/*
- * Copyright (c) 2010-2014 ARM Limited
+ * Copyright (c) 2010-2019 ARM Limited
* All rights reserved
*
* The license below extends only to copyright in the software and shall
* Authors: Andreas Hansson
* Ani Udipi
* Neha Agarwal
+ * Omar Naji
+ * Wendy Elsasser
+ * Radhika Jagtap
*/
+#include "mem/dram_ctrl.hh"
+
#include "base/bitfield.hh"
#include "base/trace.hh"
#include "debug/DRAM.hh"
#include "debug/DRAMPower.hh"
#include "debug/DRAMState.hh"
#include "debug/Drain.hh"
-#include "mem/dram_ctrl.hh"
+#include "debug/QOS.hh"
#include "sim/system.hh"
using namespace std;
using namespace Data;
DRAMCtrl::DRAMCtrl(const DRAMCtrlParams* p) :
- AbstractMemory(p),
- port(name() + ".port", *this),
+ QoS::MemCtrl(p),
+ port(name() + ".port", *this), isTimingMode(false),
retryRdReq(false), retryWrReq(false),
- busState(READ),
- nextReqEvent(this), respondEvent(this), activateEvent(this),
- prechargeEvent(this), refreshEvent(this), powerEvent(this),
- drainManager(NULL),
+ nextReqEvent([this]{ processNextReqEvent(); }, name()),
+ respondEvent([this]{ processRespondEvent(); }, name()),
deviceSize(p->device_size),
deviceBusWidth(p->device_bus_width), burstLength(p->burst_length),
deviceRowBufferSize(p->device_rowbuffer_size),
burstSize((devicesPerRank * burstLength * deviceBusWidth) / 8),
rowBufferSize(devicesPerRank * deviceRowBufferSize),
columnsPerRowBuffer(rowBufferSize / burstSize),
- columnsPerStripe(range.granularity() / burstSize),
+ columnsPerStripe(range.interleaved() ? range.granularity() / burstSize : 1),
ranksPerChannel(p->ranks_per_channel),
bankGroupsPerRank(p->bank_groups_per_rank),
bankGroupArch(p->bank_groups_per_rank > 0),
writeLowThreshold(writeBufferSize * p->write_low_thresh_perc / 100.0),
minWritesPerSwitch(p->min_writes_per_switch),
writesThisTime(0), readsThisTime(0),
- tCK(p->tCK), tWTR(p->tWTR), tRTW(p->tRTW), tCS(p->tCS), tBURST(p->tBURST),
+ tCK(p->tCK), tRTW(p->tRTW), tCS(p->tCS), tBURST(p->tBURST),
+ tCCD_L_WR(p->tCCD_L_WR),
tCCD_L(p->tCCD_L), tRCD(p->tRCD), tCL(p->tCL), tRP(p->tRP), tRAS(p->tRAS),
tWR(p->tWR), tRTP(p->tRTP), tRFC(p->tRFC), tREFI(p->tREFI), tRRD(p->tRRD),
- tRRD_L(p->tRRD_L), tXAW(p->tXAW), activationLimit(p->activation_limit),
+ tRRD_L(p->tRRD_L), tXAW(p->tXAW), tXP(p->tXP), tXS(p->tXS),
+ activationLimit(p->activation_limit), rankToRankDly(tCS + tBURST),
+ wrToRdDly(tCL + tBURST + p->tWTR), rdToWrDly(tRTW + tBURST),
memSchedPolicy(p->mem_sched_policy), addrMapping(p->addr_mapping),
pageMgmt(p->page_policy),
maxAccessesPerRow(p->max_accesses_per_row),
frontendLatency(p->static_frontend_latency),
backendLatency(p->static_backend_latency),
- busBusyUntil(0), refreshDueAt(0), refreshState(REF_IDLE),
- pwrStateTrans(PWR_IDLE), pwrState(PWR_IDLE), prevArrival(0),
- nextReqTime(0), pwrStateTick(0), numBanksActive(0),
- activeRank(0), timeStampOffset(0)
+ nextBurstAt(0), prevArrival(0),
+ nextReqTime(0),
+ stats(*this),
+ activeRank(0), timeStampOffset(0),
+ lastStatsResetTick(0), enableDRAMPowerdown(p->enable_dram_powerdown)
{
- // create the bank states based on the dimensions of the ranks and
- // banks
- banks.resize(ranksPerChannel);
+ // sanity check the ranks since we rely on bit slicing for the
+ // address decoding
+ fatal_if(!isPowerOf2(ranksPerChannel), "DRAM rank count of %d is not "
+ "allowed, must be a power of two\n", ranksPerChannel);
+
+ 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);
+
- //create list of drampower objects. For each rank 1 drampower instance.
for (int i = 0; i < ranksPerChannel; i++) {
- DRAMPower drampower = DRAMPower(p, false);
- rankPower.emplace_back(drampower);
- }
-
- actTicks.resize(ranksPerChannel);
- for (size_t c = 0; c < ranksPerChannel; ++c) {
- banks[c].resize(banksPerRank);
- actTicks[c].resize(activationLimit, 0);
- }
-
- // set the bank indices
- for (int r = 0; r < ranksPerChannel; r++) {
- for (int b = 0; b < banksPerRank; b++) {
- banks[r][b].rank = r;
- banks[r][b].bank = b;
- // GDDR addressing of banks to BG is linear.
- // Here we assume that all DRAM generations address bank groups as
- // follows:
- if (bankGroupArch) {
- // Simply assign lower bits to bank group in order to
- // rotate across bank groups as banks are incremented
- // e.g. with 4 banks per bank group and 16 banks total:
- // banks 0,4,8,12 are in bank group 0
- // banks 1,5,9,13 are in bank group 1
- // banks 2,6,10,14 are in bank group 2
- // banks 3,7,11,15 are in bank group 3
- banks[r][b].bankgr = b % bankGroupsPerRank;
- } else {
- // No bank groups; simply assign to bank number
- banks[r][b].bankgr = b;
- }
- }
+ Rank* rank = new Rank(*this, p, i);
+ ranks.push_back(rank);
}
// perform a basic check of the write thresholds
rowsPerBank = capacity / (rowBufferSize * banksPerRank * ranksPerChannel);
- // a bit of sanity checks on the interleaving
- if (range.interleaved()) {
- if (channels != range.stripes())
- fatal("%s has %d interleaved address stripes but %d channel(s)\n",
- name(), range.stripes(), channels);
-
- if (addrMapping == Enums::RoRaBaChCo) {
- if (rowBufferSize != range.granularity()) {
- fatal("Channel interleaving of %s doesn't match RoRaBaChCo "
- "address map\n", name());
- }
- } else if (addrMapping == Enums::RoRaBaCoCh ||
- addrMapping == Enums::RoCoRaBaCh) {
- // for the interleavings with channel bits in the bottom,
- // if the system uses a channel striping granularity that
- // is larger than the DRAM burst size, then map the
- // sequential accesses within a stripe to a number of
- // columns in the DRAM, effectively placing some of the
- // lower-order column bits as the least-significant bits
- // of the address (above the ones denoting the burst size)
- assert(columnsPerStripe >= 1);
-
- // channel striping has to be done at a granularity that
- // is equal or larger to a cache line
- if (system()->cacheLineSize() > range.granularity()) {
- fatal("Channel interleaving of %s must be at least as large "
- "as the cache line size\n", name());
- }
-
- // ...and equal or smaller than the row-buffer size
- if (rowBufferSize < range.granularity()) {
- fatal("Channel interleaving of %s must be at most as large "
- "as the row-buffer size\n", name());
- }
- // this is essentially the check above, so just to be sure
- assert(columnsPerStripe <= columnsPerRowBuffer);
- }
- }
-
// some basic sanity checks
if (tREFI <= tRP || tREFI <= tRFC) {
fatal("tREFI (%d) must be larger than tRP (%d) and tRFC (%d)\n",
"bank groups per rank (%d) is greater than 1\n",
tCCD_L, tBURST, bankGroupsPerRank);
}
+ // tCCD_L_WR should be greater than minimal, back-to-back burst delay
+ if (tCCD_L_WR <= tBURST) {
+ fatal("tCCD_L_WR (%d) should be larger than tBURST (%d) when "
+ "bank groups per rank (%d) is greater than 1\n",
+ tCCD_L_WR, tBURST, bankGroupsPerRank);
+ }
// tRRD_L is greater than minimal, same bank group ACT-to-ACT delay
// some datasheets might specify it equal to tRRD
if (tRRD_L < tRRD) {
void
DRAMCtrl::init()
{
- AbstractMemory::init();
+ MemCtrl::init();
if (!port.isConnected()) {
fatal("DRAMCtrl %s is unconnected!\n", name());
} else {
port.sendRangeChange();
}
+
+ // a bit of sanity checks on the interleaving, save it for here to
+ // ensure that the system pointer is initialised
+ if (range.interleaved()) {
+ if (channels != range.stripes())
+ fatal("%s has %d interleaved address stripes but %d channel(s)\n",
+ name(), range.stripes(), channels);
+
+ if (addrMapping == Enums::RoRaBaChCo) {
+ if (rowBufferSize != range.granularity()) {
+ fatal("Channel interleaving of %s doesn't match RoRaBaChCo "
+ "address map\n", name());
+ }
+ } else if (addrMapping == Enums::RoRaBaCoCh ||
+ addrMapping == Enums::RoCoRaBaCh) {
+ // for the interleavings with channel bits in the bottom,
+ // if the system uses a channel striping granularity that
+ // is larger than the DRAM burst size, then map the
+ // sequential accesses within a stripe to a number of
+ // columns in the DRAM, effectively placing some of the
+ // lower-order column bits as the least-significant bits
+ // of the address (above the ones denoting the burst size)
+ assert(columnsPerStripe >= 1);
+
+ // channel striping has to be done at a granularity that
+ // is equal or larger to a cache line
+ if (system()->cacheLineSize() > range.granularity()) {
+ fatal("Channel interleaving of %s must be at least as large "
+ "as the cache line size\n", name());
+ }
+
+ // ...and equal or smaller than the row-buffer size
+ if (rowBufferSize < range.granularity()) {
+ fatal("Channel interleaving of %s must be at most as large "
+ "as the row-buffer size\n", name());
+ }
+ // this is essentially the check above, so just to be sure
+ assert(columnsPerStripe <= columnsPerRowBuffer);
+ }
+ }
}
void
DRAMCtrl::startup()
{
- // timestamp offset should be in clock cycles for DRAMPower
- timeStampOffset = divCeil(curTick(), tCK);
- // update the start tick for the precharge accounting to the
- // current tick
- pwrStateTick = curTick();
+ // remember the memory system mode of operation
+ isTimingMode = system()->isTimingMode();
- // shift the bus busy time sufficiently far ahead that we never
- // have to worry about negative values when computing the time for
- // the next request, this will add an insignificant bubble at the
- // start of simulation
- busBusyUntil = curTick() + tRP + tRCD + tCL;
+ if (isTimingMode) {
+ // timestamp offset should be in clock cycles for DRAMPower
+ timeStampOffset = divCeil(curTick(), tCK);
- // kick off the refresh, and give ourselves enough time to
- // precharge
- schedule(refreshEvent, curTick() + tREFI - tRP);
+ // update the start tick for the precharge accounting to the
+ // current tick
+ for (auto r : ranks) {
+ r->startup(curTick() + tREFI - tRP);
+ }
+
+ // shift the bus busy time sufficiently far ahead that we never
+ // have to worry about negative values when computing the time for
+ // the next request, this will add an insignificant bubble at the
+ // start of simulation
+ nextBurstAt = curTick() + tRP + tRCD;
+ }
}
Tick
{
DPRINTF(DRAM, "recvAtomic: %s 0x%x\n", pkt->cmdString(), pkt->getAddr());
+ panic_if(pkt->cacheResponding(), "Should not see packets where cache "
+ "is responding");
+
// do the actual memory access and turn the packet into a response
access(pkt);
Tick latency = 0;
- if (!pkt->memInhibitAsserted() && pkt->hasData()) {
