/*
- * Copyright (c) 2010-2018 ARM Limited
+ * Copyright (c) 2010-2019 ARM Limited
* All rights reserved
*
* The license below extends only to copyright in the software and shall
#include "debug/DRAMPower.hh"
#include "debug/DRAMState.hh"
#include "debug/Drain.hh"
+#include "debug/QOS.hh"
#include "sim/system.hh"
using namespace std;
using namespace Data;
DRAMCtrl::DRAMCtrl(const DRAMCtrlParams* p) :
- AbstractMemory(p),
+ QoS::MemCtrl(p),
port(name() + ".port", *this), isTimingMode(false),
retryRdReq(false), retryWrReq(false),
- busState(READ),
- busStateNext(READ),
nextReqEvent([this]{ processNextReqEvent(); }, name()),
respondEvent([this]{ processRespondEvent(); }, name()),
deviceSize(p->device_size),
frontendLatency(p->static_frontend_latency),
backendLatency(p->static_backend_latency),
nextBurstAt(0), prevArrival(0),
- nextReqTime(0), activeRank(0), timeStampOffset(0),
- lastStatsResetTick(0)
+ nextReqTime(0),
+ stats(*this),
+ activeRank(0), timeStampOffset(0),
+ lastStatsResetTick(0), enableDRAMPowerdown(p->enable_dram_powerdown)
{
// sanity check the ranks since we rely on bit slicing for the
// address decoding
fatal_if(!isPowerOf2(burstSize), "DRAM burst size %d is not allowed, "
"must be a power of two\n", burstSize);
+ readQueue.resize(p->qos_priorities);
+ writeQueue.resize(p->qos_priorities);
+
for (int i = 0; i < ranksPerChannel; i++) {
Rank* rank = new Rank(*this, p, i);
void
DRAMCtrl::init()
{
- AbstractMemory::init();
+ MemCtrl::init();
if (!port.isConnected()) {
fatal("DRAMCtrl %s is unconnected!\n", name());
DRAMCtrl::readQueueFull(unsigned int neededEntries) const
{
DPRINTF(DRAM, "Read queue limit %d, current size %d, entries needed %d\n",
- readBufferSize, readQueue.size() + respQueue.size(),
+ readBufferSize, totalReadQueueSize + respQueue.size(),
neededEntries);
- return
- (readQueue.size() + respQueue.size() + neededEntries) > readBufferSize;
+ auto rdsize_new = totalReadQueueSize + respQueue.size() + neededEntries;
+ return rdsize_new > readBufferSize;
}
bool
DRAMCtrl::writeQueueFull(unsigned int neededEntries) const
{
DPRINTF(DRAM, "Write queue limit %d, current size %d, entries needed %d\n",
- writeBufferSize, writeQueue.size(), neededEntries);
- return (writeQueue.size() + neededEntries) > writeBufferSize;
+ writeBufferSize, totalWriteQueueSize, neededEntries);
+
+ auto wrsize_new = (totalWriteQueueSize + neededEntries);
+ return wrsize_new > writeBufferSize;
}
DRAMCtrl::DRAMPacket*
-DRAMCtrl::decodeAddr(PacketPtr pkt, Addr dramPktAddr, unsigned size,
- bool isRead)
+DRAMCtrl::decodeAddr(const PacketPtr pkt, Addr dramPktAddr, unsigned size,
+ bool isRead) const
{
// decode the address based on the address mapping scheme, with
// Ro, Ra, Co, Ba and Ch denoting row, rank, column, bank and
for (int cnt = 0; cnt < pktCount; ++cnt) {
unsigned size = std::min((addr | (burstSize - 1)) + 1,
pkt->getAddr() + pkt->getSize()) - addr;
- readPktSize[ceilLog2(size)]++;
- readBursts++;
+ stats.readPktSize[ceilLog2(size)]++;
+ stats.readBursts++;
+ stats.masterReadAccesses[pkt->masterId()]++;
// First check write buffer to see if the data is already at
// the controller
// if the burst address is not present then there is no need
// looking any further
if (isInWriteQueue.find(burst_addr) != isInWriteQueue.end()) {
- for (const auto& p : writeQueue) {
- // check if the read is subsumed in the write queue
- // packet we are looking at
- if (p->addr <= addr && (addr + size) <= (p->addr + p->size)) {
- foundInWrQ = true;
- servicedByWrQ++;
- pktsServicedByWrQ++;
- DPRINTF(DRAM, "Read to addr %lld with size %d serviced by "
- "write queue\n", addr, size);
- bytesReadWrQ += burstSize;
- break;
+ for (const auto& vec : writeQueue) {
+ for (const auto& p : vec) {
+ // check if the read is subsumed in the write queue
+ // packet we are looking at
+ if (p->addr <= addr &&
+ ((addr + size) <= (p->addr + p->size))) {
+
+ foundInWrQ = true;
+ stats.servicedByWrQ++;
+ pktsServicedByWrQ++;
+ DPRINTF(DRAM,
+ "Read to addr %lld with size %d serviced by "
+ "write queue\n",
+ addr, size);
+ stats.bytesReadWrQ += burstSize;
+ break;
+ }
}
}
}
dram_pkt->burstHelper = burst_helper;
assert(!readQueueFull(1));
- rdQLenPdf[readQueue.size() + respQueue.size()]++;
+ stats.rdQLenPdf[totalReadQueueSize + respQueue.size()]++;
DPRINTF(DRAM, "Adding to read queue\n");
- readQueue.push_back(dram_pkt);
+ readQueue[dram_pkt->qosValue()].push_back(dram_pkt);
- // increment read entries of the rank
++dram_pkt->rankRef.readEntries;
+ // log packet
+ logRequest(MemCtrl::READ, pkt->masterId(), pkt->qosValue(),
+ dram_pkt->addr, 1);
+
// Update stats
- avgRdQLen = readQueue.size() + respQueue.size();
+ stats.avgRdQLen = totalReadQueueSize + respQueue.size();
}
// Starting address of next dram pkt (aligend to burstSize boundary)
for (int cnt = 0; cnt < pktCount; ++cnt) {
unsigned size = std::min((addr | (burstSize - 1)) + 1,
pkt->getAddr() + pkt->getSize()) - addr;
- writePktSize[ceilLog2(size)]++;
- writeBursts++;
+ stats.writePktSize[ceilLog2(size)]++;
+ stats.writeBursts++;
+ stats.masterWriteAccesses[pkt->masterId()]++;
// see if we can merge with an existing item in the write
// queue and keep track of whether we have merged or not
if (!merged) {
DRAMPacket* dram_pkt = decodeAddr(pkt, addr, size, false);
- assert(writeQueue.size() < writeBufferSize);
- wrQLenPdf[writeQueue.size()]++;
+ assert(totalWriteQueueSize < writeBufferSize);
+ stats.wrQLenPdf[totalWriteQueueSize]++;
DPRINTF(DRAM, "Adding to write queue\n");
- writeQueue.push_back(dram_pkt);
+ writeQueue[dram_pkt->qosValue()].push_back(dram_pkt);
isInWriteQueue.insert(burstAlign(addr));
- assert(writeQueue.size() == isInWriteQueue.size());
+
+ // log packet
+ logRequest(MemCtrl::WRITE, pkt->masterId(), pkt->qosValue(),
+ dram_pkt->addr, 1);
+
+ assert(totalWriteQueueSize == isInWriteQueue.size());
// Update stats
- avgWrQLen = writeQueue.size();
+ stats.avgWrQLen = totalWriteQueueSize;
// increment write entries of the rank
++dram_pkt->rankRef.writeEntries;
// keep track of the fact that this burst effectively
// disappeared as it was merged with an existing one
- mergedWrBursts++;
+ stats.mergedWrBursts++;
}
// Starting address of next dram pkt (aligend to burstSize boundary)
}
void
-DRAMCtrl::printQs() const {
+DRAMCtrl::printQs() const
+{
+#if TRACING_ON
DPRINTF(DRAM, "===READ QUEUE===\n\n");
- for (auto i = readQueue.begin() ; i != readQueue.end() ; ++i) {
