网站建设面谈销售话术,网站建设中提示页面,网站建设租用服务器,一个网站不兼容ie怎么做代码流程
下图就是全部. 剩下文字部分是细节补充,但是内容不变: bash调用python,用python配置好configuration, 一个cpu每个tick运行一次,requestport发出pkt.
bash 启动 python文件并配置
./build/NULL/gem5.debug configs/example/garnet_synth_traffic.py \--num-cpus…代码流程
下图就是全部. 剩下文字部分是细节补充,但是内容不变: bash调用python,用python配置好configuration, 一个cpu每个tick运行一次,requestport发出pkt.
bash 启动 python文件并配置
./build/NULL/gem5.debug configs/example/garnet_synth_traffic.py \--num-cpus16 \--num-dirs16 \--networkgarnet \--topologyMesh_XY \--mesh-rows4 \--sim-cycles1000000 --inj-vnet0 \--syntheticuniform_random \--injectionrate1 \--single-sender-id0代码启动 garnet_synth_traffic.py
代码直接用了 GarnetSyntheticTraffic()函数.
cpus [GarnetSyntheticTraffic(num_packets_maxargs.num_packets_max,single_senderargs.single_sender_id,single_destargs.single_dest_id,sim_cyclesargs.sim_cycles,traffic_typeargs.synthetic,inj_rateargs.injectionrate,inj_vnetargs.inj_vnet,precisionargs.precision,num_destargs.num_dirs,)for i in range(args.num_cpus)
] 打印看看cpu类型
for cpu in cpus:print(yzzzzdebugcpus , cpu.type, m5.curTick(),cpu.inj_rate,cpu.inj_vnet,cpu.num_dest)可以看到cpu.type是 GarnetSyntheticTraffic. GarnetSyntheticTraffic()函数来自 src/cpu/testers/garnet_synthetic_traffic/GarnetSyntheticTraffic.py
GarnetSyntheticTraffic.py 代码定义了很多 python 里可以 cpu.num_dest 之类调用的子类.
class GarnetSyntheticTraffic(ClockedObject):type GarnetSyntheticTrafficcxx_header (cpu/testers/garnet_synthetic_traffic/GarnetSyntheticTraffic.hh)cxx_class gem5::GarnetSyntheticTrafficblock_offset Param.Int(6, block offset in bits)num_dest Param.Int(1, Number of Destinations)memory_size Param.Int(65536, memory size)sim_cycles Param.Int(1000, Number of simulation cycles)num_packets_max Param.Int(-1,Max number of packets to send. \Default is to keep sending till simulation ends,)single_sender Param.Int(-1,Send only from this node. \By default every node sends,)single_dest Param.Int(-1,Send only to this dest. \Default depends on traffic_type,)traffic_type Param.String(uniform_random, Traffic type)inj_rate Param.Float(0.1, Packet injection rate)inj_vnet Param.Int(-1,Vnet to inject in. \0 and 1 are 1-flit, 2 is 5-flit. \Default is to inject in all three vnets,)precision Param.Int(3,Number of digits of precision \after decimal point,)response_limit Param.Cycles(5000000,Cycles before exiting \due to lack of progress,)test RequestPort(Port to the memory system to test)system Param.System(Parent.any, System we belong to)
然后cpu变成了system的一部分,system System(cpucpus, mem_ranges[AddrRange(args.mem_size)]) 注意,这里print(\nyzzzzdebugsystem ,system.mem_mode )还是atomic. system变成了root的一部分 root Root(full_systemFalse, systemsystem) root.system.mem_mode “timing” 这里额外设置为timing.
cpp代码 , cpu每一个tick执行一次 tick()
src/cpu/testers/garnet_synthetic_traffic/GarnetSyntheticTraffic.hh // main simulation loop (one cycle)void tick();void
GarnetSyntheticTraffic::tick(){...if (senderEnable)generatePkt();
}void
GarnetSyntheticTraffic::generatePkt()
{...sendPkt(pkt);
}
void
GarnetSyntheticTraffic::sendPkt(PacketPtr pkt)
{if (!cachePort.sendTimingReq(pkt)) {retryPkt pkt; // RubyPort will retry sending}std::coutcoutyzzzzzdebug cachePort simCycles curTick() std::endl;numPacketsSent;
}tick()变成了 cachePort.sendTimingReq(pkt).
cachePort.sendTimingReq(pkt) 到底是什么
RequestPort发送一次pkt
通过 cacheport-CpuPort-RequestPort, tick()函数调用 generatePkt() 函数,再调用sendTimingReq.