+ if (pkt->hasData()) {
// this value is not supposed to be accurate, just enough to
// keep things going, mimic a closed page
latency = tRP + tRCD + tCL;
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
rank = addr % ranksPerChannel;
addr = addr / ranksPerChannel;
- // lastly, get the row bits
+ // lastly, get the row bits, no need to remove them from addr
row = addr % rowsPerBank;
- addr = addr / rowsPerBank;
} else if (addrMapping == Enums::RoRaBaCoCh) {
// take out the lower-order column bits
addr = addr / columnsPerStripe;
rank = addr % ranksPerChannel;
addr = addr / ranksPerChannel;
- // lastly, get the row bits
+ // lastly, get the row bits, no need to remove them from addr
row = addr % rowsPerBank;
- addr = addr / rowsPerBank;
} else if (addrMapping == Enums::RoCoRaBaCh) {
// optimise for closed page mode and utilise maximum
// parallelism of the DRAM (at the cost of power)
// next, the higher-order column bites
addr = addr / (columnsPerRowBuffer / columnsPerStripe);
- // lastly, get the row bits
+ // lastly, get the row bits, no need to remove them from addr
row = addr % rowsPerBank;
- addr = addr / rowsPerBank;
} else
panic("Unknown address mapping policy chosen!");
// later
uint16_t bank_id = banksPerRank * rank + bank;
return new DRAMPacket(pkt, isRead, rank, bank, row, bank_id, dramPktAddr,
- size, banks[rank][bank]);
+ size, ranks[rank]->banks[bank], *ranks[rank]);
}
void
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
bool foundInWrQ = false;
- for (auto i = writeQueue.begin(); i != writeQueue.end(); ++i) {
- // check if the read is subsumed in the write entry we are
- // looking at
- if ((*i)->addr <= addr &&
- (addr + size) <= ((*i)->addr + (*i)->size)) {
- foundInWrQ = true;
- servicedByWrQ++;
- pktsServicedByWrQ++;
- DPRINTF(DRAM, "Read to addr %lld with size %d serviced by "
- "write queue\n", addr, size);
- bytesReadWrQ += burstSize;
- break;
+ Addr burst_addr = burstAlign(addr);
+ // 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& 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);
+
+ ++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 so we
- // can stop at that point and also avoid enqueueing a new
- // request
- bool merged = false;
- auto w = writeQueue.begin();
-
- while(!merged && w != writeQueue.end()) {
- // either of the two could be first, if they are the same
- // it does not matter which way we go
- if ((*w)->addr >= addr) {
- // the existing one starts after the new one, figure
- // out where the new one ends with respect to the
- // existing one
- if ((addr + size) >= ((*w)->addr + (*w)->size)) {
- // check if the existing one is completely
- // subsumed in the new one
- DPRINTF(DRAM, "Merging write covering existing burst\n");
- merged = true;
- // update both the address and the size
- (*w)->addr = addr;
- (*w)->size = size;
- } else if ((addr + size) >= (*w)->addr &&
- ((*w)->addr + (*w)->size - addr) <= burstSize) {
- // the new one is just before or partially
- // overlapping with the existing one, and together
- // they fit within a burst
- DPRINTF(DRAM, "Merging write before existing burst\n");
- merged = true;
- // the existing queue item needs to be adjusted with
- // respect to both address and size
- (*w)->size = (*w)->addr + (*w)->size - addr;
- (*w)->addr = addr;
- }
- } else {
- // the new one starts after the current one, figure
- // out where the existing one ends with respect to the
- // new one
- if (((*w)->addr + (*w)->size) >= (addr + size)) {
- // check if the new one is completely subsumed in the
- // existing one
- DPRINTF(DRAM, "Merging write into existing burst\n");
- merged = true;
- // no adjustments necessary
- } else if (((*w)->addr + (*w)->size) >= addr &&
- (addr + size - (*w)->addr) <= burstSize) {
- // the existing one is just before or partially
- // overlapping with the new one, and together
- // they fit within a burst
- DPRINTF(DRAM, "Merging write after existing burst\n");
- merged = true;
- // the address is right, and only the size has
- // to be adjusted
- (*w)->size = addr + size - (*w)->addr;
- }
- }
- ++w;
- }
+ // queue and keep track of whether we have merged or not
+ bool merged = isInWriteQueue.find(burstAlign(addr)) !=
+ isInWriteQueue.end();
// if the item was not merged we need to create a new write
// and enqueue it
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));
+
+ // 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;
} else {
+ DPRINTF(DRAM, "Merging write burst with existing queue entry\n");
+
// 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
DRAMCtrl::recvTimingReq(PacketPtr pkt)
{
- /// @todo temporary hack to deal with memory corruption issues until
- /// 4-phase transactions are complete
- for (int x = 0; x < pendingDelete.size(); x++)
- delete pendingDelete[x];
- pendingDelete.clear();
-
// This is where we enter from the outside world
DPRINTF(DRAM, "recvTimingReq: request %s addr %lld size %d\n",
pkt->cmdString(), pkt->getAddr(), pkt->getSize());
- // simply drop inhibited packets for now
- if (pkt->memInhibitAsserted()) {
- DPRINTF(DRAM, "Inhibited packet -- Dropping it now\n");
- pendingDelete.push_back(pkt);
- return true;
- }
+ panic_if(pkt->cacheResponding(), "Should not see packets where cache "
+ "is responding");
+
+ panic_if(!(pkt->isRead() || pkt->isWrite()),
+ "Should only see read and writes at memory controller\n");
// 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()) {
- assert(size != 0);
- if (readQueueFull(dram_pkt_count)) {
- DPRINTF(DRAM, "Read queue full, not accepting\n");
- // remember that we have to retry this port
- retryRdReq = true;
- numRdRetry++;
- return false;
- } else {
- addToReadQueue(pkt, dram_pkt_count);
- readReqs++;
- bytesReadSys += size;
- }
- } else if (pkt->isWrite()) {
+ if (pkt->isWrite()) {
assert(size != 0);
if (writeQueueFull(dram_pkt_count)) {
DPRINTF(DRAM, "Write queue full, not accepting\n");
// remember that we have to retry this port
retryWrReq = true;
- numWrRetry++;
+ stats.numWrRetry++;
return false;
} else {
addToWriteQueue(pkt, dram_pkt_count);
- writeReqs++;
- bytesWrittenSys += size;
+ stats.writeReqs++;
+ stats.bytesWrittenSys += size;
}
} else {
- DPRINTF(DRAM,"Neither read nor write, ignore timing\n");
- neitherReadNorWrite++;
- accessAndRespond(pkt, 1);
+ assert(pkt->isRead());
+ assert(size != 0);
+ if (readQueueFull(dram_pkt_count)) {
+ DPRINTF(DRAM, "Read queue full, not accepting\n");
+ // remember that we have to retry this port
+ retryRdReq = true;
+ stats.numRdRetry++;
+ return false;
+ } else {
+ addToReadQueue(pkt, dram_pkt_count);
+ stats.readReqs++;
+ stats.bytesReadSys += size;
+ }
}
return true;
DRAMPacket* dram_pkt = respQueue.front();
+ // if a read has reached its ready-time, decrement the number of reads
+ // At this point the packet has been handled and there is a possibility
+ // to switch to low-power mode if no other packet is available
+ --dram_pkt->rankRef.readEntries;
+ DPRINTF(DRAM, "number of read entries for rank %d is %d\n",
+ dram_pkt->rank, dram_pkt->rankRef.readEntries);
+
+ // counter should at least indicate one outstanding request
+ // for this read
+ assert(dram_pkt->rankRef.outstandingEvents > 0);
+ // read response received, decrement count
+ --dram_pkt->rankRef.outstandingEvents;
+
+ // at this moment should not have transitioned to a low-power state
+ assert((dram_pkt->rankRef.pwrState != PWR_SREF) &&
+ (dram_pkt->rankRef.pwrState != PWR_PRE_PDN) &&
+ (dram_pkt->rankRef.pwrState != PWR_ACT_PDN));
+
+ // 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 && enableDRAMPowerdown) {
+ // verify that there are no events scheduled
+ assert(!dram_pkt->rankRef.activateEvent.scheduled());
+ assert(!dram_pkt->rankRef.prechargeEvent.scheduled());
+
+ // if coming from active state, schedule power event to
+ // active power-down else go to precharge power-down
+ DPRINTF(DRAMState, "Rank %d sleep at tick %d; current power state is "
+ "%d\n", dram_pkt->rank, curTick(), dram_pkt->rankRef.pwrState);
+
+ // default to ACT power-down unless already in IDLE state
+ // could be in IDLE if PRE issued before data returned
+ PowerState next_pwr_state = PWR_ACT_PDN;
+ if (dram_pkt->rankRef.pwrState == PWR_IDLE) {
+ next_pwr_state = PWR_PRE_PDN;
+ }
+
+ dram_pkt->rankRef.powerDownSleep(next_pwr_state, curTick());
+ }
+
if (dram_pkt->burstHelper) {
// it is a split packet
dram_pkt->burstHelper->burstsServiced++;
schedule(respondEvent, respQueue.front()->readyTime);
} else {
// if there is nothing left in any queue, signal a drain
- if (writeQueue.empty() && readQueue.empty() &&
- drainManager) {
+ if (drainState() == DrainState::Draining &&
+ !totalWriteQueueSize && !totalReadQueueSize && allRanksDrained()) {
+
DPRINTF(Drain, "DRAM controller done draining\n");
- drainManager->signalDrainDone();
- drainManager = NULL;
+ signalDrainDone();
}
}
// so if there is a read that was forced to wait, retry now
if (retryRdReq) {
retryRdReq = false;
- port.sendRetry();
+ port.sendRetryReq();
}
}
-void
-DRAMCtrl::chooseNext(std::deque<DRAMPacket*>& queue, bool switched_cmd_type)
+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());
-
- if (queue.size() == 1) {
- DPRINTF(DRAM, "Single request, nothing to do\n");
- return;
- }
+ // This method does the arbitration between requests.
- if (memSchedPolicy == Enums::fcfs) {
- // Do nothing, since the correct request is already head
- } else if (memSchedPolicy == Enums::frfcfs) {
- reorderQueue(queue, switched_cmd_type);
- } else
- panic("No scheduling policy chosen\n");
+ 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()) {
+ 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");
+ }
+ }
+ return ret;
}
-void
-DRAMCtrl::reorderQueue(std::deque<DRAMPacket*>& queue, bool switched_cmd_type)
+DRAMCtrl::DRAMPacketQueue::iterator
+DRAMCtrl::chooseNextFRFCFS(DRAMPacketQueue& queue, Tick extra_col_delay)
{
- // Only determine this when needed
- uint64_t earliest_banks = 0;
-
- // Search for row hits first, if no row hit is found then schedule the
- // packet to one of the earliest banks available
+ // Only determine this if needed
+ vector<uint32_t> earliest_banks(ranksPerChannel, 0);
+
+ // Has minBankPrep been called to populate earliest_banks?
+ bool filled_earliest_banks = false;
+ // can the PRE/ACT sequence be done without impacting utlization?