- DPRINTF(DRAM, "Read %lu\n", (*i)->addr);
+ for (const auto& queue : readQueue) {
+ for (const auto& packet : queue) {
+ DPRINTF(DRAM, "Read %lu\n", packet->addr);
+ }
}
+
DPRINTF(DRAM, "\n===RESP QUEUE===\n\n");
- for (auto i = respQueue.begin() ; i != respQueue.end() ; ++i) {
- DPRINTF(DRAM, "Response %lu\n", (*i)->addr);
+ for (const auto& packet : respQueue) {
+ DPRINTF(DRAM, "Response %lu\n", packet->addr);
}
+
DPRINTF(DRAM, "\n===WRITE QUEUE===\n\n");
- for (auto i = writeQueue.begin() ; i != writeQueue.end() ; ++i) {
- DPRINTF(DRAM, "Write %lu\n", (*i)->addr);
+ for (const auto& queue : writeQueue) {
+ for (const auto& packet : queue) {
+ DPRINTF(DRAM, "Write %lu\n", packet->addr);
+ }
}
+#endif // TRACING_ON
}
bool
// Calc avg gap between requests
if (prevArrival != 0) {
- totGap += curTick() - prevArrival;
+ stats.totGap += curTick() - prevArrival;
}
prevArrival = curTick();
unsigned offset = pkt->getAddr() & (burstSize - 1);
unsigned int dram_pkt_count = divCeil(offset + size, burstSize);
+ // run the QoS scheduler and assign a QoS priority value to the packet
+ qosSchedule( { &readQueue, &writeQueue }, burstSize, pkt);
+
// check local buffers and do not accept if full
- if (pkt->isRead()) {
+ if (pkt->isWrite()) {
assert(size != 0);
- if (readQueueFull(dram_pkt_count)) {
- DPRINTF(DRAM, "Read queue full, not accepting\n");
+ if (writeQueueFull(dram_pkt_count)) {
+ DPRINTF(DRAM, "Write queue full, not accepting\n");
// remember that we have to retry this port
- retryRdReq = true;
- numRdRetry++;
+ retryWrReq = true;
+ stats.numWrRetry++;
return false;
} else {
- addToReadQueue(pkt, dram_pkt_count);
- readReqs++;
- bytesReadSys += size;
+ addToWriteQueue(pkt, dram_pkt_count);
+ stats.writeReqs++;
+ stats.bytesWrittenSys += size;
}
} else {
- assert(pkt->isWrite());
+ assert(pkt->isRead());
assert(size != 0);
- if (writeQueueFull(dram_pkt_count)) {
- DPRINTF(DRAM, "Write queue full, not accepting\n");
+ if (readQueueFull(dram_pkt_count)) {
+ DPRINTF(DRAM, "Read queue full, not accepting\n");
// remember that we have to retry this port
- retryWrReq = true;
- numWrRetry++;
+ retryRdReq = true;
+ stats.numRdRetry++;
return false;
} else {
- addToWriteQueue(pkt, dram_pkt_count);
- writeReqs++;
- bytesWrittenSys += size;
+ addToReadQueue(pkt, dram_pkt_count);
+ stats.readReqs++;
+ stats.bytesReadSys += size;
}
}
// track if this is the last packet before idling
// and that there are no outstanding commands to this rank
if (dram_pkt->rankRef.isQueueEmpty() &&
- dram_pkt->rankRef.outstandingEvents == 0) {
+ dram_pkt->rankRef.outstandingEvents == 0 && enableDRAMPowerdown) {
// verify that there are no events scheduled
assert(!dram_pkt->rankRef.activateEvent.scheduled());
assert(!dram_pkt->rankRef.prechargeEvent.scheduled());
} else {
// if there is nothing left in any queue, signal a drain
if (drainState() == DrainState::Draining &&
- writeQueue.empty() && readQueue.empty() && allRanksDrained()) {
+ !totalWriteQueueSize && !totalReadQueueSize && allRanksDrained()) {
DPRINTF(Drain, "DRAM controller done draining\n");
signalDrainDone();
}
}
-bool
-DRAMCtrl::chooseNext(std::deque<DRAMPacket*>& queue, Tick extra_col_delay)
+DRAMCtrl::DRAMPacketQueue::iterator
+DRAMCtrl::chooseNext(DRAMPacketQueue& queue, Tick extra_col_delay)
{
- // This method does the arbitration between requests. The chosen
- // packet is simply moved to the head of the queue. The other
- // methods know that this is the place to look. For example, with
- // FCFS, this method does nothing
- assert(!queue.empty());
-
- // bool to indicate if a packet to an available rank is found
- bool found_packet = false;
- if (queue.size() == 1) {
- DRAMPacket* dram_pkt = queue.front();
- // available rank corresponds to state refresh idle
- if (ranks[dram_pkt->rank]->inRefIdleState()) {
- found_packet = true;
- DPRINTF(DRAM, "Single request, going to a free rank\n");
- } else {
- DPRINTF(DRAM, "Single request, going to a busy rank\n");
- }
- return found_packet;
- }
+ // This method does the arbitration between requests.
- if (memSchedPolicy == Enums::fcfs) {
- // check if there is a packet going to a free rank
- for (auto i = queue.begin(); i != queue.end() ; ++i) {
- DRAMPacket* dram_pkt = *i;
+ DRAMCtrl::DRAMPacketQueue::iterator ret = queue.end();
+
+ if (!queue.empty()) {
+ if (queue.size() == 1) {
+ // available rank corresponds to state refresh idle
+ DRAMPacket* dram_pkt = *(queue.begin());
if (ranks[dram_pkt->rank]->inRefIdleState()) {
- queue.erase(i);
- queue.push_front(dram_pkt);
- found_packet = true;
- break;
+ ret = queue.begin();
+ DPRINTF(DRAM, "Single request, going to a free rank\n");
+ } else {
+ DPRINTF(DRAM, "Single request, going to a busy rank\n");
}
+ } else if (memSchedPolicy == Enums::fcfs) {
+ // check if there is a packet going to a free rank
+ for (auto i = queue.begin(); i != queue.end(); ++i) {
+ DRAMPacket* dram_pkt = *i;
+ if (ranks[dram_pkt->rank]->inRefIdleState()) {
+ ret = i;
+ break;
+ }
+ }
+ } else if (memSchedPolicy == Enums::frfcfs) {
+ ret = chooseNextFRFCFS(queue, extra_col_delay);
+ } else {
+ panic("No scheduling policy chosen\n");
}
- } else if (memSchedPolicy == Enums::frfcfs) {
- found_packet = reorderQueue(queue, extra_col_delay);
- } else
- panic("No scheduling policy chosen\n");
- return found_packet;
+ }
+ return ret;
}
-bool
-DRAMCtrl::reorderQueue(std::deque<DRAMPacket*>& queue, Tick extra_col_delay)
+DRAMCtrl::DRAMPacketQueue::iterator
+DRAMCtrl::chooseNextFRFCFS(DRAMPacketQueue& queue, Tick extra_col_delay)
{
// Only determine this if needed
vector<uint32_t> earliest_banks(ranksPerChannel, 0);
for (auto i = queue.begin(); i != queue.end() ; ++i) {
DRAMPacket* dram_pkt = *i;
const Bank& bank = dram_pkt->bankRef;
- const Tick col_allowed_at = dram_pkt->isRead ? bank.rdAllowedAt :
- bank.wrAllowedAt;
+ const Tick col_allowed_at = dram_pkt->isRead() ? bank.rdAllowedAt :
+ bank.wrAllowedAt;
+
+ DPRINTF(DRAM, "%s checking packet in bank %d\n",
+ __func__, dram_pkt->bankRef.bank);
// check if rank is not doing a refresh and thus is available, if not,
// jump to the next packet
if (dram_pkt->rankRef.inRefIdleState()) {
+
+ DPRINTF(DRAM,
+ "%s bank %d - Rank %d available\n", __func__,
+ dram_pkt->bankRef.bank, dram_pkt->rankRef.rank);
+
// check if it is a row hit
if (bank.openRow == dram_pkt->row) {
// no additional rank-to-rank or same bank-group
