit - val prober = new ProbeUnit - val mshr = new MSHRFile() - val flusher = new FlushUnit(lines) + val wb = new WritebackUnit(co) + val prober = new ProbeUnit(co) + val mshr = new MSHRFile(co) + val flusher = new FlushUnit(lines, co) val replay_amo_val = mshr.io.data_req.valid && mshr.io.data_req.bits.cmd(3).toBool // reset and flush unit @@ -856,10 +848,10 @@ class HellaCacheUniproc extends HellaCache with FourStateCoherence { val early_tag_nack = !meta_arb.io.in(3).ready val cpu_req_ppn = Mux(prober.io.mshr_req.valid, prober.io.address >> UFix(PGIDX_BITS-OFFSET_BITS), io.cpu.req_ppn) val cpu_req_tag = Cat(cpu_req_ppn, r_cpu_req_idx)(tagmsb,taglsb) - val tag_match_arr = (0 until NWAYS).map( w => isValid(meta.io.resp(w).state) && (meta.io.resp(w).tag === cpu_req_tag)) + val tag_match_arr = (0 until NWAYS).map( w => co.isValid(meta.io.resp(w).state) && (meta.io.resp(w).tag === cpu_req_tag)) val tag_match = Cat(Bits(0),tag_match_arr:_*).orR val tag_match_way_oh = Cat(Bits(0),tag_match_arr.reverse:_*)(NWAYS-1, 0) //TODO: use Vec - val tag_hit_arr = (0 until NWAYS).map( w => isHit(r_cpu_req_cmd, meta.io.resp(w).state) && (meta.io.resp(w).tag === cpu_req_tag)) + val tag_hit_arr = (0 until NWAYS).map( w => co.isHit(r_cpu_req_cmd, meta.io.resp(w).state) && (meta.io.resp(w).tag === cpu_req_tag)) val tag_hit = Cat(Bits(0),tag_hit_arr:_*).orR val meta_resp_way_oh = Mux(meta.io.way_en === ~UFix(0, NWAYS), tag_match_way_oh, meta.io.way_en) val data_resp_way_oh = Mux(data.io.way_en === ~UFix(0, NWAYS), tag_match_way_oh, data.io.way_en) @@ -894,7 +886,7 @@ class HellaCacheUniproc extends HellaCache with FourStateCoherence { data_arb.io.in(0).bits.wmask := ~UFix(0, MEM_DATA_BITS/8) data_arb.io.in(0).bits.data := io.mem.xact_rep.bits.data data_arb.io.in(0).bits.way_en := mshr.io.mem_resp_way_oh - data_arb.io.in(0).valid := io.mem.xact_rep.valid && (io.mem.xact_rep.bits.x_type === xactReplyReadShared || io.mem.xact_rep.bits.x_type === xactReplyReadExclusive) + data_arb.io.in(0).valid := io.mem.xact_rep.valid && co.messageUpdatesDataArray(io.mem.xact_rep.bits) // load hits data_arb.io.in(4).bits.offset := io.cpu.req_idx(offsetmsb,ramindexlsb) @@ -924,7 +916,7 @@ class HellaCacheUniproc extends HellaCache with FourStateCoherence { p_store_valid := p_store_valid && !drain_store || (r_cpu_req_val && tag_hit && r_req_store && mshr.io.req.ready && !nack_hit) || p_amo // tag update after a store to an exclusive clean line. - val new_hit_state = newStateOnHit(r_cpu_req_cmd, meta_resp_mux.state) + val new_hit_state = co.newStateOnHit(r_cpu_req_cmd, meta_resp_mux.state) val set_hit_state = r_cpu_req_val && tag_hit && meta_resp_mux.state != new_hit_state meta.io.state_req.bits.rw := Bool(true) meta.io.state_req.bits.idx := Reg(r_cpu_req_idx(indexmsb,indexlsb)) @@ -948,7 +940,7 @@ class HellaCacheUniproc extends HellaCache with FourStateCoherence { // miss handling mshr.io.req.valid := r_cpu_req_val && r_req_readwrite && !nack_hit || flusher.io.mshr_req.valid mshr.io.req.bits.tag_miss := !tag_hit || flusher.io.mshr_req.valid - mshr.io.req.bits.old_dirty := needsWriteback(meta_wb_mux.state) && (!tag_match || flusher.io.mshr_req.valid) // don't wb upgrades + mshr.io.req.bits.old_dirty := co.needsWriteback(meta_wb_mux.state) && (!tag_match || flusher.io.mshr_req.valid) // don't wb upgrades mshr.io.req.bits.old_tag := meta_wb_mux.tag mshr.io.req.bits.tag := cpu_req_tag mshr.io.req.bits.idx := r_cpu_req_idx(indexmsb,indexlsb) diff --git a/rocket/src/main/scala/tile.scala b/rocket/src/main/scala/tile.scala index 1b7f6e05..ef368a85 100644 --- a/rocket/src/main/scala/tile.scala +++ b/rocket/src/main/scala/tile.scala @@ -4,7 +4,7 @@ import Chisel._ import Node._ import Constants._ -class Tile extends Component +class Tile(co: CoherencePolicyWithUncached) extends Component { val io = new Bundle { val tilelink = new ioTileLink @@ -12,8 +12,8 @@ class Tile extends Component } val cpu = new rocketProc(resetSignal = io.host.reset) - val icache = new rocketICache(128, 4) // 128 sets x 4 ways (32KB) - val dcache = new HellaCacheUniproc + val icache = new rocketICache(128, 4, co) // 128 sets x 4 ways (32KB) + val dcache = new HellaCache(co) val arbiter = new rocketMemArbiter(2 + (if (HAVE_VEC) 1 else 0)) arbiter.io.requestor(0) <> dcache.io.mem @@ -30,7 +30,7 @@ class Tile extends Component if (HAVE_VEC) { - val vicache = new rocketICache(128, 1) // 128 sets x 1 ways (8KB) + val vicache = new rocketICache(128, 1, co) // 128 sets x 1 ways (8KB) arbiter.io.requestor(2) <> vicache.io.mem cpu.io.vimem <> vicache.io.cpu } diff --git a/rocket/src/main/scala/top.scala b/rocket/src/main/scala/top.scala index a74c9ff4..bc90cf99 100644 --- a/rocket/src/main/scala/top.scala +++ b/rocket/src/main/scala/top.scala @@ -20,10 +20,11 @@ class Top extends Component val htif_width = 8 val io = new ioTop(htif_width) - val tile = new Tile - val htif = new rocketHTIF(htif_width, 1) + val co = new FourStateCoherence + val tile = new Tile(co) + val htif = new rocketHTIF(htif_width, 1, co) - val hub = new CoherenceHubBroadcast(2) + val hub = new CoherenceHubBroadcast(2, co) hub.io.tiles(0) <> tile.io.tilelink hub.io.tiles(1) <> htif.io.mem diff --git a/rocket/src/main/scala/uncore.scala b/rocket/src/main/scala/uncore.scala index 14164f7f..a8fc9432 100644 --- a/rocket/src/main/scala/uncore.scala +++ b/rocket/src/main/scala/uncore.scala @@ -50,7 +50,7 @@ class ioTileLink extends Bundle { val xact_finish = (new ioDecoupled) { new TransactionFinish } } -class XactTracker(ntiles: Int, id: Int) extends Component with FourStateCoherence { +class XactTracker(ntiles: Int, id: Int, co: CoherencePolicy) extends Component { val io = new Bundle { val alloc_req = (new ioDecoupled) { new TrackerAllocReq }.flip val p_data = (new ioPipe) { new TrackerProbeData }.flip @@ -140,7 +140,7 @@ class XactTracker(ntiles: Int, id: Int) extends Component with FourStateCoherenc io.sharer_count := UFix(ntiles) // TODO: Broadcast only io.x_type := x_type_ - io.mem_req_cmd.valid := Bool(false) + io.mem_req_cmd.valid := Bool(false) io.mem_req_cmd.bits.rw := Bool(false) io.mem_req_cmd.bits.addr := addr_ io.mem_req_cmd.bits.tag := UFix(id) @@ -148,7 +148,7 @@ class XactTracker(ntiles: Int, id: Int) extends Component with FourStateCoherenc io.mem_req_data.bits.data := UFix(0) io.mem_req_lock := Bool(false) io.probe_req.valid := Bool(false) - io.probe_req.bits.p_type := getProbeRequestType(x_type_, UFix(0)) + io.probe_req.bits.p_type := co.getProbeRequestType(x_type_, UFix(0)) io.probe_req.bits.global_xact_id := UFix(id) io.probe_req.bits.address := addr_ io.push_p_req := Bits(0, width = ntiles) @@ -167,8 +167,8 @@ class XactTracker(ntiles: Int, id: Int) extends Component with FourStateCoherenc x_type_ := io.alloc_req.bits.xact_init.x_type init_tile_id_ := io.alloc_req.bits.tile_id tile_xact_id_ := io.alloc_req.bits.xact_init.tile_xact_id - x_init_data_needs_write := hasData(io.alloc_req.bits.xact_init) - x_needs_read := needsMemRead(io.alloc_req.bits.xact_init.x_type, UFix(0)) + x_init_data_needs_write := co.messageHasData(io.alloc_req.bits.xact_init) + x_needs_read := co.needsMemRead(io.alloc_req.bits.xact_init.x_type, UFix(0)) if(ntiles > 1) p_rep_count := UFix(ntiles-1) val p_req_initial_flags = ~( UFix(1) << io.alloc_req.bits.tile_id ) //TODO: Broadcast only p_req_flags := p_req_initial_flags @@ -226,7 +226,7 @@ class XactTracker(ntiles: Int, id: Int) extends Component with FourStateCoherenc } . elsewhen (x_needs_read) { doMemReqRead(io.mem_req_cmd, x_needs_read) } . otherwise { - state := Mux(needsAckReply(x_type_, UFix(0)), s_ack, s_busy) + state := Mux(co.needsAckReply(x_type_, UFix(0)), s_ack, s_busy) } } is(s_ack) { @@ -241,17 +241,17 @@ class XactTracker(ntiles: Int, id: Int) extends Component with FourStateCoherenc } } -abstract class CoherenceHub(ntiles: Int) extends Component with CoherencePolicy { +abstract class CoherenceHub(ntiles: Int, co: CoherencePolicy) extends Component { val io = new Bundle { val tiles = Vec(ntiles) { new ioTileLink() }.flip val mem = new ioMem } } -class CoherenceHubNull extends CoherenceHub(1) with ThreeStateIncoherence +class CoherenceHubNull(co: ThreeStateIncoherence) extends CoherenceHub(1, co) { val x_init = io.tiles(0).xact_init - val is_write = x_init.bits.x_type === xactInitWriteback + val is_write = x_init.bits.x_type === co.xactInitWriteback x_init.ready := io.mem.req_cmd.ready && !(is_write && io.mem.resp.valid) //stall write req/resp to handle previous read resp io.mem.req_cmd.valid := x_init.valid && !(is_write && io.mem.resp.valid) io.mem.req_cmd.bits.rw := is_write @@ -260,7 +260,7 @@ class CoherenceHubNull extends CoherenceHub(1) with ThreeStateIncoherence io.mem.req_data <> io.tiles(0).xact_init_data val x_rep = io.tiles(0).xact_rep - x_rep.bits.x_type := Mux(io.mem.resp.valid, xactReplyData, xactReplyAck) + x_rep.bits.x_type := Mux(io.mem.resp.valid, co.xactReplyData, co.xactReplyAck) x_rep.bits.tile_xact_id := Mux(io.mem.resp.valid, io.mem.resp.bits.tag, x_init.bits.tile_xact_id) x_rep.bits.global_xact_id := UFix(0) // don't care x_rep.bits.data := io.mem.resp.bits.data @@ -275,9 +275,9 @@ class CoherenceHubNull extends CoherenceHub(1) with ThreeStateIncoherence } -class CoherenceHubBroadcast(ntiles: Int) extends CoherenceHub(ntiles) with FourStateCoherence +class CoherenceHubBroadcast(ntiles: Int, co: CoherencePolicy) extends CoherenceHub(ntiles, co) { - val trackerList = (0 until NGLOBAL_XACTS).map(new XactTracker(ntiles, _)) + val trackerList = (0 until NGLOBAL_XACTS).map(new XactTracker(ntiles, _, co)) val busy_arr = Vec(NGLOBAL_XACTS){ Wire(){Bool()} } val addr_arr = Vec(NGLOBAL_XACTS){ Wire(){Bits(width=PADDR_BITS-OFFSET_BITS)} } @@ -344,12 +344,12 @@ class CoherenceHubBroadcast(ntiles: Int) extends CoherenceHub(ntiles) with FourS rep.bits.require_ack := Bool(true) rep.valid := Bool(false) when(io.mem.resp.valid && (UFix(j) === init_tile_id_arr(mem_idx))) { - rep.bits.x_type := getTransactionReplyType(x_type_arr(mem_idx), sh_count_arr(mem_idx)) + rep.bits.x_type := co.getTransactionReplyType(x_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 := Bool(true) } . otherwise { - rep.bits.x_type := getTransactionReplyType(x_type_arr(ack_idx), sh_count_arr(ack_idx)) + rep.bits.x_type := co.getTransactionReplyType(x_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 when (UFix(j) === init_tile_id_arr(ack_idx)) { @@ -417,16 +417,16 @@ class CoherenceHubBroadcast(ntiles: Int) extends CoherenceHub(ntiles) with FourS val conflicts = Vec(NGLOBAL_XACTS) { Wire() { Bool() } } for( i <- 0 until NGLOBAL_XACTS) { val t = trackerList(i).io - conflicts(i) := t.busy && x_init.valid && isCoherenceConflict(t.addr, x_init.bits.address) + conflicts(i) := t.busy && x_init.valid && co.isCoherenceConflict(t.addr, x_init.bits.address) } x_abort.bits.tile_xact_id := x_init.bits.tile_xact_id - want_to_abort_arr(j) := x_init.valid && (conflicts.toBits.orR || busy_arr.toBits.andR || (!x_init_data_dep_list(j).io.enq.ready && hasData(x_init.bits))) + want_to_abort_arr(j) := x_init.valid && (conflicts.toBits.orR || busy_arr.toBits.andR || (!x_init_data_dep_list(j).io.enq.ready && co.messageHasData(x_init.bits))) x_abort.valid := Bool(false) switch(abort_state_arr(j)) { is(s_idle) { when(want_to_abort_arr(j)) { - when(hasData(x_init.bits)) { + when(co.messageHasData(x_init.bits)) { abort_state_arr(j) := s_abort_drain } . otherwise { abort_state_arr(j) := s_abort_send @@ -478,7 +478,7 @@ class CoherenceHubBroadcast(ntiles: Int) extends CoherenceHub(ntiles) with FourS init_arb.io.in(j).bits.tile_id := UFix(j) val pop_x_inits = trackerList.map(_.io.pop_x_init(j).toBool) val do_pop = foldR(pop_x_inits)(_||_) - x_init_data_dep_list(j).io.enq.valid := do_pop && hasData(x_init.bits) && (abort_state_arr(j) === s_idle) + x_init_data_dep_list(j).io.enq.valid := do_pop && co.messageHasData(x_init.bits) && (abort_state_arr(j) === s_idle) x_init_data_dep_list(j).io.enq.bits.global_xact_id := OHToUFix(pop_x_inits) x_init.ready := (abort_state_arr(j) === s_abort_complete) || do_pop x_init_data.ready := (abort_state_arr(j) === s_abort_drain) || foldR(trackerList.map(_.io.pop_x_init_data(j).toBool))(_||_)