+ bool hidden_bank_prep = false;
+
+ // search for seamless row hits first, if no seamless row hit is
+ // found then determine if there are other packets that can be issued
+ // without incurring additional bus delay due to bank timing
+ // Will select closed rows first to enable more open row possibilies
+ // in future selections
+ bool found_hidden_bank = false;
+
+ // remember if we found a row hit, not seamless, but bank prepped
+ // and ready
+ bool found_prepped_pkt = false;
+
+ // if we have no row hit, prepped or not, and no seamless packet,
+ // just go for the earliest possible
bool found_earliest_pkt = false;
- bool found_prepped_diff_rank_pkt = false;
- auto selected_pkt_it = queue.begin();
+
+ auto selected_pkt_it = queue.end();
+
+ // time we need to issue a column command to be seamless
+ const Tick min_col_at = std::max(nextBurstAt + extra_col_delay, curTick());
for (auto i = queue.begin(); i != queue.end() ; ++i) {
DRAMPacket* dram_pkt = *i;
const Bank& bank = dram_pkt->bankRef;
- // Check if it is a row hit
- if (bank.openRow == dram_pkt->row) {
- if (dram_pkt->rank == activeRank || switched_cmd_type) {
- // FCFS within the hits, giving priority to commands
- // that access the same rank as the previous burst
- // to minimize bus turnaround delays
- // Only give rank prioity when command type is not changing
- DPRINTF(DRAM, "Row buffer hit\n");
- selected_pkt_it = i;
- break;
- } else if (!found_prepped_diff_rank_pkt) {
- // found row hit for command on different rank than prev burst
- selected_pkt_it = i;
- found_prepped_diff_rank_pkt = true;
- }
- } else if (!found_earliest_pkt & !found_prepped_diff_rank_pkt) {
- // No row hit and
- // haven't found an entry with a row hit to a new rank
- if (earliest_banks == 0)
- // Determine entries with earliest bank prep delay
- // Function will give priority to commands that access the
- // same rank as previous burst and can prep the bank seamlessly
- earliest_banks = minBankPrep(queue, switched_cmd_type);
-
- // FCFS - Bank is first available bank
- if (bits(earliest_banks, dram_pkt->bankId, dram_pkt->bankId)) {
- // Remember the packet to be scheduled to one of the earliest
- // banks available, FCFS amongst the earliest banks
- selected_pkt_it = i;
- found_earliest_pkt = true;
+ 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
+ // delays, or we switched read/write and might as well
+ // go for the row hit
+ if (col_allowed_at <= min_col_at) {
+ // FCFS within the hits, giving priority to
+ // commands that can issue seamlessly, without
+ // additional delay, such as same rank accesses
+ // and/or different bank-group accesses
+ DPRINTF(DRAM, "%s Seamless row buffer hit\n", __func__);
+ selected_pkt_it = i;
+ // no need to look through the remaining queue entries
+ break;
+ } else if (!found_hidden_bank && !found_prepped_pkt) {
+ // if we did not find a packet to a closed row that can
+ // issue the bank commands without incurring delay, and
+ // did not yet find a packet to a prepped row, remember
+ // the current one
+ selected_pkt_it = i;
+ found_prepped_pkt = true;
+ DPRINTF(DRAM, "%s Prepped row buffer hit\n", __func__);
+ }
+ } else if (!found_earliest_pkt) {
+ // if we have not initialised the bank status, do it
+ // now, and only once per scheduling decisions
+ if (!filled_earliest_banks) {
+ // determine entries with earliest bank delay
+ std::tie(earliest_banks, hidden_bank_prep) =
+ minBankPrep(queue, min_col_at);
+ filled_earliest_banks = true;
+ }
+
+ // bank is amongst first available banks
+ // minBankPrep will give priority to packets that can
+ // issue seamlessly
+ if (bits(earliest_banks[dram_pkt->rank],
+ dram_pkt->bank, dram_pkt->bank)) {
+ found_earliest_pkt = true;
+ found_hidden_bank = hidden_bank_prep;
+
+ // give priority to packets that can issue
+ // bank commands 'behind the scenes'
+ // any additional delay if any will be due to
+ // col-to-col command requirements
+ if (hidden_bank_prep || !found_prepped_pkt)
+ selected_pkt_it = i;
+ }
}
+ } else {
+ DPRINTF(DRAM, "%s bank %d - Rank %d not available\n", __func__,
+ dram_pkt->bankRef.bank, dram_pkt->rankRef.rank);
}
}
- DRAMPacket* selected_pkt = *selected_pkt_it;
- queue.erase(selected_pkt_it);
- queue.push_front(selected_pkt);
+ if (selected_pkt_it == queue.end()) {
+ DPRINTF(DRAM, "%s no available ranks found\n", __func__);
+ }
+
+ return selected_pkt_it;
}
void
if (needsResponse) {
// access already turned the packet into a response
assert(pkt->isResponse());
-
- // @todo someone should pay for this
- pkt->firstWordDelay = pkt->lastWordDelay = 0;
+ // response_time consumes the static latency and is charged also
+ // with headerDelay that takes into account the delay provided by
+ // the xbar and also the payloadDelay that takes into account the
+ // number of data beats.
+ Tick response_time = curTick() + static_latency + pkt->headerDelay +
+ pkt->payloadDelay;
+ // Here we reset the timing of the packet before sending it out.
+ pkt->headerDelay = pkt->payloadDelay = 0;
// queue the packet in the response queue to be sent out after
// the static latency has passed
- port.schedTimingResp(pkt, curTick() + static_latency);
+ port.schedTimingResp(pkt, response_time);
} else {
// @todo the packet is going to be deleted, and the DRAMPacket
// is still having a pointer to it
- pendingDelete.push_back(pkt);
+ pendingDelete.reset(pkt);
}
DPRINTF(DRAM, "Done\n");
}
void
-DRAMCtrl::activateBank(Bank& bank, Tick act_tick, uint32_t row)
+DRAMCtrl::activateBank(Rank& rank_ref, Bank& bank_ref,
+ Tick act_tick, uint32_t row)
{
- // get the rank index from the bank
- uint8_t rank = bank.rank;
-
- assert(actTicks[rank].size() == activationLimit);
+ assert(rank_ref.actTicks.size() == activationLimit);
DPRINTF(DRAM, "Activate at tick %d\n", act_tick);
// update the open row
- assert(bank.openRow == Bank::NO_ROW);
- bank.openRow = row;
+ assert(bank_ref.openRow == Bank::NO_ROW);
+ bank_ref.openRow = row;
// start counting anew, this covers both the case when we
// auto-precharged, and when this access is forced to
// precharge
- bank.bytesAccessed = 0;
- bank.rowAccesses = 0;
+ bank_ref.bytesAccessed = 0;
+ bank_ref.rowAccesses = 0;
- ++numBanksActive;
- assert(numBanksActive <= banksPerRank * ranksPerChannel);
+ ++rank_ref.numBanksActive;
+ assert(rank_ref.numBanksActive <= banksPerRank);
DPRINTF(DRAM, "Activate bank %d, rank %d at tick %lld, now got %d active\n",
- bank.bank, bank.rank, act_tick, numBanksActive);
+ bank_ref.bank, rank_ref.rank, act_tick,
+ ranks[rank_ref.rank]->numBanksActive);
- rankPower[bank.rank].powerlib.doCommand(MemCommand::ACT, bank.bank,
- divCeil(act_tick, tCK) -
- timeStampOffset);
+ rank_ref.cmdList.push_back(Command(MemCommand::ACT, bank_ref.bank,
+ act_tick));
DPRINTF(DRAMPower, "%llu,ACT,%d,%d\n", divCeil(act_tick, tCK) -
- timeStampOffset, bank.bank, bank.rank);
+ timeStampOffset, bank_ref.bank, rank_ref.rank);
// The next access has to respect tRAS for this bank
- bank.preAllowedAt = act_tick + tRAS;
+ bank_ref.preAllowedAt = act_tick + tRAS;
- // Respect the row-to-column command delay
- bank.colAllowedAt = std::max(act_tick + tRCD, bank.colAllowedAt);
+ // Respect the row-to-column command delay for both read and write cmds
+ bank_ref.rdAllowedAt = std::max(act_tick + tRCD, bank_ref.rdAllowedAt);
+ bank_ref.wrAllowedAt = std::max(act_tick + tRCD, bank_ref.wrAllowedAt);
// start by enforcing tRRD
- for(int i = 0; i < banksPerRank; i++) {
+ for (int i = 0; i < banksPerRank; i++) {
// next activate to any bank in this rank must not happen
// before tRRD
- if (bankGroupArch && (bank.bankgr == banks[rank][i].bankgr)) {
+ if (bankGroupArch && (bank_ref.bankgr == rank_ref.banks[i].bankgr)) {
// bank group architecture requires longer delays between
// ACT commands within the same bank group. Use tRRD_L
// in this case
- banks[rank][i].actAllowedAt = std::max(act_tick + tRRD_L,
- banks[rank][i].actAllowedAt);
+ rank_ref.banks[i].actAllowedAt = std::max(act_tick + tRRD_L,
+ rank_ref.banks[i].actAllowedAt);
} else {
// use shorter tRRD value when either
// 1) bank group architecture is not supportted
// 2) bank is in a different bank group
- banks[rank][i].actAllowedAt = std::max(act_tick + tRRD,
- banks[rank][i].actAllowedAt);
+ rank_ref.banks[i].actAllowedAt = std::max(act_tick + tRRD,
+ rank_ref.banks[i].actAllowedAt);
}
}
// next, we deal with tXAW, if the activation limit is disabled
// then we directly schedule an activate power event
- if (!actTicks[rank].empty()) {
+ if (!rank_ref.actTicks.empty()) {
// sanity check
- if (actTicks[rank].back() &&
- (act_tick - actTicks[rank].back()) < tXAW) {
+ if (rank_ref.actTicks.back() &&
+ (act_tick - rank_ref.actTicks.back()) < tXAW) {
panic("Got %d activates in window %d (%llu - %llu) which "
"is smaller than %llu\n", activationLimit, act_tick -
- actTicks[rank].back(), act_tick, actTicks[rank].back(),
- tXAW);
+ rank_ref.actTicks.back(), act_tick,
+ rank_ref.actTicks.back(), tXAW);
}
// shift the times used for the book keeping, the last element
// (highest index) is the oldest one and hence the lowest value
- actTicks[rank].pop_back();
+ rank_ref.actTicks.pop_back();
// record an new activation (in the future)
- actTicks[rank].push_front(act_tick);
+ rank_ref.actTicks.push_front(act_tick);
// cannot activate more than X times in time window tXAW, push the
// next one (the X + 1'st activate) to be tXAW away from the
// oldest in our window of X
- if (actTicks[rank].back() &&
- (act_tick - actTicks[rank].back()) < tXAW) {
+ if (rank_ref.actTicks.back() &&
+ (act_tick - rank_ref.actTicks.back()) < tXAW) {
DPRINTF(DRAM, "Enforcing tXAW with X = %d, next activate "
"no earlier than %llu\n", activationLimit,
- actTicks[rank].back() + tXAW);
- for(int j = 0; j < banksPerRank; j++)
+ rank_ref.actTicks.back() + tXAW);
+ for (int j = 0; j < banksPerRank; j++)
// next activate must not happen before end of window
- banks[rank][j].actAllowedAt =
- std::max(actTicks[rank].back() + tXAW,
- banks[rank][j].actAllowedAt);
+ rank_ref.banks[j].actAllowedAt =
+ std::max(rank_ref.actTicks.back() + tXAW,
+ rank_ref.banks[j].actAllowedAt);
}
}
// at the point when this activate takes place, make sure we
// transition to the active power state
- if (!activateEvent.scheduled())
- schedule(activateEvent, act_tick);
- else if (activateEvent.when() > act_tick)
+ if (!rank_ref.activateEvent.scheduled())
+ schedule(rank_ref.activateEvent, act_tick);
+ else if (rank_ref.activateEvent.when() > act_tick)
// move it sooner in time
- reschedule(activateEvent, act_tick);
-}
-
-void
-DRAMCtrl::processActivateEvent()
-{
- // we should transition to the active state as soon as any bank is active
- if (pwrState != PWR_ACT)
- // note that at this point numBanksActive could be back at
- // zero again due to a precharge scheduled in the future
- schedulePowerEvent(PWR_ACT, curTick());
+ reschedule(rank_ref.activateEvent, act_tick);
}
void
-DRAMCtrl::prechargeBank(Bank& bank, Tick pre_at, bool trace)
+DRAMCtrl::prechargeBank(Rank& rank_ref, Bank& bank, Tick pre_at, bool trace)
{
// make sure the bank has an open row
assert(bank.openRow != Bank::NO_ROW);
// 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;
bank.actAllowedAt = std::max(bank.actAllowedAt, pre_done_at);
- assert(numBanksActive != 0);
- --numBanksActive;
+ assert(rank_ref.numBanksActive != 0);
+ --rank_ref.numBanksActive;
DPRINTF(DRAM, "Precharging bank %d, rank %d at tick %lld, now got "
- "%d active\n", bank.bank, bank.rank, pre_at, numBanksActive);
+ "%d active\n", bank.bank, rank_ref.rank, pre_at,
+ rank_ref.numBanksActive);
if (trace) {
- rankPower[bank.rank].powerlib.doCommand(MemCommand::PRE, bank.bank,
- divCeil(pre_at, tCK) -
- timeStampOffset);
+ rank_ref.cmdList.push_back(Command(MemCommand::PRE, bank.bank,
+ pre_at));
DPRINTF(DRAMPower, "%llu,PRE,%d,%d\n", divCeil(pre_at, tCK) -
- timeStampOffset, bank.bank, bank.rank);
+ timeStampOffset, bank.bank, rank_ref.rank);
}
// if we look at the current number of active banks we might be
// tempted to think the DRAM is now idle, however this can be
// would have reached the idle state, so schedule an event and
// rather check once we actually make it to the point in time when
// the (last) precharge takes place
- if (!prechargeEvent.scheduled())
- schedule(prechargeEvent, pre_done_at);
- else if (prechargeEvent.when() < pre_done_at)
- reschedule(prechargeEvent, pre_done_at);
-}
-
-void
-DRAMCtrl::processPrechargeEvent()
-{
- // if we reached zero, then special conditions apply as we track
- // if all banks are precharged for the power models
- if (numBanksActive == 0) {
- // we should transition to the idle state when the last bank
- // is precharged
- schedulePowerEvent(PWR_IDLE, curTick());
+ if (!rank_ref.prechargeEvent.scheduled()) {
+ schedule(rank_ref.prechargeEvent, pre_done_at);
+ // New event, increment count
+ ++rank_ref.outstandingEvents;
+ } else if (rank_ref.prechargeEvent.when() < pre_done_at) {
+ reschedule(rank_ref.prechargeEvent, pre_done_at);
}
}
DPRINTF(DRAM, "Timing access to addr %lld, rank/bank/row %d %d %d\n",
dram_pkt->addr, dram_pkt->rank, dram_pkt->bank, dram_pkt->row);
+ // get the rank
+ Rank& rank = dram_pkt->rankRef;
+
+ // are we in or transitioning to a low-power state and have not scheduled
+ // a power-up event?