// commands that can issue seamlessly, without
// additional delay, such as same rank accesses
// and/or different bank-group accesses
- DPRINTF(DRAM, "Seamless row buffer hit\n");
+ DPRINTF(DRAM, "%s Seamless row buffer hit\n", __func__);
selected_pkt_it = i;
// no need to look through the remaining queue entries
break;
// the current one
selected_pkt_it = i;
found_prepped_pkt = true;
- DPRINTF(DRAM, "Prepped row buffer hit\n");
+ DPRINTF(DRAM, "%s Prepped row buffer hit\n", __func__);
}
} else if (!found_earliest_pkt) {
// if we have not initialised the bank status, do it
selected_pkt_it = i;
}
}
+ } else {
+ DPRINTF(DRAM, "%s bank %d - Rank %d not available\n", __func__,
+ dram_pkt->bankRef.bank, dram_pkt->rankRef.rank);
}
}
- if (selected_pkt_it != queue.end()) {
- DRAMPacket* selected_pkt = *selected_pkt_it;
- queue.erase(selected_pkt_it);
- queue.push_front(selected_pkt);
- return true;
+ if (selected_pkt_it == queue.end()) {
+ DPRINTF(DRAM, "%s no available ranks found\n", __func__);
}
- return false;
+ return selected_pkt_it;
}
void
// queue the packet in the response queue to be sent out after
// the static latency has passed
- port.schedTimingResp(pkt, response_time, true);
+ port.schedTimingResp(pkt, response_time);
} else {
// @todo the packet is going to be deleted, and the DRAMPacket
// is still having a pointer to it
// sample the bytes per activate here since we are closing
// the page
- bytesPerActivate.sample(bank.bytesAccessed);
+ stats.bytesPerActivate.sample(bank.bytesAccessed);
bank.openRow = Bank::NO_ROW;
}
// respect any constraints on the command (e.g. tRCD or tCCD)
- const Tick col_allowed_at = dram_pkt->isRead ?
+ const Tick col_allowed_at = dram_pkt->isRead() ?
bank.rdAllowedAt : bank.wrAllowedAt;
// we need to wait until the bus is available before we can issue
// tCCD_L is default requirement for same BG timing
// tCCD_L_WR is required for write-to-write
// Need to also take bus turnaround delays into account
- dly_to_rd_cmd = dram_pkt->isRead ?
+ dly_to_rd_cmd = dram_pkt->isRead() ?
tCCD_L : std::max(tCCD_L, wrToRdDly);
- dly_to_wr_cmd = dram_pkt->isRead ?
+ dly_to_wr_cmd = dram_pkt->isRead() ?
std::max(tCCD_L, rdToWrDly) : tCCD_L_WR;
} else {
// tBURST is default requirement for diff BG timing
// Need to also take bus turnaround delays into account
- dly_to_rd_cmd = dram_pkt->isRead ? tBURST : wrToRdDly;
- dly_to_wr_cmd = dram_pkt->isRead ? rdToWrDly : tBURST;
+ dly_to_rd_cmd = dram_pkt->isRead() ? tBURST : wrToRdDly;
+ dly_to_wr_cmd = dram_pkt->isRead() ? rdToWrDly : tBURST;
}
} else {
// different rank is by default in a different bank group and
// time before a precharge, in the case of a read, respect the
// read to precharge constraint
bank.preAllowedAt = std::max(bank.preAllowedAt,
- dram_pkt->isRead ? cmd_at + tRTP :
+ dram_pkt->isRead() ? cmd_at + tRTP :
dram_pkt->readyTime + tWR);
// increment the bytes accessed and the accesses per row
bool got_bank_conflict = false;
// either look at the read queue or write queue
- const deque<DRAMPacket*>& queue = dram_pkt->isRead ? readQueue :
- writeQueue;
- auto p = queue.begin();
- // make sure we are not considering the packet that we are
- // currently dealing with (which is the head of the queue)
- ++p;
-
- // keep on looking until we find a hit or reach the end of the queue
- // 1) if a hit is found, then both open and close adaptive policies keep
- // the page open
- // 2) if no hit is found, got_bank_conflict is set to true if a bank
- // conflict request is waiting in the queue
- while (!got_more_hits && p != queue.end()) {
- bool same_rank_bank = (dram_pkt->rank == (*p)->rank) &&
- (dram_pkt->bank == (*p)->bank);
- bool same_row = dram_pkt->row == (*p)->row;
- got_more_hits |= same_rank_bank && same_row;
- got_bank_conflict |= same_rank_bank && !same_row;
- ++p;
+ const std::vector<DRAMPacketQueue>& queue =
+ dram_pkt->isRead() ? readQueue : writeQueue;
+
+ for (uint8_t i = 0; i < numPriorities(); ++i) {
+ auto p = queue[i].begin();
+ // keep on looking until we find a hit or reach the end of the queue
+ // 1) if a hit is found, then both open and close adaptive policies keep
+ // the page open
+ // 2) if no hit is found, got_bank_conflict is set to true if a bank
+ // conflict request is waiting in the queue
+ // 3) make sure we are not considering the packet that we are
+ // currently dealing with
+ while (!got_more_hits && p != queue[i].end()) {
+ if (dram_pkt != (*p)) {
+ bool same_rank_bank = (dram_pkt->rank == (*p)->rank) &&
+ (dram_pkt->bank == (*p)->bank);
+
+ bool same_row = dram_pkt->row == (*p)->row;
+ got_more_hits |= same_rank_bank && same_row;
+ got_bank_conflict |= same_rank_bank && !same_row;
+ }
+ ++p;
+ }
+
+ if (got_more_hits)
+ break;
}
// auto pre-charge when either
}
// DRAMPower trace command to be written
- std::string mem_cmd = dram_pkt->isRead ? "RD" : "WR";
+ std::string mem_cmd = dram_pkt->isRead() ? "RD" : "WR";
// MemCommand required for DRAMPower library
MemCommand::cmds command = (mem_cmd == "RD") ? MemCommand::RD :
nextReqTime = nextBurstAt - (tRP + tRCD);
// Update the stats and schedule the next request
- if (dram_pkt->isRead) {
+ if (dram_pkt->isRead()) {
++readsThisTime;
if (row_hit)
- readRowHits++;
- bytesReadDRAM += burstSize;
- perBankRdBursts[dram_pkt->bankId]++;
+ stats.readRowHits++;
+ stats.bytesReadDRAM += burstSize;
+ stats.perBankRdBursts[dram_pkt->bankId]++;
// Update latency stats
- totMemAccLat += dram_pkt->readyTime - dram_pkt->entryTime;
- totBusLat += tBURST;
- totQLat += cmd_at - dram_pkt->entryTime;
+ stats.totMemAccLat += dram_pkt->readyTime - dram_pkt->entryTime;
+ stats.masterReadTotalLat[dram_pkt->masterId()] +=
+ dram_pkt->readyTime - dram_pkt->entryTime;
+
+ stats.totBusLat += tBURST;
+ stats.totQLat += cmd_at - dram_pkt->entryTime;
+ stats.masterReadBytes[dram_pkt->masterId()] += dram_pkt->size;
} else {
++writesThisTime;
if (row_hit)
- writeRowHits++;
- bytesWritten += burstSize;
- perBankWrBursts[dram_pkt->bankId]++;
+ stats.writeRowHits++;
+ stats.bytesWritten += burstSize;
+ stats.perBankWrBursts[dram_pkt->bankId]++;
+ stats.masterWriteBytes[dram_pkt->masterId()] += dram_pkt->size;
+ stats.masterWriteTotalLat[dram_pkt->masterId()] +=
+ dram_pkt->readyTime - dram_pkt->entryTime;
}
}
void
DRAMCtrl::processNextReqEvent()
{
+ // transition is handled by QoS algorithm if enabled
+ if (turnPolicy) {
+ // select bus state - only done if QoS algorithms are in use