+ // if so, wake up from power down to issue RD/WR burst
+ if (rank.inLowPowerState) {
+ assert(rank.pwrState != PWR_SREF);
+ rank.scheduleWakeUpEvent(tXP);
+ }
+
// get the bank
Bank& bank = dram_pkt->bankRef;
// for the state we need to track if it is a row hit or not
bool row_hit = true;
- // respect any constraints on the command (e.g. tRCD or tCCD)
- Tick cmd_at = std::max(bank.colAllowedAt, curTick());
-
// Determine the access latency and update the bank state
if (bank.openRow == dram_pkt->row) {
// nothing to do
// If there is a page open, precharge it.
if (bank.openRow != Bank::NO_ROW) {
- prechargeBank(bank, std::max(bank.preAllowedAt, curTick()));
+ prechargeBank(rank, bank, std::max(bank.preAllowedAt, curTick()));
}
// next we need to account for the delay in activating the
// Record the activation and deal with all the global timing
// constraints caused be a new activation (tRRD and tXAW)
- activateBank(bank, act_tick, dram_pkt->row);
-
- // issue the command as early as possible
- cmd_at = bank.colAllowedAt;
+ activateBank(rank, bank, act_tick, dram_pkt->row);
}
+ // respect any constraints on the command (e.g. tRCD or tCCD)
+ 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
- // the command
- cmd_at = std::max(cmd_at, busBusyUntil - tCL);
+ // the command; need minimum of tBURST between commands
+ Tick cmd_at = std::max({col_allowed_at, nextBurstAt, curTick()});
// update the packet ready time
dram_pkt->readyTime = cmd_at + tCL + tBURST;
- // only one burst can use the bus at any one point in time
- assert(dram_pkt->readyTime - busBusyUntil >= tBURST);
-
// update the time for the next read/write burst for each
- // bank (add a max with tCCD/tCCD_L here)
- Tick cmd_dly;
- for(int j = 0; j < ranksPerChannel; j++) {
- for(int i = 0; i < banksPerRank; i++) {
+ // bank (add a max with tCCD/tCCD_L/tCCD_L_WR here)
+ Tick dly_to_rd_cmd;
+ Tick dly_to_wr_cmd;
+ for (int j = 0; j < ranksPerChannel; j++) {
+ for (int i = 0; i < banksPerRank; i++) {
// next burst to same bank group in this rank must not happen
// before tCCD_L. Different bank group timing requirement is
// tBURST; Add tCS for different ranks
if (dram_pkt->rank == j) {
- if (bankGroupArch && (bank.bankgr == banks[j][i].bankgr)) {
+ if (bankGroupArch &&
+ (bank.bankgr == ranks[j]->banks[i].bankgr)) {
// bank group architecture requires longer delays between
// RD/WR burst commands to the same bank group.
- // Use tCCD_L in this case
- cmd_dly = tCCD_L;
+ // 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() ?
+ tCCD_L : std::max(tCCD_L, wrToRdDly);
+ dly_to_wr_cmd = dram_pkt->isRead() ?
+ std::max(tCCD_L, rdToWrDly) : tCCD_L_WR;
} else {
- // use tBURST (equivalent to tCCD_S), the shorter
- // cas-to-cas delay value, when either:
- // 1) bank group architecture is not supportted
- // 2) bank is in a different bank group
- cmd_dly = tBURST;
+ // 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;
}
} else {
- // different rank is by default in a different bank group
- // use tBURST (equivalent to tCCD_S), which is the shorter
- // cas-to-cas delay in this case
- // Add tCS to account for rank-to-rank bus delay requirements
- cmd_dly = tBURST + tCS;
+ // different rank is by default in a different bank group and
+ // doesn't require longer tCCD or additional RTW, WTR delays
+ // Need to account for rank-to-rank switching with tCS
+ dly_to_wr_cmd = rankToRankDly;
+ dly_to_rd_cmd = rankToRankDly;
}
- banks[j][i].colAllowedAt = std::max(cmd_at + cmd_dly,
- banks[j][i].colAllowedAt);
+ ranks[j]->banks[i].rdAllowedAt = std::max(cmd_at + dly_to_rd_cmd,
+ ranks[j]->banks[i].rdAllowedAt);
+ ranks[j]->banks[i].wrAllowedAt = std::max(cmd_at + dly_to_wr_cmd,
+ ranks[j]->banks[i].wrAllowedAt);
}
}
// 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 have found required condition or
- // reached the end
- while (!(got_more_hits &&
- (got_bank_conflict || pageMgmt == Enums::close_adaptive)) &&
- 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 :
MemCommand::WR;
+ // Update bus state to reflect when previous command was issued
+ nextBurstAt = cmd_at + tBURST;
+
+ DPRINTF(DRAM, "Access to %lld, ready at %lld next burst at %lld.\n",
+ dram_pkt->addr, dram_pkt->readyTime, nextBurstAt);
+
+ dram_pkt->rankRef.cmdList.push_back(Command(command, dram_pkt->bank,
+ cmd_at));
+
+ DPRINTF(DRAMPower, "%llu,%s,%d,%d\n", divCeil(cmd_at, tCK) -
+ timeStampOffset, mem_cmd, dram_pkt->bank, dram_pkt->rank);
+
// if this access should use auto-precharge, then we are
- // closing the row
+ // closing the row after the read/write burst
if (auto_precharge) {
// if auto-precharge push a PRE command at the correct tick to the
// list used by DRAMPower library to calculate power
- prechargeBank(bank, std::max(curTick(), bank.preAllowedAt));
+ prechargeBank(rank, bank, std::max(curTick(), bank.preAllowedAt));
DPRINTF(DRAM, "Auto-precharged bank: %d\n", dram_pkt->bankId);
}
- // Update bus state
- busBusyUntil = dram_pkt->readyTime;
-
- DPRINTF(DRAM, "Access to %lld, ready at %lld bus busy until %lld.\n",
- dram_pkt->addr, dram_pkt->readyTime, busBusyUntil);
-
- rankPower[dram_pkt->rank].powerlib.doCommand(command, dram_pkt->bank,
- divCeil(cmd_at, tCK) -
- timeStampOffset);
-
- DPRINTF(DRAMPower, "%llu,%s,%d,%d\n", divCeil(cmd_at, tCK) -
- timeStampOffset, mem_cmd, dram_pkt->bank, dram_pkt->rank);
-
// Update the minimum timing between the requests, this is a
// conservative estimate of when we have to schedule the next
// request to not introduce any unecessary bubbles. In most cases
// we will wake up sooner than we have to.
- nextReqTime = busBusyUntil - (tRP + tRCD + tCL);
+ 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()
{
- // pre-emptively set to false. Overwrite if in READ_TO_WRITE
- // or WRITE_TO_READ state
- bool switched_cmd_type = false;
- if (busState == READ_TO_WRITE) {
- 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
- busState = WRITE;
- switched_cmd_type = true;
- } else if (busState == WRITE_TO_READ) {
- DPRINTF(DRAM, "Switching to reads after %d writes with %d writes "
- "waiting\n", writesThisTime, writeQueue.size());
-
- wrPerTurnAround.sample(writesThisTime);
- writesThisTime = 0;
-
- busState = READ;
- switched_cmd_type = true;
- }
-
- if (refreshState != REF_IDLE) {
- // if a refresh waiting for this event loop to finish, then hand
- // over now, and do not schedule a new nextReqEvent
- if (refreshState == REF_DRAIN) {
- DPRINTF(DRAM, "Refresh drain done, now precharging\n");
+ // transition is handled by QoS algorithm if enabled
+ if (turnPolicy) {
+ // select bus state - only done if QoS algorithms are in use
+ busStateNext = selectNextBusState();
+ }
- refreshState = REF_PRE;
+ // 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;
+ }
+ }
- // hand control back to the refresh event loop
- schedule(refreshEvent, curTick());
+ // 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()) {
+ if (r->pwrState != PWR_SREF) {
+ // rank is busy refreshing
+ DPRINTF(DRAMState, "Rank %d is not available\n", r->rank);
+ busyRanks++;
+
+ // let the rank know that if it was waiting to drain, it
+ // is now done and ready to proceed
+ r->checkDrainDone();
+ }
+
+ // check if we were in self-refresh and haven't started
+ // to transition out
+ if ((r->pwrState == PWR_SREF) && r->inLowPowerState) {
+ DPRINTF(DRAMState, "Rank %d is in self-refresh\n", r->rank);
+ // if we have commands queued to this rank and we don't have
+ // a minimum number of active commands enqueued,
+ // exit self-refresh
+ if (r->forceSelfRefreshExit()) {
+ DPRINTF(DRAMState, "rank %d was in self refresh and"
+ " should wake up\n", r->rank);
+ //wake up from self-refresh
+ r->scheduleWakeUpEvent(tXS);
+ // things are brought back into action once a refresh is
+ // performed after self-refresh
+ // continue with selection for other ranks
+ }
+ }
}
+ }
- // let the refresh finish before issuing any further requests
+ if (busyRanks == ranksPerChannel) {
+ // if all ranks are refreshing wait for them to finish
+ // and stall this state machine without taking any further
+ // action, and do not schedule a new nextReqEvent
return;
}
// 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() &&
- (drainManager || writeQueue.size() > writeLowThreshold)) {
+ if (!(totalWriteQueueSize == 0) &&
+ (drainState() == DrainState::Draining ||
+ totalWriteQueueSize > writeLowThreshold)) {
+ DPRINTF(DRAM, "Switching to writes due to read queue empty\n");
switch_to_writes = true;
} else {
// check if we are drained
- if (respQueue.empty () && drainManager) {
+ // not done draining until in PWR_IDLE state
+ // ensuring all banks are closed and
+ // have exited low power states
+ if (drainState() == DrainState::Draining &&
+ respQueue.empty() && allRanksDrained()) {
+
DPRINTF(Drain, "DRAM controller done draining\n");
- drainManager->signalDrainDone();
- drainManager = NULL;
+ signalDrainDone();
}
// nothing to do, not even any point in scheduling an
return;
}
} else {
- // Figure out which read request goes next, and move it to the
- // front of the read queue
- chooseNext(readQueue, switched_cmd_type);
-
- DRAMPacket* dram_pkt = readQueue.front();
-
- // here we get a bit creative and shift the bus busy time not
- // just the tWTR, but also a CAS latency to capture the fact
- // that we are allowed to prepare a new bank, but not issue a
- // read command until after tWTR, in essence we capture a
- // bubble on the data bus that is tWTR + tCL
- if (switched_cmd_type && dram_pkt->rank == activeRank) {
- busBusyUntil += tWTR + tCL;
+
+ 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;
+ }
}
- doDRAMAccess(dram_pkt);
+ // 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 (!read_found) {
+ DPRINTF(DRAM, "No Reads Found - exiting\n");
+ return;
+ }
- // At this point we're done dealing with the request
- readQueue.pop_front();
+ auto dram_pkt = *to_read;
+ assert(dram_pkt->rankRef.inRefIdleState());
+
+ doDRAMAccess(dram_pkt);
+
+ // 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
// draining), or because the writes hit the hight threshold
if (switch_to_writes) {
// transition to writing
- busState = READ_TO_WRITE;
+ busStateNext = WRITE;
}
} else {
- chooseNext(writeQueue, switched_cmd_type);
- DRAMPacket* dram_pkt = writeQueue.front();
- // sanity check
- assert(dram_pkt->size <= burstSize);
- // add a bubble to the data bus, as defined by the
- // tRTW when access is to the same rank as previous burst
- // Different rank timing is handled with tCS, which is
- // applied to colAllowedAt
- if (switched_cmd_type && dram_pkt->rank == activeRank) {
- busBusyUntil += tRTW;
- }
+ bool write_found = false;
+ DRAMPacketQueue::iterator to_write;
+ uint8_t prio = numPriorities();
- doDRAMAccess(dram_pkt);
+ for (auto queue = writeQueue.rbegin();
+ queue != writeQueue.rend(); ++queue) {
- writeQueue.pop_front();
- delete dram_pkt;
+ prio--;
- // 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 &&
- !drainManager) ||
- (!readQueue.empty() && writesThisTime >= minWritesPerSwitch)) {
- // turn the bus back around for reads again
- busState = WRITE_TO_READ;
+ DPRINTF(QOS,
+ "DRAM controller checking WRITE queue [%d] priority [%d elements]\n",
+ prio, queue->size());
- // note that the we switch back to reads also in the idle
- // case, which eventually will check for any draining and
- // also pause any further scheduling if there is really
- // nothing to do
+ // 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;
+ }
}
- }
- schedule(nextReqEvent, std::max(nextReqTime, curTick()));
+ // 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 (!write_found) {
+ DPRINTF(DRAM, "No Writes Found - exiting\n");
+ return;
+ }
+
+ auto dram_pkt = *to_write;
+
+ assert(dram_pkt->rankRef.inRefIdleState());
+ // sanity check
+ assert(dram_pkt->size <= burstSize);
+
+ doDRAMAccess(dram_pkt);
+
+ // removed write from queue, decrement count
+ --dram_pkt->rankRef.writeEntries;
+
+ // Schedule write done event to decrement event count
+ // after the readyTime has been reached
+ // Only schedule latest write event to minimize events
+ // required; only need to ensure that final event scheduled covers
+ // the time that writes are outstanding and bus is active
+ // to holdoff power-down entry events
+ if (!dram_pkt->rankRef.writeDoneEvent.scheduled()) {
+ schedule(dram_pkt->rankRef.writeDoneEvent, dram_pkt->readyTime);
+ // New event, increment count
+ ++dram_pkt->rankRef.outstandingEvents;
+
+ } else if (dram_pkt->rankRef.writeDoneEvent.when() <
+ 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.
+ 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;
+
+ // note that the we switch back to reads also in the idle
+ // case, which eventually will check for any draining and
+ // also pause any further scheduling if there is really
+ // nothing to do
+ }
+ }
+ // It is possible that a refresh to another rank kicks things back into
+ // action before reaching this point.
+ if (!nextReqEvent.scheduled())
+ schedule(nextReqEvent, std::max(nextReqTime, curTick()));
// If there is space available and we have writes waiting then let
// 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.sendRetry();
+ port.sendRetryReq();
}
}
-uint64_t
-DRAMCtrl::minBankPrep(const deque<DRAMPacket*>& queue,
- bool switched_cmd_type) const
+pair<vector<uint32_t>, bool>
+DRAMCtrl::minBankPrep(const DRAMPacketQueue& queue,
+ Tick min_col_at) const
{
- uint64_t bank_mask = 0;
Tick min_act_at = MaxTick;
+ vector<uint32_t> bank_mask(ranksPerChannel, 0);
- uint64_t bank_mask_same_rank = 0;
- Tick min_act_at_same_rank = MaxTick;
+ // latest Tick for which ACT can occur without incurring additoinal
+ // delay on the data bus
+ const Tick hidden_act_max = std::max(min_col_at - tRCD, curTick());
- // Give precedence to commands that access same rank as previous command
- bool same_rank_match = false;
+ // Flag condition when burst can issue back-to-back with previous burst
+ bool found_seamless_bank = false;
+
+ // Flag condition when bank can be opened without incurring additional
+ // delay on the data bus
+ bool hidden_bank_prep = false;
// determine if we have queued transactions targetting the
// bank in question
vector<bool> got_waiting(ranksPerChannel * banksPerRank, false);
- for (auto p = queue.begin(); p != queue.end(); ++p) {
- got_waiting[(*p)->bankId] = true;
+ for (const auto& p : queue) {
+ if (p->rankRef.inRefIdleState())
+ got_waiting[p->bankId] = true;
}
+ // Find command with optimal bank timing
+ // Will prioritize commands that can issue seamlessly.
for (int i = 0; i < ranksPerChannel; i++) {
for (int j = 0; j < banksPerRank; j++) {
- uint8_t bank_id = i * banksPerRank + j;
+ uint16_t bank_id = i * banksPerRank + j;
// if we have waiting requests for the bank, and it is
// amongst the first available, update the mask
if (got_waiting[bank_id]) {
+ // make sure this rank is not currently refreshing.
+ assert(ranks[i]->inRefIdleState());
// simplistic approximation of when the bank can issue
// an activate, ignoring any rank-to-rank switching
// cost in this calculation
- Tick act_at = banks[i][j].openRow == Bank::NO_ROW ?
- banks[i][j].actAllowedAt :
- std::max(banks[i][j].preAllowedAt, curTick()) + tRP;
-
- // prioritize commands that access the
- // same rank as previous burst
- // Calculate bank mask separately for the case and
- // evaluate after loop iterations complete
- if (i == activeRank && ranksPerChannel > 1) {
- if (act_at <= min_act_at_same_rank) {
- // reset same rank bank mask if new minimum is found
- // and previous minimum could not immediately send ACT
- if (act_at < min_act_at_same_rank &&
- min_act_at_same_rank > curTick())
- bank_mask_same_rank = 0;
-
- // Set flag indicating that a same rank
- // opportunity was found
- same_rank_match = true;
-
- // set the bit corresponding to the available bank
- replaceBits(bank_mask_same_rank, bank_id, bank_id, 1);
- min_act_at_same_rank = act_at;
- }
- } else {
- if (act_at <= min_act_at) {
- // reset bank mask if new minimum is found
- // and either previous minimum could not immediately send ACT
- if (act_at < min_act_at && min_act_at > curTick())
- bank_mask = 0;
- // set the bit corresponding to the available bank
- replaceBits(bank_mask, bank_id, bank_id, 1);
- min_act_at = act_at;
+ Tick act_at = ranks[i]->banks[j].openRow == Bank::NO_ROW ?
+ std::max(ranks[i]->banks[j].actAllowedAt, curTick()) :
+ std::max(ranks[i]->banks[j].preAllowedAt, curTick()) + tRP;
+
+ // When is the earliest the R/W burst can issue?
+ const Tick col_allowed_at = (busState == READ) ?
+ ranks[i]->banks[j].rdAllowedAt :
+ ranks[i]->banks[j].wrAllowedAt;
+ Tick col_at = std::max(col_allowed_at, act_at + tRCD);
+
+ // bank can issue burst back-to-back (seamlessly) with
+ // previous burst
+ bool new_seamless_bank = col_at <= min_col_at;
+
+ // if we found a new seamless bank or we have no
+ // seamless banks, and got a bank with an earlier
+ // activate time, it should be added to the bit mask
+ if (new_seamless_bank ||
+ (!found_seamless_bank && act_at <= min_act_at)) {
+ // if we did not have a seamless bank before, and
+ // we do now, reset the bank mask, also reset it
+ // if we have not yet found a seamless bank and
+ // the activate time is smaller than what we have
+ // seen so far
+ if (!found_seamless_bank &&
+ (new_seamless_bank || act_at < min_act_at)) {
+ std::fill(bank_mask.begin(), bank_mask.end(), 0);
}
+
+ found_seamless_bank |= new_seamless_bank;
+
+ // ACT can occur 'behind the scenes'
+ hidden_bank_prep = act_at <= hidden_act_max;
+
+ // set the bit corresponding to the available bank
+ replaceBits(bank_mask[i], j, j, 1);
+ min_act_at = act_at;
}
}
}
}
- // Determine the earliest time when the next burst can issue based
- // on the current busBusyUntil delay.
- // Offset by tRCD to correlate with ACT timing variables
- Tick min_cmd_at = busBusyUntil - tCL - tRCD;
+ return make_pair(bank_mask, hidden_bank_prep);
+}
- // Prioritize same rank accesses that can issue B2B
- // Only optimize for same ranks when the command type
- // does not change; do not want to unnecessarily incur tWTR
- //
- // Resulting FCFS prioritization Order is:
- // 1) Commands that access the same rank as previous burst
- // and can prep the bank seamlessly.
- // 2) Commands (any rank) with earliest bank prep
- if (!switched_cmd_type && same_rank_match &&
- min_act_at_same_rank <= min_cmd_at) {
- bank_mask = bank_mask_same_rank;
+DRAMCtrl::Rank::Rank(DRAMCtrl& _memory, const DRAMCtrlParams* _p, int rank)
+ : EventManager(&_memory), memory(_memory),
+ pwrStateTrans(PWR_IDLE), pwrStatePostRefresh(PWR_IDLE),
+ pwrStateTick(0), refreshDueAt(0), pwrState(PWR_IDLE),
+ refreshState(REF_IDLE), inLowPowerState(false), rank(rank),
+ readEntries(0), writeEntries(0), outstandingEvents(0),
+ wakeUpAllowedAt(0), power(_p, false), banks(_p->banks_per_rank),
+ numBanksActive(0), actTicks(_p->activation_limit, 0),
+ writeDoneEvent([this]{ processWriteDoneEvent(); }, name()),
+ activateEvent([this]{ processActivateEvent(); }, name()),
+ prechargeEvent([this]{ processPrechargeEvent(); }, name()),
+ refreshEvent([this]{ processRefreshEvent(); }, name()),
+ powerEvent([this]{ processPowerEvent(); }, name()),
+ wakeUpEvent([this]{ processWakeUpEvent(); }, name()),
+ stats(_memory, *this)
+{
+ for (int b = 0; b < _p->banks_per_rank; b++) {
+ banks[b].bank = b;
+ // GDDR addressing of banks to BG is linear.