+ busStateNext = selectNextBusState();
+ }
+
+ // detect bus state change
+ bool switched_cmd_type = (busState != busStateNext);
+ // record stats
+ recordTurnaroundStats();
+
+ DPRINTF(DRAM, "QoS Turnarounds selected state %s %s\n",
+ (busState==MemCtrl::READ)?"READ":"WRITE",
+ switched_cmd_type?"[turnaround triggered]":"");
+
+ if (switched_cmd_type) {
+ if (busState == READ) {
+ DPRINTF(DRAM,
+ "Switching to writes after %d reads with %d reads "
+ "waiting\n", readsThisTime, totalReadQueueSize);
+ stats.rdPerTurnAround.sample(readsThisTime);
+ readsThisTime = 0;
+ } else {
+ DPRINTF(DRAM,
+ "Switching to reads after %d writes with %d writes "
+ "waiting\n", writesThisTime, totalWriteQueueSize);
+ stats.wrPerTurnAround.sample(writesThisTime);
+ writesThisTime = 0;
+ }
+ }
+
+ // updates current state
+ busState = busStateNext;
+
+ // check ranks for refresh/wakeup - uses busStateNext, so done after turnaround
+ // decisions
int busyRanks = 0;
for (auto r : ranks) {
if (!r->inRefIdleState()) {
return;
}
- // pre-emptively set to false. Overwrite if in transitioning to
- // a new state
- bool switched_cmd_type = false;
- if (busState != busStateNext) {
- if (busState == READ) {
- DPRINTF(DRAM, "Switching to writes after %d reads with %d reads "
- "waiting\n", readsThisTime, readQueue.size());
-
- // sample and reset the read-related stats as we are now
- // transitioning to writes, and all reads are done
- rdPerTurnAround.sample(readsThisTime);
- readsThisTime = 0;
-
- // now proceed to do the actual writes
- switched_cmd_type = true;
- } else {
- DPRINTF(DRAM, "Switching to reads after %d writes with %d writes "
- "waiting\n", writesThisTime, writeQueue.size());
-
- wrPerTurnAround.sample(writesThisTime);
- writesThisTime = 0;
-
- switched_cmd_type = true;
- }
- // update busState to match next state until next transition
- busState = busStateNext;
- }
-
// when we get here it is either a read or a write
if (busState == READ) {
// track if we should switch or not
bool switch_to_writes = false;
- if (readQueue.empty()) {
+ if (totalReadQueueSize == 0) {
// In the case there is no read request to go next,
// trigger writes if we have passed the low threshold (or
// if we are draining)
- if (!writeQueue.empty() &&
+ if (!(totalWriteQueueSize == 0) &&
(drainState() == DrainState::Draining ||
- writeQueue.size() > writeLowThreshold)) {
+ totalWriteQueueSize > writeLowThreshold)) {
+ DPRINTF(DRAM, "Switching to writes due to read queue empty\n");
switch_to_writes = true;
} else {
// check if we are drained
return;
}
} else {
- // bool to check if there is a read to a free rank
- bool found_read = false;
- // Figure out which read request goes next, and move it to the
- // front of the read queue
- // If we are changing command type, incorporate the minimum
- // bus turnaround delay which will be tCS (different rank) case
- found_read = chooseNext(readQueue, switched_cmd_type ? tCS : 0);
+ bool read_found = false;
+ DRAMPacketQueue::iterator to_read;
+ uint8_t prio = numPriorities();
+
+ for (auto queue = readQueue.rbegin();
+ queue != readQueue.rend(); ++queue) {
+
+ prio--;
+
+ DPRINTF(QOS,
+ "DRAM controller checking READ queue [%d] priority [%d elements]\n",
+ prio, queue->size());
+
+ // Figure out which read request goes next
+ // If we are changing command type, incorporate the minimum
+ // bus turnaround delay which will be tCS (different rank) case
+ to_read = chooseNext((*queue), switched_cmd_type ? tCS : 0);
+
+ if (to_read != queue->end()) {
+ // candidate read found
+ read_found = true;
+ break;
+ }
+ }
// if no read to an available rank is found then return
// at this point. There could be writes to the available ranks
// which are above the required threshold. However, to
// avoid adding more complexity to the code, return and wait
// for a refresh event to kick things into action again.
- if (!found_read)
+ if (!read_found) {
+ DPRINTF(DRAM, "No Reads Found - exiting\n");
return;
+ }
+
+ auto dram_pkt = *to_read;
- DRAMPacket* dram_pkt = readQueue.front();
assert(dram_pkt->rankRef.inRefIdleState());
doDRAMAccess(dram_pkt);
- // At this point we're done dealing with the request
- readQueue.pop_front();
-
// Every respQueue which will generate an event, increment count
++dram_pkt->rankRef.outstandingEvents;
-
// sanity check
assert(dram_pkt->size <= burstSize);
assert(dram_pkt->readyTime >= curTick());
+ // log the response
+ logResponse(MemCtrl::READ, (*to_read)->masterId(),
+ dram_pkt->qosValue(), dram_pkt->getAddr(), 1,
+ dram_pkt->readyTime - dram_pkt->entryTime);
+
+
// Insert into response queue. It will be sent back to the
- // requestor at its readyTime
+ // requester at its readyTime
if (respQueue.empty()) {
assert(!respondEvent.scheduled());
schedule(respondEvent, dram_pkt->readyTime);
respQueue.push_back(dram_pkt);
// we have so many writes that we have to transition
- if (writeQueue.size() > writeHighThreshold) {
+ if (totalWriteQueueSize > writeHighThreshold) {
switch_to_writes = true;
}
+
+ // remove the request from the queue - the iterator is no longer valid .
+ readQueue[dram_pkt->qosValue()].erase(to_read);
}
// switching to writes, either because the read queue is empty
busStateNext = WRITE;
}
} else {
- // bool to check if write to free rank is found
- bool found_write = false;
- // If we are changing command type, incorporate the minimum
- // bus turnaround delay
- found_write = chooseNext(writeQueue,
- switched_cmd_type ? std::min(tRTW, tCS) : 0);
+ bool write_found = false;
+ DRAMPacketQueue::iterator to_write;
+ uint8_t prio = numPriorities();
+
+ for (auto queue = writeQueue.rbegin();
+ queue != writeQueue.rend(); ++queue) {
+
+ prio--;
+
+ DPRINTF(QOS,
+ "DRAM controller checking WRITE queue [%d] priority [%d elements]\n",
+ prio, queue->size());
+
+ // If we are changing command type, incorporate the minimum
+ // bus turnaround delay
+ to_write = chooseNext((*queue),
+ switched_cmd_type ? std::min(tRTW, tCS) : 0);
+
+ if (to_write != queue->end()) {
+ write_found = true;
+ break;
+ }
+ }
// if there are no writes to a rank that is available to service
// requests (i.e. rank is in refresh idle state) are found then
// return. There could be reads to the available ranks. However, to
// avoid adding more complexity to the code, return at this point and
// wait for a refresh event to kick things into action again.
- if (!found_write)
+ if (!write_found) {
+ DPRINTF(DRAM, "No Writes Found - exiting\n");
return;
+ }
+
+ auto dram_pkt = *to_write;
- DRAMPacket* dram_pkt = writeQueue.front();
assert(dram_pkt->rankRef.inRefIdleState());
// sanity check
assert(dram_pkt->size <= burstSize);
doDRAMAccess(dram_pkt);
- writeQueue.pop_front();
-
// removed write from queue, decrement count
--dram_pkt->rankRef.writeEntries;
++dram_pkt->rankRef.outstandingEvents;
} else if (dram_pkt->rankRef.writeDoneEvent.when() <
- dram_pkt-> readyTime) {
+ dram_pkt->readyTime) {
+
reschedule(dram_pkt->rankRef.writeDoneEvent, dram_pkt->readyTime);
}
isInWriteQueue.erase(burstAlign(dram_pkt->addr));
+
+ // log the response
+ logResponse(MemCtrl::WRITE, dram_pkt->masterId(),
+ dram_pkt->qosValue(), dram_pkt->getAddr(), 1,
+ dram_pkt->readyTime - dram_pkt->entryTime);
+
+
+ // remove the request from the queue - the iterator is no longer valid
+ writeQueue[dram_pkt->qosValue()].erase(to_write);
+
delete dram_pkt;
// If we emptied the write queue, or got sufficiently below the
// threshold (using the minWritesPerSwitch as the hysteresis) and
// are not draining, or we have reads waiting and have done enough
// writes, then switch to reads.
- if (writeQueue.empty() ||
- (writeQueue.size() + minWritesPerSwitch < writeLowThreshold &&
- drainState() != DrainState::Draining) ||
- (!readQueue.empty() && writesThisTime >= minWritesPerSwitch)) {
+ bool below_threshold =
+ totalWriteQueueSize + minWritesPerSwitch < writeLowThreshold;
+
+ if (totalWriteQueueSize == 0 ||
+ (below_threshold && drainState() != DrainState::Draining) ||
+ (totalReadQueueSize && writesThisTime >= minWritesPerSwitch)) {
+
// turn the bus back around for reads again
busStateNext = READ;
// them retry. This is done here to ensure that the retry does not
// cause a nextReqEvent to be scheduled before we do so as part of
// the next request processing
- if (retryWrReq && writeQueue.size() < writeBufferSize) {
+ if (retryWrReq && totalWriteQueueSize < writeBufferSize) {
retryWrReq = false;
port.sendRetryReq();
}
}
pair<vector<uint32_t>, bool>
-DRAMCtrl::minBankPrep(const deque<DRAMPacket*>& queue,
+DRAMCtrl::minBankPrep(const DRAMPacketQueue& queue,
Tick min_col_at) const
{
Tick min_act_at = MaxTick;
prechargeEvent([this]{ processPrechargeEvent(); }, name()),
refreshEvent([this]{ processRefreshEvent(); }, name()),
powerEvent([this]{ processPowerEvent(); }, name()),
- wakeUpEvent([this]{ processWakeUpEvent(); }, name())
+ wakeUpEvent([this]{ processWakeUpEvent(); }, name()),
+ stats(_memory, *this)
{
for (int b = 0; b < _p->banks_per_rank; b++) {
banks[b].bank = b;
if (numBanksActive == 0) {
// no reads to this rank in the Q and no pending
// RD/WR or refresh commands
- if (isQueueEmpty() && outstandingEvents == 0) {
+ if (isQueueEmpty() && outstandingEvents == 0 &&
+ memory.enableDRAMPowerdown) {
// should still be in ACT state since bank still open
assert(pwrState == PWR_ACT);
// Force PRE power-down if there are no outstanding commands
// in Q after refresh.
- } else if (isQueueEmpty()) {
+ } else if (isQueueEmpty() && memory.enableDRAMPowerdown) {
// still have refresh event outstanding but there should
// be no other events outstanding
assert(outstandingEvents == 1);
PowerState prev_state = pwrState;
// update the accounting
- pwrStateTime[prev_state] += duration;
+ stats.memoryStateTime[prev_state] += duration;
// track to total idle time
if ((prev_state == PWR_PRE_PDN) || (prev_state == PWR_ACT_PDN) ||
(prev_state == PWR_SREF)) {
- totalIdleTime += duration;
+ stats.totalIdleTime += duration;
}
pwrState = pwrStateTrans;
// will issue refresh immediately upon entry
if (pwrStatePostRefresh == PWR_PRE_PDN && isQueueEmpty() &&
(memory.drainState() != DrainState::Draining) &&
- (memory.drainState() != DrainState::Drained)) {
+ (memory.drainState() != DrainState::Drained) &&
+ memory.enableDRAMPowerdown) {
DPRINTF(DRAMState, "Rank %d bypassing refresh and transitioning "
"to self refresh at %11u tick\n", rank, curTick());
powerDownSleep(PWR_SREF, curTick());
// The energy components inside the power lib are calculated over
// the window so accumulate into the corresponding gem5 stat
- actEnergy += energy.act_energy * memory.devicesPerRank;
- preEnergy += energy.pre_energy * memory.devicesPerRank;
- readEnergy += energy.read_energy * memory.devicesPerRank;
- writeEnergy += energy.write_energy * memory.devicesPerRank;
- refreshEnergy += energy.ref_energy * memory.devicesPerRank;
- actBackEnergy += energy.act_stdby_energy * memory.devicesPerRank;
- preBackEnergy += energy.pre_stdby_energy * memory.devicesPerRank;
- actPowerDownEnergy += energy.f_act_pd_energy * memory.devicesPerRank;
- prePowerDownEnergy += energy.f_pre_pd_energy * memory.devicesPerRank;
- selfRefreshEnergy += energy.sref_energy * memory.devicesPerRank;
+ stats.actEnergy += energy.act_energy * memory.devicesPerRank;
+ stats.preEnergy += energy.pre_energy * memory.devicesPerRank;
+ stats.readEnergy += energy.read_energy * memory.devicesPerRank;
+ stats.writeEnergy += energy.write_energy * memory.devicesPerRank;
+ stats.refreshEnergy += energy.ref_energy * memory.devicesPerRank;
+ stats.actBackEnergy += energy.act_stdby_energy * memory.devicesPerRank;
+ stats.preBackEnergy += energy.pre_stdby_energy * memory.devicesPerRank;
+ stats.actPowerDownEnergy += energy.f_act_pd_energy * memory.devicesPerRank;
+ stats.prePowerDownEnergy += energy.f_pre_pd_energy * memory.devicesPerRank;
+ stats.selfRefreshEnergy += energy.sref_energy * memory.devicesPerRank;
// Accumulate window energy into the total energy.
- totalEnergy += energy.window_energy * memory.devicesPerRank;
+ stats.totalEnergy += energy.window_energy * memory.devicesPerRank;
// Average power must not be accumulated but calculated over the time
// since last stats reset. SimClock::Frequency is tick period not tick
// frequency.