+ // Here we assume that all DRAM generations address bank groups as
+ // follows:
+ if (_p->bank_groups_per_rank > 0) {
+ // Simply assign lower bits to bank group in order to
+ // rotate across bank groups as banks are incremented
+ // e.g. with 4 banks per bank group and 16 banks total:
+ // banks 0,4,8,12 are in bank group 0
+ // banks 1,5,9,13 are in bank group 1
+ // banks 2,6,10,14 are in bank group 2
+ // banks 3,7,11,15 are in bank group 3
+ banks[b].bankgr = b % _p->bank_groups_per_rank;
+ } else {
+ // No bank groups; simply assign to bank number
+ banks[b].bankgr = b;
+ }
}
+}
+
+void
+DRAMCtrl::Rank::startup(Tick ref_tick)
+{
+ assert(ref_tick > curTick());
+
+ pwrStateTick = curTick();
+
+ // kick off the refresh, and give ourselves enough time to
+ // precharge
+ schedule(refreshEvent, ref_tick);
+}
- return bank_mask;
+void
+DRAMCtrl::Rank::suspend()
+{
+ deschedule(refreshEvent);
+
+ // Update the stats
+ updatePowerStats();
+
+ // don't automatically transition back to LP state after next REF
+ pwrStatePostRefresh = PWR_IDLE;
+}
+
+bool
+DRAMCtrl::Rank::isQueueEmpty() const
+{
+ // check commmands in Q based on current bus direction
+ bool no_queued_cmds = ((memory.busStateNext == READ) && (readEntries == 0))
+ || ((memory.busStateNext == WRITE) &&
+ (writeEntries == 0));
+ return no_queued_cmds;
+}
+
+void
+DRAMCtrl::Rank::checkDrainDone()
+{
+ // if this rank was waiting to drain it is now able to proceed to
+ // precharge
+ if (refreshState == REF_DRAIN) {
+ DPRINTF(DRAM, "Refresh drain done, now precharging\n");
+
+ refreshState = REF_PD_EXIT;
+
+ // hand control back to the refresh event loop
+ schedule(refreshEvent, curTick());
+ }
}
void
-DRAMCtrl::processRefreshEvent()
+DRAMCtrl::Rank::flushCmdList()
+{
+ // at the moment sort the list of commands and update the counters
+ // for DRAMPower libray when doing a refresh
+ sort(cmdList.begin(), cmdList.end(), DRAMCtrl::sortTime);
+
+ auto next_iter = cmdList.begin();
+ // push to commands to DRAMPower
+ for ( ; next_iter != cmdList.end() ; ++next_iter) {
+ Command cmd = *next_iter;
+ if (cmd.timeStamp <= curTick()) {
+ // Move all commands at or before curTick to DRAMPower
+ power.powerlib.doCommand(cmd.type, cmd.bank,
+ divCeil(cmd.timeStamp, memory.tCK) -
+ memory.timeStampOffset);
+ } else {
+ // done - found all commands at or before curTick()
+ // next_iter references the 1st command after curTick
+ break;
+ }
+ }
+ // reset cmdList to only contain commands after curTick
+ // if there are no commands after curTick, updated cmdList will be empty
+ // in this case, next_iter is cmdList.end()
+ cmdList.assign(next_iter, cmdList.end());
+}
+
+void
+DRAMCtrl::Rank::processActivateEvent()
+{
+ // we should transition to the active state as soon as any bank is active
+ if (pwrState != PWR_ACT)
+ // note that at this point numBanksActive could be back at
+ // zero again due to a precharge scheduled in the future
+ schedulePowerEvent(PWR_ACT, curTick());
+}
+
+void
+DRAMCtrl::Rank::processPrechargeEvent()
+{
+ // counter should at least indicate one outstanding request
+ // for this precharge
+ assert(outstandingEvents > 0);
+ // precharge complete, decrement count
+ --outstandingEvents;
+
+ // if we reached zero, then special conditions apply as we track
+ // if all banks are precharged for the power models
+ if (numBanksActive == 0) {
+ // no reads to this rank in the Q and no pending
+ // RD/WR or refresh commands
+ if (isQueueEmpty() && outstandingEvents == 0 &&
+ memory.enableDRAMPowerdown) {
+ // should still be in ACT state since bank still open
+ assert(pwrState == PWR_ACT);
+
+ // All banks closed - switch to precharge power down state.
+ DPRINTF(DRAMState, "Rank %d sleep at tick %d\n",
+ rank, curTick());
+ powerDownSleep(PWR_PRE_PDN, curTick());
+ } else {
+ // we should transition to the idle state when the last bank
+ // is precharged
+ schedulePowerEvent(PWR_IDLE, curTick());
+ }
+ }
+}
+
+void
+DRAMCtrl::Rank::processWriteDoneEvent()
+{
+ // counter should at least indicate one outstanding request
+ // for this write
+ assert(outstandingEvents > 0);
+ // Write transfer on bus has completed
+ // decrement per rank counter
+ --outstandingEvents;
+}
+
+void
+DRAMCtrl::Rank::processRefreshEvent()
{
// when first preparing the refresh, remember when it was due
- if (refreshState == REF_IDLE) {
+ if ((refreshState == REF_IDLE) || (refreshState == REF_SREF_EXIT)) {
// remember when the refresh is due
refreshDueAt = curTick();
// proceed to drain
refreshState = REF_DRAIN;
+ // make nonzero while refresh is pending to ensure
+ // power down and self-refresh are not entered
+ ++outstandingEvents;
+
DPRINTF(DRAM, "Refresh due\n");
}
- // let any scheduled read or write go ahead, after which it will
+ // let any scheduled read or write to the same rank go ahead,
+ // after which it will
// hand control back to this event loop
if (refreshState == REF_DRAIN) {
- if (nextReqEvent.scheduled()) {
+ // if a request is at the moment being handled and this request is
+ // accessing the current rank then wait for it to finish
+ if ((rank == memory.activeRank)
+ && (memory.nextReqEvent.scheduled())) {
// hand control over to the request loop until it is
// evaluated next
DPRINTF(DRAM, "Refresh awaiting draining\n");
+ return;
+ } else {
+ refreshState = REF_PD_EXIT;
+ }
+ }
+
+ // at this point, ensure that rank is not in a power-down state
+ if (refreshState == REF_PD_EXIT) {
+ // if rank was sleeping and we have't started exit process,
+ // wake-up for refresh
+ if (inLowPowerState) {
+ DPRINTF(DRAM, "Wake Up for refresh\n");
+ // save state and return after refresh completes
+ scheduleWakeUpEvent(memory.tXP);
return;
} else {
refreshState = REF_PRE;
// at this point, ensure that all banks are precharged
if (refreshState == REF_PRE) {
- // precharge any active bank if we are not already in the idle
- // state
- if (pwrState != PWR_IDLE) {
+ // precharge any active bank
+ if (numBanksActive != 0) {
// at the moment, we use a precharge all even if there is
// only a single bank open
DPRINTF(DRAM, "Precharging all\n");
// first determine when we can precharge
Tick pre_at = curTick();
- for (int i = 0; i < ranksPerChannel; i++) {
- for (int j = 0; j < banksPerRank; j++) {
- // respect both causality and any existing bank
- // constraints, some banks could already have a
- // (auto) precharge scheduled
- pre_at = std::max(banks[i][j].preAllowedAt, pre_at);
- }
+
+ for (auto &b : banks) {
+ // respect both causality and any existing bank
+ // constraints, some banks could already have a
+ // (auto) precharge scheduled
+ pre_at = std::max(b.preAllowedAt, pre_at);
}
- // make sure all banks are precharged, and for those that
+ // make sure all banks per rank are precharged, and for those that
// already are, update their availability
- Tick act_allowed_at = pre_at + tRP;
-
- for (int i = 0; i < ranksPerChannel; i++) {
- for (int j = 0; j < banksPerRank; j++) {
- if (banks[i][j].openRow != Bank::NO_ROW) {
- prechargeBank(banks[i][j], pre_at, false);
- } else {
- banks[i][j].actAllowedAt =
- std::max(banks[i][j].actAllowedAt, act_allowed_at);
- banks[i][j].preAllowedAt =
- std::max(banks[i][j].preAllowedAt, pre_at);
- }
+ Tick act_allowed_at = pre_at + memory.tRP;
+
+ for (auto &b : banks) {
+ if (b.openRow != Bank::NO_ROW) {
+ memory.prechargeBank(*this, b, pre_at, false);
+ } else {
+ b.actAllowedAt = std::max(b.actAllowedAt, act_allowed_at);
+ b.preAllowedAt = std::max(b.preAllowedAt, pre_at);
}
+ }
- // at the moment this affects all ranks
- rankPower[i].powerlib.doCommand(MemCommand::PREA, 0,
- divCeil(pre_at, tCK) -
- timeStampOffset);
+ // precharge all banks in rank
+ cmdList.push_back(Command(MemCommand::PREA, 0, pre_at));
- DPRINTF(DRAMPower, "%llu,PREA,0,%d\n", divCeil(pre_at, tCK) -
- timeStampOffset, i);
- }
- } else {
+ DPRINTF(DRAMPower, "%llu,PREA,0,%d\n",
+ divCeil(pre_at, memory.tCK) -
+ memory.timeStampOffset, rank);
+ } else if ((pwrState == PWR_IDLE) && (outstandingEvents == 1)) {
+ // Banks are closed, have transitioned to IDLE state, and
+ // no outstanding ACT,RD/WR,Auto-PRE sequence scheduled
DPRINTF(DRAM, "All banks already precharged, starting refresh\n");
- // go ahead and kick the power state machine into gear if
+ // go ahead and kick the power state machine into gear since
// we are already idle
schedulePowerEvent(PWR_REF, curTick());
+ } else {
+ // banks state is closed but haven't transitioned pwrState to IDLE
+ // or have outstanding ACT,RD/WR,Auto-PRE sequence scheduled
+ // should have outstanding precharge event in this case
+ assert(prechargeEvent.scheduled());
+ // will start refresh when pwrState transitions to IDLE
}
- refreshState = REF_RUN;
assert(numBanksActive == 0);
// wait for all banks to be precharged, at which point the
}
// last but not least we perform the actual refresh
- if (refreshState == REF_RUN) {
+ if (refreshState == REF_START) {
// should never get here with any banks active
assert(numBanksActive == 0);
assert(pwrState == PWR_REF);
- Tick ref_done_at = curTick() + tRFC;
-
- for (int i = 0; i < ranksPerChannel; i++) {
- for (int j = 0; j < banksPerRank; j++) {
- banks[i][j].actAllowedAt = ref_done_at;
- }
+ Tick ref_done_at = curTick() + memory.tRFC;
- // at the moment this affects all ranks
- rankPower[i].powerlib.doCommand(MemCommand::REF, 0,
- divCeil(curTick(), tCK) -
- timeStampOffset);
-
- // at the moment sort the list of commands and update the counters
- // for DRAMPower libray when doing a refresh
- sort(rankPower[i].powerlib.cmdList.begin(),
- rankPower[i].powerlib.cmdList.end(), DRAMCtrl::sortTime);
-
- // update the counters for DRAMPower, passing false to
- // indicate that this is not the last command in the
- // list. DRAMPower requires this information for the
- // correct calculation of the background energy at the end
- // of the simulation. Ideally we would want to call this
- // function with true once at the end of the
- // simulation. However, the discarded energy is extremly
- // small and does not effect the final results.
- rankPower[i].powerlib.updateCounters(false);
-
- // call the energy function
- rankPower[i].powerlib.calcEnergy();
-
- // Update the stats
- updatePowerStats(i);
-
- DPRINTF(DRAMPower, "%llu,REF,0,%d\n", divCeil(curTick(), tCK) -
- timeStampOffset, i);
+ for (auto &b : banks) {
+ b.actAllowedAt = ref_done_at;
}
+ // at the moment this affects all ranks
+ cmdList.push_back(Command(MemCommand::REF, 0, curTick()));
+
+ // Update the stats
+ updatePowerStats();
+
+ DPRINTF(DRAMPower, "%llu,REF,0,%d\n", divCeil(curTick(), memory.tCK) -
+ memory.timeStampOffset, rank);
+
+ // Update for next refresh
+ refreshDueAt += memory.tREFI;
+
// make sure we did not wait so long that we cannot make up
// for it
- if (refreshDueAt + tREFI < ref_done_at) {
+ if (refreshDueAt < ref_done_at) {
fatal("Refresh was delayed so long we cannot catch up\n");
}
- // compensate for the delay in actually performing the refresh
- // when scheduling the next one
- schedule(refreshEvent, refreshDueAt + tREFI - tRP);
+ // Run the refresh and schedule event to transition power states
+ // when refresh completes
+ refreshState = REF_RUN;
+ schedule(refreshEvent, ref_done_at);
+ return;
+ }
+
+ if (refreshState == REF_RUN) {
+ // should never get here with any banks active
+ assert(numBanksActive == 0);
+ assert(pwrState == PWR_REF);
assert(!powerEvent.scheduled());
- // move to the idle power state once the refresh is done, this
- // will also move the refresh state machine to the refresh
- // idle state
- schedulePowerEvent(PWR_IDLE, ref_done_at);
+ if ((memory.drainState() == DrainState::Draining) ||
+ (memory.drainState() == DrainState::Drained)) {
+ // if draining, do not re-enter low-power mode.
+ // simply go to IDLE and wait
+ schedulePowerEvent(PWR_IDLE, curTick());
+ } else {
+ // At the moment, we sleep when the refresh ends and wait to be
+ // woken up again if previously in a low-power state.
+ if (pwrStatePostRefresh != PWR_IDLE) {
+ // power State should be power Refresh
+ assert(pwrState == PWR_REF);
+ DPRINTF(DRAMState, "Rank %d sleeping after refresh and was in "
+ "power state %d before refreshing\n", rank,
+ pwrStatePostRefresh);
+ powerDownSleep(pwrState, curTick());
+
+ // Force PRE power-down if there are no outstanding commands
+ // in Q after refresh.
+ } else if (isQueueEmpty() && memory.enableDRAMPowerdown) {
+ // still have refresh event outstanding but there should
+ // be no other events outstanding
+ assert(outstandingEvents == 1);
+ DPRINTF(DRAMState, "Rank %d sleeping after refresh but was NOT"
+ " in a low power state before refreshing\n", rank);
+ powerDownSleep(PWR_PRE_PDN, curTick());
- DPRINTF(DRAMState, "Refresh done at %llu and next refresh at %llu\n",
- ref_done_at, refreshDueAt + tREFI);
+ } else {
+ // move to the idle power state once the refresh is done, this
+ // will also move the refresh state machine to the refresh
+ // idle state
+ schedulePowerEvent(PWR_IDLE, curTick());
+ }
+ }
+
+ // At this point, we have completed the current refresh.
+ // In the SREF bypass case, we do not get to this state in the
+ // refresh STM and therefore can always schedule next event.
+ // Compensate for the delay in actually performing the refresh
+ // when scheduling the next one
+ schedule(refreshEvent, refreshDueAt - memory.tRP);
+
+ DPRINTF(DRAMState, "Refresh done at %llu and next refresh"
+ " at %llu\n", curTick(), refreshDueAt);
}
}
void
-DRAMCtrl::schedulePowerEvent(PowerState pwr_state, Tick tick)
+DRAMCtrl::Rank::schedulePowerEvent(PowerState pwr_state, Tick tick)
{
// respect causality
assert(tick >= curTick());
}
void
-DRAMCtrl::processPowerEvent()
+DRAMCtrl::Rank::powerDownSleep(PowerState pwr_state, Tick tick)
+{
+ // if low power state is active low, schedule to active low power state.
+ // in reality tCKE is needed to enter active low power. This is neglected
+ // here and could be added in the future.
+ if (pwr_state == PWR_ACT_PDN) {
+ schedulePowerEvent(pwr_state, tick);
+ // push command to DRAMPower
+ cmdList.push_back(Command(MemCommand::PDN_F_ACT, 0, tick));
+ DPRINTF(DRAMPower, "%llu,PDN_F_ACT,0,%d\n", divCeil(tick,
+ memory.tCK) - memory.timeStampOffset, rank);
+ } else if (pwr_state == PWR_PRE_PDN) {
+ // if low power state is precharge low, schedule to precharge low
+ // power state. In reality tCKE is needed to enter active low power.
+ // This is neglected here.
+ schedulePowerEvent(pwr_state, tick);
+ //push Command to DRAMPower
+ cmdList.push_back(Command(MemCommand::PDN_F_PRE, 0, tick));
+ DPRINTF(DRAMPower, "%llu,PDN_F_PRE,0,%d\n", divCeil(tick,
+ memory.tCK) - memory.timeStampOffset, rank);
+ } else if (pwr_state == PWR_REF) {
+ // if a refresh just occurred
+ // transition to PRE_PDN now that all banks are closed
+ // precharge power down requires tCKE to enter. For simplicity
+ // this is not considered.
+ schedulePowerEvent(PWR_PRE_PDN, tick);
+ //push Command to DRAMPower
+ cmdList.push_back(Command(MemCommand::PDN_F_PRE, 0, tick));
+ DPRINTF(DRAMPower, "%llu,PDN_F_PRE,0,%d\n", divCeil(tick,
+ memory.tCK) - memory.timeStampOffset, rank);
+ } else if (pwr_state == PWR_SREF) {
+ // should only enter SREF after PRE-PD wakeup to do a refresh
+ assert(pwrStatePostRefresh == PWR_PRE_PDN);
+ // self refresh requires time tCKESR to enter. For simplicity,
+ // this is not considered.
+ schedulePowerEvent(PWR_SREF, tick);
+ // push Command to DRAMPower
+ cmdList.push_back(Command(MemCommand::SREN, 0, tick));
+ DPRINTF(DRAMPower, "%llu,SREN,0,%d\n", divCeil(tick,
+ memory.tCK) - memory.timeStampOffset, rank);
+ }
+ // Ensure that we don't power-down and back up in same tick
+ // Once we commit to PD entry, do it and wait for at least 1tCK
+ // This could be replaced with tCKE if/when that is added to the model
+ wakeUpAllowedAt = tick + memory.tCK;
+
+ // Transitioning to a low power state, set flag
+ inLowPowerState = true;
+}
+
+void
+DRAMCtrl::Rank::scheduleWakeUpEvent(Tick exit_delay)
+{
+ Tick wake_up_tick = std::max(curTick(), wakeUpAllowedAt);
+
+ DPRINTF(DRAMState, "Scheduling wake-up for rank %d at tick %d\n",
+ rank, wake_up_tick);
+
+ // if waking for refresh, hold previous state
+ // else reset state back to IDLE
+ if (refreshState == REF_PD_EXIT) {
+ pwrStatePostRefresh = pwrState;
+ } else {
+ // don't automatically transition back to LP state after next REF
+ pwrStatePostRefresh = PWR_IDLE;
+ }
+
+ // schedule wake-up with event to ensure entry has completed before
+ // we try to wake-up
+ schedule(wakeUpEvent, wake_up_tick);
+
+ for (auto &b : banks) {
+ // respect both causality and any existing bank
+ // constraints, some banks could already have a
+ // (auto) precharge scheduled
+ b.wrAllowedAt = std::max(wake_up_tick + exit_delay, b.wrAllowedAt);
+ b.rdAllowedAt = std::max(wake_up_tick + exit_delay, b.rdAllowedAt);
+ b.preAllowedAt = std::max(wake_up_tick + exit_delay, b.preAllowedAt);
+ b.actAllowedAt = std::max(wake_up_tick + exit_delay, b.actAllowedAt);
+ }
+ // Transitioning out of low power state, clear flag
+ inLowPowerState = false;
+
+ // push to DRAMPower
+ // use pwrStateTrans for cases where we have a power event scheduled
+ // to enter low power that has not yet been processed
+ if (pwrStateTrans == PWR_ACT_PDN) {
+ cmdList.push_back(Command(MemCommand::PUP_ACT, 0, wake_up_tick));
+ DPRINTF(DRAMPower, "%llu,PUP_ACT,0,%d\n", divCeil(wake_up_tick,
+ memory.tCK) - memory.timeStampOffset, rank);
+
+ } else if (pwrStateTrans == PWR_PRE_PDN) {
+ cmdList.push_back(Command(MemCommand::PUP_PRE, 0, wake_up_tick));
+ DPRINTF(DRAMPower, "%llu,PUP_PRE,0,%d\n", divCeil(wake_up_tick,
+ memory.tCK) - memory.timeStampOffset, rank);
+ } else if (pwrStateTrans == PWR_SREF) {
+ cmdList.push_back(Command(MemCommand::SREX, 0, wake_up_tick));
+ DPRINTF(DRAMPower, "%llu,SREX,0,%d\n", divCeil(wake_up_tick,
+ memory.tCK) - memory.timeStampOffset, rank);
+ }
+}
+
+void
+DRAMCtrl::Rank::processWakeUpEvent()
+{
+ // Should be in a power-down or self-refresh state
+ assert((pwrState == PWR_ACT_PDN) || (pwrState == PWR_PRE_PDN) ||
+ (pwrState == PWR_SREF));
+
+ // Check current state to determine transition state
+ if (pwrState == PWR_ACT_PDN) {
+ // banks still open, transition to PWR_ACT
+ schedulePowerEvent(PWR_ACT, curTick());
+ } else {
+ // transitioning from a precharge power-down or self-refresh state
+ // banks are closed - transition to PWR_IDLE
+ schedulePowerEvent(PWR_IDLE, curTick());
+ }
+}
+
+void
+DRAMCtrl::Rank::processPowerEvent()
{
+ assert(curTick() >= pwrStateTick);
// remember where we were, and for how long
Tick duration = curTick() - pwrStateTick;
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)) {
+ stats.totalIdleTime += duration;
+ }
pwrState = pwrStateTrans;
pwrStateTick = curTick();
- if (pwrState == PWR_IDLE) {
- DPRINTF(DRAMState, "All banks precharged\n");
+ // if rank was refreshing, make sure to start scheduling requests again
+ if (prev_state == PWR_REF) {
+ // bus IDLED prior to REF
+ // counter should be one for refresh command only
+ assert(outstandingEvents == 1);
+ // REF complete, decrement count and go back to IDLE
+ --outstandingEvents;
+ refreshState = REF_IDLE;
+
+ DPRINTF(DRAMState, "Was refreshing for %llu ticks\n", duration);
+ // if moving back to power-down after refresh
+ if (pwrState != PWR_IDLE) {
+ assert(pwrState == PWR_PRE_PDN);
+ DPRINTF(DRAMState, "Switching to power down state after refreshing"
+ " rank %d at %llu tick\n", rank, curTick());
+ }
- // if we were refreshing, make sure we start scheduling requests again
- if (prev_state == PWR_REF) {
- DPRINTF(DRAMState, "Was refreshing for %llu ticks\n", duration);
- assert(pwrState == PWR_IDLE);
+ // completed refresh event, ensure next request is scheduled
+ if (!memory.nextReqEvent.scheduled()) {
+ DPRINTF(DRAM, "Scheduling next request after refreshing"
+ " rank %d\n", rank);
+ schedule(memory.nextReqEvent, curTick());
+ }
+ }
- // kick things into action again
- refreshState = REF_IDLE;
- assert(!nextReqEvent.scheduled());
- schedule(nextReqEvent, curTick());
- } else {
- assert(prev_state == PWR_ACT);
+ if ((pwrState == PWR_ACT) && (refreshState == REF_PD_EXIT)) {
+ // have exited ACT PD
+ assert(prev_state == PWR_ACT_PDN);
+ // go back to REF event and close banks
+ refreshState = REF_PRE;
+ schedule(refreshEvent, curTick());
+ } else if (pwrState == PWR_IDLE) {
+ DPRINTF(DRAMState, "All banks precharged\n");
+ if (prev_state == PWR_SREF) {
+ // set refresh state to REF_SREF_EXIT, ensuring inRefIdleState
+ // continues to return false during tXS after SREF exit
+ // Schedule a refresh which kicks things back into action
+ // when it finishes
+ refreshState = REF_SREF_EXIT;
+ schedule(refreshEvent, curTick() + memory.tXS);
+ } else {
// if we have a pending refresh, and are now moving to
- // the idle state, direclty transition to a refresh
- if (refreshState == REF_RUN) {
- // there should be nothing waiting at this point
- assert(!powerEvent.scheduled());
-
- // update the state in zero time and proceed below
- pwrState = PWR_REF;
+ // the idle state, directly transition to, or schedule refresh
+ if ((refreshState == REF_PRE) || (refreshState == REF_PD_EXIT)) {
+ // ensure refresh is restarted only after final PRE command.
+ // do not restart refresh if controller is in an intermediate
+ // state, after PRE_PDN exit, when banks are IDLE but an
+ // ACT is scheduled.