// energy (pJ) 1e-9
// power (mW) = ----------- * ----------
// time (tick) tick_frequency
- averagePower = (totalEnergy.value() /
- (curTick() - memory.lastStatsResetTick)) *
- (SimClock::Frequency / 1000000000.0);
+ stats.averagePower = (stats.totalEnergy.value() /
+ (curTick() - memory.lastStatsResetTick)) *
+ (SimClock::Frequency / 1000000000.0);
}
void
updatePowerStats();
// final update of power state times
- pwrStateTime[pwrState] += (curTick() - pwrStateTick);
+ stats.memoryStateTime[pwrState] += (curTick() - pwrStateTick);
pwrStateTick = curTick();
-
}
void
}
-void
-DRAMCtrl::Rank::regStats()
+DRAMCtrl::DRAMStats::DRAMStats(DRAMCtrl &_dram)
+ : Stats::Group(&_dram),
+ dram(_dram),
+
+ ADD_STAT(readReqs, "Number of read requests accepted"),
+ ADD_STAT(writeReqs, "Number of write requests accepted"),
+
+ ADD_STAT(readBursts,
+ "Number of DRAM read bursts, "
+ "including those serviced by the write queue"),
+ ADD_STAT(writeBursts,
+ "Number of DRAM write bursts, "
+ "including those merged in the write queue"),
+ ADD_STAT(servicedByWrQ,
+ "Number of DRAM read bursts serviced by the write queue"),
+ ADD_STAT(mergedWrBursts,
+ "Number of DRAM write bursts merged with an existing one"),
+
+ ADD_STAT(neitherReadNorWriteReqs,
+ "Number of requests that are neither read nor write"),
+
+ ADD_STAT(perBankRdBursts, "Per bank write bursts"),
+ ADD_STAT(perBankWrBursts, "Per bank write bursts"),
+
+ ADD_STAT(avgRdQLen, "Average read queue length when enqueuing"),
+ ADD_STAT(avgWrQLen, "Average write queue length when enqueuing"),
+
+ ADD_STAT(totQLat, "Total ticks spent queuing"),
+ ADD_STAT(totBusLat, "Total ticks spent in databus transfers"),
+ ADD_STAT(totMemAccLat,
+ "Total ticks spent from burst creation until serviced "
+ "by the DRAM"),
+ ADD_STAT(avgQLat, "Average queueing delay per DRAM burst"),
+ ADD_STAT(avgBusLat, "Average bus latency per DRAM burst"),
+ ADD_STAT(avgMemAccLat, "Average memory access latency per DRAM burst"),
+
+ ADD_STAT(numRdRetry, "Number of times read queue was full causing retry"),
+ ADD_STAT(numWrRetry, "Number of times write queue was full causing retry"),
+
+ ADD_STAT(readRowHits, "Number of row buffer hits during reads"),
+ ADD_STAT(writeRowHits, "Number of row buffer hits during writes"),
+ ADD_STAT(readRowHitRate, "Row buffer hit rate for reads"),
+ ADD_STAT(writeRowHitRate, "Row buffer hit rate for writes"),
+
+ ADD_STAT(readPktSize, "Read request sizes (log2)"),
+ ADD_STAT(writePktSize, "Write request sizes (log2)"),
+
+ ADD_STAT(rdQLenPdf, "What read queue length does an incoming req see"),
+ ADD_STAT(wrQLenPdf, "What write queue length does an incoming req see"),
+
+ ADD_STAT(bytesPerActivate, "Bytes accessed per row activation"),
+
+ ADD_STAT(rdPerTurnAround,
+ "Reads before turning the bus around for writes"),
+ ADD_STAT(wrPerTurnAround,
+ "Writes before turning the bus around for reads"),
+
+ ADD_STAT(bytesReadDRAM, "Total number of bytes read from DRAM"),
+ ADD_STAT(bytesReadWrQ, "Total number of bytes read from write queue"),
+ ADD_STAT(bytesWritten, "Total number of bytes written to DRAM"),
+ ADD_STAT(bytesReadSys, "Total read bytes from the system interface side"),
+ ADD_STAT(bytesWrittenSys,
+ "Total written bytes from the system interface side"),
+
+ ADD_STAT(avgRdBW, "Average DRAM read bandwidth in MiByte/s"),
+ ADD_STAT(avgWrBW, "Average achieved write bandwidth in MiByte/s"),
+ ADD_STAT(avgRdBWSys, "Average system read bandwidth in MiByte/s"),
+ ADD_STAT(avgWrBWSys, "Average system write bandwidth in MiByte/s"),
+ ADD_STAT(peakBW, "Theoretical peak bandwidth in MiByte/s"),
+
+ ADD_STAT(busUtil, "Data bus utilization in percentage"),
+ ADD_STAT(busUtilRead, "Data bus utilization in percentage for reads"),
+ ADD_STAT(busUtilWrite, "Data bus utilization in percentage for writes"),
+
+ ADD_STAT(totGap, "Total gap between requests"),
+ ADD_STAT(avgGap, "Average gap between requests"),
+
+ ADD_STAT(masterReadBytes, "Per-master bytes read from memory"),
+ ADD_STAT(masterWriteBytes, "Per-master bytes write to memory"),
+ ADD_STAT(masterReadRate,
+ "Per-master bytes read from memory rate (Bytes/sec)"),
+ ADD_STAT(masterWriteRate,
+ "Per-master bytes write to memory rate (Bytes/sec)"),
+ ADD_STAT(masterReadAccesses,
+ "Per-master read serviced memory accesses"),
+ ADD_STAT(masterWriteAccesses,
+ "Per-master write serviced memory accesses"),
+ ADD_STAT(masterReadTotalLat,
+ "Per-master read total memory access latency"),
+ ADD_STAT(masterWriteTotalLat,
+ "Per-master write total memory access latency"),
+ ADD_STAT(masterReadAvgLat,
+ "Per-master read average memory access latency"),
+ ADD_STAT(masterWriteAvgLat,
+ "Per-master write average memory access latency"),
+
+ ADD_STAT(pageHitRate, "Row buffer hit rate, read and write combined")
{
- pwrStateTime
- .init(6)
- .name(name() + ".memoryStateTime")
- .desc("Time in different power states");
- pwrStateTime.subname(0, "IDLE");
- pwrStateTime.subname(1, "REF");
- pwrStateTime.subname(2, "SREF");
- pwrStateTime.subname(3, "PRE_PDN");
- pwrStateTime.subname(4, "ACT");
- pwrStateTime.subname(5, "ACT_PDN");
-
- actEnergy
- .name(name() + ".actEnergy")
- .desc("Energy for activate commands per rank (pJ)");
-
- preEnergy
- .name(name() + ".preEnergy")
- .desc("Energy for precharge commands per rank (pJ)");
-
- readEnergy
- .name(name() + ".readEnergy")
- .desc("Energy for read commands per rank (pJ)");
-
- writeEnergy
- .name(name() + ".writeEnergy")
- .desc("Energy for write commands per rank (pJ)");
-
- refreshEnergy
- .name(name() + ".refreshEnergy")
- .desc("Energy for refresh commands per rank (pJ)");
-
- actBackEnergy
- .name(name() + ".actBackEnergy")
- .desc("Energy for active background per rank (pJ)");
-
- preBackEnergy
- .name(name() + ".preBackEnergy")
- .desc("Energy for precharge background per rank (pJ)");
-
- actPowerDownEnergy
- .name(name() + ".actPowerDownEnergy")
- .desc("Energy for active power-down per rank (pJ)");
-
- prePowerDownEnergy
- .name(name() + ".prePowerDownEnergy")
- .desc("Energy for precharge power-down per rank (pJ)");
-
- selfRefreshEnergy
- .name(name() + ".selfRefreshEnergy")
- .desc("Energy for self refresh per rank (pJ)");
-
- totalEnergy
- .name(name() + ".totalEnergy")
- .desc("Total energy per rank (pJ)");
-
- averagePower
- .name(name() + ".averagePower")
- .desc("Core power per rank (mW)");
-
- totalIdleTime
- .name(name() + ".totalIdleTime")
- .desc("Total Idle time Per DRAM Rank");
-
- Stats::registerDumpCallback(new RankDumpCallback(this));
- Stats::registerResetCallback(new RankResetCallback(this));
}
+
void
-DRAMCtrl::regStats()
+DRAMCtrl::DRAMStats::regStats()
{
using namespace Stats;
- AbstractMemory::regStats();
-
- for (auto r : ranks) {
- r->regStats();
- }
-
- registerResetCallback(new MemResetCallback(this));
-
- readReqs
- .name(name() + ".readReqs")
- .desc("Number of read requests accepted");
-
- writeReqs
- .name(name() + ".writeReqs")
- .desc("Number of write requests accepted");
-
- readBursts
- .name(name() + ".readBursts")
- .desc("Number of DRAM read bursts, "
- "including those serviced by the write queue");
-
- writeBursts
- .name(name() + ".writeBursts")
- .desc("Number of DRAM write bursts, "
- "including those merged in the write queue");
-
- servicedByWrQ
- .name(name() + ".servicedByWrQ")
- .desc("Number of DRAM read bursts serviced by the write queue");
-
- mergedWrBursts
- .name(name() + ".mergedWrBursts")
- .desc("Number of DRAM write bursts merged with an existing one");
-
- neitherReadNorWrite
- .name(name() + ".neitherReadNorWriteReqs")
- .desc("Number of requests that are neither read nor write");
-
- perBankRdBursts
- .init(banksPerRank * ranksPerChannel)
- .name(name() + ".perBankRdBursts")
- .desc("Per bank write bursts");
-
- perBankWrBursts
- .init(banksPerRank * ranksPerChannel)
- .name(name() + ".perBankWrBursts")
- .desc("Per bank write bursts");
-
- avgRdQLen
- .name(name() + ".avgRdQLen")
- .desc("Average read queue length when enqueuing")
+ assert(dram._system);
+ const auto max_masters = dram._system->maxMasters();
+
+ perBankRdBursts.init(dram.banksPerRank * dram.ranksPerChannel);
+ perBankWrBursts.init(dram.banksPerRank * dram.ranksPerChannel);
+
+ avgRdQLen.precision(2);
+ avgWrQLen.precision(2);
+ avgQLat.precision(2);
+ avgBusLat.precision(2);
+ avgMemAccLat.precision(2);
+
+ readRowHitRate.precision(2);
+ writeRowHitRate.precision(2);
+
+ readPktSize.init(ceilLog2(dram.burstSize) + 1);
+ writePktSize.init(ceilLog2(dram.burstSize) + 1);
+
+ rdQLenPdf.init(dram.readBufferSize);
+ wrQLenPdf.init(dram.writeBufferSize);
+
+ bytesPerActivate
+ .init(dram.maxAccessesPerRow ?
+ dram.maxAccessesPerRow : dram.rowBufferSize)
+ .flags(nozero);
+
+ rdPerTurnAround
+ .init(dram.readBufferSize)
+ .flags(nozero);
+ wrPerTurnAround
+ .init(dram.writeBufferSize)
+ .flags(nozero);
+
+ avgRdBW.precision(2);
+ avgWrBW.precision(2);
+ avgRdBWSys.precision(2);
+ avgWrBWSys.precision(2);
+ peakBW.precision(2);
+ busUtil.precision(2);
+ avgGap.precision(2);
+ busUtilWrite.precision(2);
+ pageHitRate.precision(2);
+
+
+ // per-master bytes read and written to memory
+ masterReadBytes
+ .init(max_masters)
+ .flags(nozero | nonan);
+
+ masterWriteBytes
+ .init(max_masters)
+ .flags(nozero | nonan);
+
+ // per-master bytes read and written to memory rate
+ masterReadRate
+ .flags(nozero | nonan)
+ .precision(12);
+
+ masterReadAccesses
+ .init(max_masters)
+ .flags(nozero);
+
+ masterWriteAccesses
+ .init(max_masters)
+ .flags(nozero);
+
+ masterReadTotalLat
+ .init(max_masters)
+ .flags(nozero | nonan);
+
+ masterReadAvgLat
+ .flags(nonan)
.precision(2);
- avgWrQLen
- .name(name() + ".avgWrQLen")
- .desc("Average write queue length when enqueuing")
- .precision(2);
- totQLat
- .name(name() + ".totQLat")
- .desc("Total ticks spent queuing");
+ busUtilRead
+ .precision(2);
- totBusLat
- .name(name() + ".totBusLat")
- .desc("Total ticks spent in databus transfers");
+ masterWriteRate
+ .flags(nozero | nonan)
+ .precision(12);
- totMemAccLat
- .name(name() + ".totMemAccLat")
- .desc("Total ticks spent from burst creation until serviced "
- "by the DRAM");
+ masterWriteTotalLat
+ .init(max_masters)
+ .flags(nozero | nonan);
- avgQLat
- .name(name() + ".avgQLat")
- .desc("Average queueing delay per DRAM burst")
+ masterWriteAvgLat
+ .flags(nonan)
.precision(2);
+ for (int i = 0; i < max_masters; i++) {
+ const std::string master = dram._system->getMasterName(i);
+ masterReadBytes.subname(i, master);
+ masterReadRate.subname(i, master);
+ masterWriteBytes.subname(i, master);
+ masterWriteRate.subname(i, master);
+ masterReadAccesses.subname(i, master);
+ masterWriteAccesses.subname(i, master);
+ masterReadTotalLat.subname(i, master);
+ masterReadAvgLat.subname(i, master);
+ masterWriteTotalLat.subname(i, master);
+ masterWriteAvgLat.subname(i, master);
+ }
+
+ // Formula stats
avgQLat = totQLat / (readBursts - servicedByWrQ);
-
- avgBusLat
- .name(name() + ".avgBusLat")
- .desc("Average bus latency per DRAM burst")
- .precision(2);
-
avgBusLat = totBusLat / (readBursts - servicedByWrQ);
-
- avgMemAccLat
- .name(name() + ".avgMemAccLat")
- .desc("Average memory access latency per DRAM burst")
- .precision(2);
-
avgMemAccLat = totMemAccLat / (readBursts - servicedByWrQ);
- numRdRetry
- .name(name() + ".numRdRetry")
- .desc("Number of times read queue was full causing retry");
-
- numWrRetry
- .name(name() + ".numWrRetry")
- .desc("Number of times write queue was full causing retry");
-
- readRowHits
- .name(name() + ".readRowHits")
- .desc("Number of row buffer hits during reads");
-
- writeRowHits
- .name(name() + ".writeRowHits")
- .desc("Number of row buffer hits during writes");
-
- readRowHitRate
- .name(name() + ".readRowHitRate")
- .desc("Row buffer hit rate for reads")
- .precision(2);
-
readRowHitRate = (readRowHits / (readBursts - servicedByWrQ)) * 100;
-
- writeRowHitRate
- .name(name() + ".writeRowHitRate")
- .desc("Row buffer hit rate for writes")
- .precision(2);
-
writeRowHitRate = (writeRowHits / (writeBursts - mergedWrBursts)) * 100;
- readPktSize
- .init(ceilLog2(burstSize) + 1)
- .name(name() + ".readPktSize")
- .desc("Read request sizes (log2)");
-
- writePktSize
- .init(ceilLog2(burstSize) + 1)
- .name(name() + ".writePktSize")
- .desc("Write request sizes (log2)");
-
- rdQLenPdf
- .init(readBufferSize)
- .name(name() + ".rdQLenPdf")
- .desc("What read queue length does an incoming req see");
-
- wrQLenPdf
- .init(writeBufferSize)
- .name(name() + ".wrQLenPdf")
- .desc("What write queue length does an incoming req see");
-
- bytesPerActivate
- .init(maxAccessesPerRow)
- .name(name() + ".bytesPerActivate")
- .desc("Bytes accessed per row activation")
- .flags(nozero);
-
- rdPerTurnAround
- .init(readBufferSize)
- .name(name() + ".rdPerTurnAround")
- .desc("Reads before turning the bus around for writes")
- .flags(nozero);
-
- wrPerTurnAround
- .init(writeBufferSize)
- .name(name() + ".wrPerTurnAround")
- .desc("Writes before turning the bus around for reads")
- .flags(nozero);
-
- bytesReadDRAM
- .name(name() + ".bytesReadDRAM")
- .desc("Total number of bytes read from DRAM");
-
- bytesReadWrQ
- .name(name() + ".bytesReadWrQ")
- .desc("Total number of bytes read from write queue");
-
- bytesWritten
- .name(name() + ".bytesWritten")
- .desc("Total number of bytes written to DRAM");
-
- bytesReadSys
- .name(name() + ".bytesReadSys")
- .desc("Total read bytes from the system interface side");
-
- bytesWrittenSys
- .name(name() + ".bytesWrittenSys")
- .desc("Total written bytes from the system interface side");
-
- avgRdBW
- .name(name() + ".avgRdBW")
- .desc("Average DRAM read bandwidth in MiByte/s")
- .precision(2);
-
avgRdBW = (bytesReadDRAM / 1000000) / simSeconds;
-
- avgWrBW
- .name(name() + ".avgWrBW")
- .desc("Average achieved write bandwidth in MiByte/s")
- .precision(2);
-
avgWrBW = (bytesWritten / 1000000) / simSeconds;
-
- avgRdBWSys
- .name(name() + ".avgRdBWSys")
- .desc("Average system read bandwidth in MiByte/s")