+ if (!activateEvent.scheduled()) {
+ // there should be nothing waiting at this point
+ assert(!powerEvent.scheduled());
+ if (refreshState == REF_PD_EXIT) {
+ // exiting PRE PD, will be in IDLE until tXP expires
+ // and then should transition to PWR_REF state
+ assert(prev_state == PWR_PRE_PDN);
+ schedulePowerEvent(PWR_REF, curTick() + memory.tXP);
+ } else if (refreshState == REF_PRE) {
+ // can directly move to PWR_REF state and proceed below
+ pwrState = PWR_REF;
+ }
+ } else {
+ // must have PRE scheduled to transition back to IDLE
+ // and re-kick off refresh
+ assert(prechargeEvent.scheduled());
+ }
}
}
}
- // we transition to the refresh state, let the refresh state
- // machine know of this state update and let it deal with the
- // scheduling of the next power state transition as well as the
- // following refresh
+ // transition to the refresh state and re-start refresh process
+ // refresh state machine will schedule the next power state transition
if (pwrState == PWR_REF) {
- DPRINTF(DRAMState, "Refreshing\n");
- // kick the refresh event loop into action again, and that
- // in turn will schedule a transition to the idle power
- // state once the refresh is done
- assert(refreshState == REF_RUN);
- processRefreshEvent();
+ // completed final PRE for refresh or exiting power-down
+ assert(refreshState == REF_PRE || refreshState == REF_PD_EXIT);
+
+ // exited PRE PD for refresh, with no pending commands
+ // bypass auto-refresh and go straight to SREF, where memory
+ // will issue refresh immediately upon entry
+ if (pwrStatePostRefresh == PWR_PRE_PDN && isQueueEmpty() &&
+ (memory.drainState() != DrainState::Draining) &&
+ (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());
+
+ // Since refresh was bypassed, remove event by decrementing count
+ assert(outstandingEvents == 1);
+ --outstandingEvents;
+
+ // reset state back to IDLE temporarily until SREF is entered
+ pwrState = PWR_IDLE;
+
+ // Not bypassing refresh for SREF entry
+ } else {
+ DPRINTF(DRAMState, "Refreshing\n");
+
+ // there should be nothing waiting at this point
+ assert(!powerEvent.scheduled());
+
+ // kick the refresh event loop into action again, and that
+ // in turn will schedule a transition to the idle power
+ // state once the refresh is done
+ schedule(refreshEvent, curTick());
+
+ // Banks transitioned to IDLE, start REF
+ refreshState = REF_START;
+ }
}
+
}
void
-DRAMCtrl::updatePowerStats(uint8_t rank)
+DRAMCtrl::Rank::updatePowerStats()
{
- // Get the energy and power from DRAMPower
- Data::MemoryPowerModel::Energy energy =
- rankPower[rank].powerlib.getEnergy();
- Data::MemoryPowerModel::Power power =
- rankPower[rank].powerlib.getPower();
-
- actEnergy[rank] = energy.act_energy * devicesPerRank;
- preEnergy[rank] = energy.pre_energy * devicesPerRank;
- readEnergy[rank] = energy.read_energy * devicesPerRank;
- writeEnergy[rank] = energy.write_energy * devicesPerRank;
- refreshEnergy[rank] = energy.ref_energy * devicesPerRank;
- actBackEnergy[rank] = energy.act_stdby_energy * devicesPerRank;
- preBackEnergy[rank] = energy.pre_stdby_energy * devicesPerRank;
- totalEnergy[rank] = energy.total_energy * devicesPerRank;
- averagePower[rank] = power.average_power * devicesPerRank;
+ // All commands up to refresh have completed
+ // flush cmdList to DRAMPower
+ flushCmdList();
+
+ // Call the function that calculates window energy at intermediate update
+ // events like at refresh, stats dump as well as at simulation exit.
+ // Window starts at the last time the calcWindowEnergy function was called
+ // and is upto current time.
+ power.powerlib.calcWindowEnergy(divCeil(curTick(), memory.tCK) -
+ memory.timeStampOffset);
+
+ // Get the energy from DRAMPower
+ Data::MemoryPowerModel::Energy energy = power.powerlib.getEnergy();
+
+ // The energy components inside the power lib are calculated over
+ // the window so accumulate into the corresponding gem5 stat
+ 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.
+ 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
+ stats.averagePower = (stats.totalEnergy.value() /
+ (curTick() - memory.lastStatsResetTick)) *
+ (SimClock::Frequency / 1000000000.0);
}
void
-DRAMCtrl::regStats()
+DRAMCtrl::Rank::computeStats()
{
- using namespace Stats;
-
- AbstractMemory::regStats();
-
- 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");
+ DPRINTF(DRAM,"Computing stats due to a dump callback\n");
- writeBursts
- .name(name() + ".writeBursts")
- .desc("Number of DRAM write bursts, "
- "including those merged in the write queue");
+ // Update the stats
+ updatePowerStats();
- servicedByWrQ
- .name(name() + ".servicedByWrQ")
- .desc("Number of DRAM read bursts serviced by the write queue");
+ // final update of power state times
+ stats.memoryStateTime[pwrState] += (curTick() - pwrStateTick);
+ pwrStateTick = curTick();
+}
- mergedWrBursts
- .name(name() + ".mergedWrBursts")
- .desc("Number of DRAM write bursts merged with an existing one");
+void
+DRAMCtrl::Rank::resetStats() {
+ // The only way to clear the counters in DRAMPower is to call
+ // calcWindowEnergy function as that then calls clearCounters. The
+ // clearCounters method itself is private.
+ power.powerlib.calcWindowEnergy(divCeil(curTick(), memory.tCK) -
+ memory.timeStampOffset);
- 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");
+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")
+{
+}
- perBankWrBursts
- .init(banksPerRank * ranksPerChannel)
- .name(name() + ".perBankWrBursts")
- .desc("Per bank write bursts");
+void
+DRAMCtrl::DRAMStats::regStats()
+{
+ using namespace Stats;
- 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);
- avgQLat = totQLat / (readBursts - servicedByWrQ);
-
- avgBusLat
- .name(name() + ".avgBusLat")
- .desc("Average bus latency per DRAM burst")
- .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 = 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
- .name(name() + ".peakBW")
- .desc("Theoretical peak bandwidth in MiByte/s")
- .precision(2);
-
- peakBW = (SimClock::Frequency / tBURST) * burstSize / 1000000;
-
- busUtil
- .name(name() + ".busUtil")
- .desc("Data bus utilization in percentage")
- .precision(2);
+ peakBW = (SimClock::Frequency / dram.tBURST) * dram.burstSize / 1000000;
busUtil = (avgRdBW + avgWrBW) / peakBW * 100;
- totGap
- .name(name() + ".totGap")
- .desc("Total gap between requests");
+ avgGap = totGap / (readReqs + writeReqs);
- avgGap
- .name(name() + ".avgGap")
- .desc("Average gap between requests")
- .precision(2);
+ busUtilRead = avgRdBW / peakBW * 100;
+ busUtilWrite = avgWrBW / peakBW * 100;
- avgGap = totGap / (readReqs + writeReqs);
+ pageHitRate = (writeRowHits + readRowHits) /
+ (writeBursts - mergedWrBursts + readBursts - servicedByWrQ) * 100;
- // Stats for DRAM Power calculation based on Micron datasheet
- busUtilRead
- .name(name() + ".busUtilRead")
- .desc("Data bus utilization in percentage for reads")
- .precision(2);
+ masterReadRate = masterReadBytes / simSeconds;
+ masterWriteRate = masterWriteBytes / simSeconds;
+ masterReadAvgLat = masterReadTotalLat / masterReadAccesses;
+ masterWriteAvgLat = masterWriteTotalLat / masterWriteAccesses;
+}
- busUtilRead = avgRdBW / peakBW * 100;
+void
+DRAMCtrl::DRAMStats::resetStats()
+{
+ dram.lastStatsResetTick = curTick();
+}
- busUtilWrite
- .name(name() + ".busUtilWrite")
- .desc("Data bus utilization in percentage for writes")
- .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")
+{
+}
- busUtilWrite = avgWrBW / 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");
+}
- pageHitRate
- .name(name() + ".pageHitRate")
- .desc("Row buffer hit rate, read and write combined")
- .precision(2);
+void
+DRAMCtrl::RankStats::resetStats()
+{
+ Stats::Group::resetStats();
- pageHitRate = (writeRowHits + readRowHits) /
- (writeBursts - mergedWrBursts + readBursts - servicedByWrQ) * 100;
+ rank.resetStats();
+}
- pwrStateTime
- .init(5)
- .name(name() + ".memoryStateTime")
- .desc("Time in different power states");
- pwrStateTime.subname(0, "IDLE");
- pwrStateTime.subname(1, "REF");
- pwrStateTime.subname(2, "PRE_PDN");
- pwrStateTime.subname(3, "ACT");
- pwrStateTime.subname(4, "ACT_PDN");
-
- actEnergy
- .init(ranksPerChannel)
- .name(name() + ".actEnergy")
- .desc("Energy for activate commands per rank (pJ)");
-
- preEnergy
- .init(ranksPerChannel)
- .name(name() + ".preEnergy")
- .desc("Energy for precharge commands per rank (pJ)");
-
- readEnergy
- .init(ranksPerChannel)
- .name(name() + ".readEnergy")
- .desc("Energy for read commands per rank (pJ)");
-
- writeEnergy
- .init(ranksPerChannel)
- .name(name() + ".writeEnergy")
- .desc("Energy for write commands per rank (pJ)");
-
- refreshEnergy
- .init(ranksPerChannel)
- .name(name() + ".refreshEnergy")
- .desc("Energy for refresh commands per rank (pJ)");
-
- actBackEnergy
- .init(ranksPerChannel)
- .name(name() + ".actBackEnergy")
- .desc("Energy for active background per rank (pJ)");
-
- preBackEnergy
- .init(ranksPerChannel)
- .name(name() + ".preBackEnergy")
- .desc("Energy for precharge background per rank (pJ)");
-
- totalEnergy
- .init(ranksPerChannel)
- .name(name() + ".totalEnergy")
- .desc("Total energy per rank (pJ)");
-
- averagePower
- .init(ranksPerChannel)
- .name(name() + ".averagePower")
- .desc("Core power per rank (mW)");
+void
+DRAMCtrl::RankStats::preDumpStats()
+{
+ Stats::Group::preDumpStats();
+
+ 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;
}
}
-unsigned int
-DRAMCtrl::drain(DrainManager *dm)
+DrainState
+DRAMCtrl::drain()
{
- unsigned int count = port.drain(dm);
-
// 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());
- ++count;
- drainManager = dm;
- // the only part that is not drained automatically over time
+ // 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());
}
+
+ // also need to kick off events to exit self-refresh
+ for (auto r : ranks) {
+ // force self-refresh exit, which in turn will issue auto-refresh
+ if (r->pwrState == PWR_SREF) {
+ DPRINTF(DRAM,"Rank%d: Forcing self-refresh wakeup in drain\n",
+ r->rank);
+ r->scheduleWakeUpEvent(tXS);
+ }
+ }
+
+ return DrainState::Draining;
+ } else {
+ return DrainState::Drained;
+ }
+}
+
+bool
+DRAMCtrl::allRanksDrained() const
+{
+ // true until proven false
+ bool all_ranks_drained = true;
+ for (auto r : ranks) {
+ // then verify that the power state is IDLE ensuring all banks are
+ // closed and rank is not in a low power state. Also verify that rank
+ // is idle from a refresh point of view.
+ all_ranks_drained = r->inPwrIdleState() && r->inRefIdleState() &&
+ all_ranks_drained;
+ }
+ return all_ranks_drained;
+}
+
+void
+DRAMCtrl::drainResume()
+{
+ if (!isTimingMode && system()->isTimingMode()) {
+ // if we switched to timing mode, kick things into action,
+ // and behave as if we restored from a checkpoint
+ startup();
+ } else if (isTimingMode && !system()->isTimingMode()) {
+ // if we switch from timing mode, stop the refresh events to
+ // not cause issues with KVM
+ for (auto r : ranks) {
+ r->suspend();
+ }
}
- if (count)
- setDrainState(Drainable::Draining);
- else
- setDrainState(Drainable::Drained);
- return count;
+ // update the mode
+ isTimingMode = system()->isTimingMode();
}
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.
projects / gem5.git / blobdiff