- .precision(2);
-
avgRdBWSys = (bytesReadSys / 1000000) / simSeconds;
-
- avgWrBWSys
- .name(name() + ".avgWrBWSys")
- .desc("Average system write bandwidth in MiByte/s")
- .precision(2);
-
avgWrBWSys = (bytesWrittenSys / 1000000) / simSeconds;
+ peakBW = (SimClock::Frequency / dram.tBURST) * dram.burstSize / 1000000;
- peakBW
- .name(name() + ".peakBW")
- .desc("Theoretical peak bandwidth in MiByte/s")
- .precision(2);
+ busUtil = (avgRdBW + avgWrBW) / peakBW * 100;
- peakBW = (SimClock::Frequency / tBURST) * burstSize / 1000000;
+ avgGap = totGap / (readReqs + writeReqs);
- busUtil
- .name(name() + ".busUtil")
- .desc("Data bus utilization in percentage")
- .precision(2);
- busUtil = (avgRdBW + avgWrBW) / peakBW * 100;
+ busUtilRead = avgRdBW / peakBW * 100;
+ busUtilWrite = avgWrBW / peakBW * 100;
- totGap
- .name(name() + ".totGap")
- .desc("Total gap between requests");
+ pageHitRate = (writeRowHits + readRowHits) /
+ (writeBursts - mergedWrBursts + readBursts - servicedByWrQ) * 100;
- avgGap
- .name(name() + ".avgGap")
- .desc("Average gap between requests")
- .precision(2);
+ masterReadRate = masterReadBytes / simSeconds;
+ masterWriteRate = masterWriteBytes / simSeconds;
+ masterReadAvgLat = masterReadTotalLat / masterReadAccesses;
+ masterWriteAvgLat = masterWriteTotalLat / masterWriteAccesses;
+}
- avgGap = totGap / (readReqs + writeReqs);
+void
+DRAMCtrl::DRAMStats::resetStats()
+{
+ dram.lastStatsResetTick = curTick();
+}
- // Stats for DRAM Power calculation based on Micron datasheet
- busUtilRead
- .name(name() + ".busUtilRead")
- .desc("Data bus utilization in percentage for reads")
- .precision(2);
+DRAMCtrl::RankStats::RankStats(DRAMCtrl &_memory, Rank &_rank)
+ : Stats::Group(&_memory, csprintf("rank%d", _rank.rank).c_str()),
+ rank(_rank),
+
+ ADD_STAT(actEnergy, "Energy for activate commands per rank (pJ)"),
+ ADD_STAT(preEnergy, "Energy for precharge commands per rank (pJ)"),
+ ADD_STAT(readEnergy, "Energy for read commands per rank (pJ)"),
+ ADD_STAT(writeEnergy, "Energy for write commands per rank (pJ)"),
+ ADD_STAT(refreshEnergy, "Energy for refresh commands per rank (pJ)"),
+ ADD_STAT(actBackEnergy, "Energy for active background per rank (pJ)"),
+ ADD_STAT(preBackEnergy, "Energy for precharge background per rank (pJ)"),
+ ADD_STAT(actPowerDownEnergy,
+ "Energy for active power-down per rank (pJ)"),
+ ADD_STAT(prePowerDownEnergy,
+ "Energy for precharge power-down per rank (pJ)"),
+ ADD_STAT(selfRefreshEnergy, "Energy for self refresh per rank (pJ)"),
+
+ ADD_STAT(totalEnergy, "Total energy per rank (pJ)"),
+ ADD_STAT(averagePower, "Core power per rank (mW)"),
+
+ ADD_STAT(totalIdleTime, "Total Idle time Per DRAM Rank"),
+ ADD_STAT(memoryStateTime, "Time in different power states")
+{
+}
- busUtilRead = avgRdBW / peakBW * 100;
+void
+DRAMCtrl::RankStats::regStats()
+{
+ Stats::Group::regStats();
+
+ memoryStateTime.init(6);
+ memoryStateTime.subname(0, "IDLE");
+ memoryStateTime.subname(1, "REF");
+ memoryStateTime.subname(2, "SREF");
+ memoryStateTime.subname(3, "PRE_PDN");
+ memoryStateTime.subname(4, "ACT");
+ memoryStateTime.subname(5, "ACT_PDN");
+}
- busUtilWrite
- .name(name() + ".busUtilWrite")
- .desc("Data bus utilization in percentage for writes")
- .precision(2);
+void
+DRAMCtrl::RankStats::resetStats()
+{
+ Stats::Group::resetStats();
- busUtilWrite = avgWrBW / peakBW * 100;
+ rank.resetStats();
+}
- pageHitRate
- .name(name() + ".pageHitRate")
- .desc("Row buffer hit rate, read and write combined")
- .precision(2);
+void
+DRAMCtrl::RankStats::preDumpStats()
+{
+ Stats::Group::preDumpStats();
- pageHitRate = (writeRowHits + readRowHits) /
- (writeBursts - mergedWrBursts + readBursts - servicedByWrQ) * 100;
+ rank.computeStats();
}
void
functionalAccess(pkt);
}
-BaseSlavePort&
-DRAMCtrl::getSlavePort(const string &if_name, PortID idx)
+Port &
+DRAMCtrl::getPort(const string &if_name, PortID idx)
{
if (if_name != "port") {
- return MemObject::getSlavePort(if_name, idx);
+ return QoS::MemCtrl::getPort(if_name, idx);
} else {
return port;
}
{
// if there is anything in any of our internal queues, keep track
// of that as well
- if (!(writeQueue.empty() && readQueue.empty() && respQueue.empty() &&
+ if (!(!totalWriteQueueSize && !totalReadQueueSize && respQueue.empty() &&
allRanksDrained())) {
DPRINTF(Drain, "DRAM controller not drained, write: %d, read: %d,"
- " resp: %d\n", writeQueue.size(), readQueue.size(),
+ " resp: %d\n", totalWriteQueueSize, totalReadQueueSize,
respQueue.size());
// the only queue that is not drained automatically over time
// is the write queue, thus kick things into action if needed
- if (!writeQueue.empty() && !nextReqEvent.scheduled()) {
+ if (!totalWriteQueueSize && !nextReqEvent.scheduled()) {
schedule(nextReqEvent, curTick());
}
}
DRAMCtrl::MemoryPort::MemoryPort(const std::string& name, DRAMCtrl& _memory)
- : QueuedSlavePort(name, &_memory, queue), queue(_memory, *this),
+ : QueuedSlavePort(name, &_memory, queue), queue(_memory, *this, true),
memory(_memory)
{ }
{
pkt->pushLabel(memory.name());
- if (!queue.checkFunctional(pkt)) {
+ if (!queue.trySatisfyFunctional(pkt)) {
// Default implementation of SimpleTimingPort::recvFunctional()
// calls recvAtomic() and throws away the latency; we can save a
// little here by just not calculating the latency.