inline bool
RequestPort::sendTimingReq(PacketPtr pkt)
{try {addTrace(pkt);bool succ TimingRequestProtocol::sendReq(_responsePort, pkt);//下面是我自己加的//std::coutcoutdebugyzzzzRequestPort::sendTimingReq succ curTick()std::endl;if (!succ)removeTrace(pkt);return succ;} catch (UnboundPortException) {reportUnbound();}
}我加了一行输出,把这行代码解除注释后,运行的命令行如下:
./build/NULL/gem5.debug configs/example/garnet_synth_traffic.py \--num-cpus16 \--num-dirs16 \--networkgarnet \--topologyMesh_XY \--mesh-rows4 \--sim-cycles1000000 --inj-vnet0 \--syntheticuniform_random \--injectionrate1 \--single-sender-id0跑出来的结果是: 可以看到,每1000 个tick,这个requestport都会发送一个pkt,而且返回的succ是1.
纯虚函数 virtual bool recvTimingReq
下一步, sendReq变成了peerrecvTimingReq(pkt); 我们发现peer-recvTimingReq是一个复杂的部分,因为他是timing.hh里的纯虚函数,是不固定的,除非我们知道派生类是什么. 纯虚函数: /*** Receive a timing request from the peer.*/virtual bool recvTimingReq(PacketPtr pkt) 0;src/mem/ruby/system/RubyPort.cc 的 RubyPort::MemResponsePort::recvTimingReq(PacketPtr pkt)
找到了! 在下方代码加入打印代码,输出的结果验证了,调用的是 RubyPort::MemResponsePort::recvTimingReq(PacketPtr pkt). 其实用vscode搜 recvTimingReq(会有很多cc文件里有例化,大概二三十个吧,给每一个都加上,编译,运行,就可以知道了. 缺点就是这个方法有点笨.
RubyPort::MemResponsePort::recvTimingReq 其实是 submit rubyrequest
bool
RubyPort::MemResponsePort::recvTimingReq(PacketPtr pkt)
{ std::coutdebugyzzzwhichrecvTimingReq?src/mem/ruby/system/rubyport.cc/memresponseportstd::endl;DPRINTF(RubyPort, Timing request for address %#x on port %d\n,pkt-getAddr(), id);if (pkt-cacheResponding())panic(RubyPort should never see request with the cacheResponding flag set\n);// ruby doesnt support cache maintenance operations at the// moment, as a workaround, we respond right awayif (pkt-req-isCacheMaintenance()) {warn_once(Cache maintenance operations are not supported in Ruby.\n);pkt-makeResponse();schedTimingResp(pkt, curTick());std::coutdebugyzzzthisReqIs pkt-req-isCacheMaintenance()std::endl;return true;}// Check for pio requests and directly send them to the dedicated// pio port.if (pkt-cmd ! MemCmd::MemSyncReq) {if (!pkt-req-isMemMgmt() !isPhysMemAddress(pkt)) {assert(owner.memRequestPort.isConnected());DPRINTF(RubyPort, Request address %#x assumed to be a pio address\n, pkt-getAddr());// Save the port in the sender state object to be used later to// route the responsepkt-pushSenderState(new SenderState(this));// send next cycleRubySystem *rs owner.m_ruby_system;owner.memRequestPort.schedTimingReq(pkt,curTick() rs-clockPeriod());std::coutdebugyzzzthisReqIs pkt-cmd ! MemCmd::MemSyncReqstd::endl;return true;}}// Save the port in the sender state object to be used later to// route the responsepkt-pushSenderState(new SenderState(this));// Submit the ruby requestRequestStatus requestStatus owner.makeRequest(pkt);// If the request successfully issued then we should return true.// Otherwise, we need to tell the port to retry at a later point// and return false.if (requestStatus RequestStatus_Issued) {DPRINTF(RubyPort, Request %s 0x%x issued\n, pkt-cmdString(),pkt-getAddr());std::coutdebugyzzzthisReqIs submit the ruby requeststd::endl;return true;}// pop off sender state as this request failed to issueSenderState *ss safe_castSenderState *(pkt-popSenderState());delete ss;if (pkt-cmd ! MemCmd::MemSyncReq) {DPRINTF(RubyPort,Request %s for address %#x did not issue because %s\n,pkt-cmdString(), pkt-getAddr(),RequestStatus_to_string(requestStatus));}addToRetryList();return false;
}submit rubyrequest 的owener是 rubyport
这里的owener是 RubyPort. 注意下图左边的两个竖线,仔细看,他是在RubyPort的public下面的. 也就是说,rubyPort下定义了class MemResponsePort,还定义了每个RubyPort都有 的makeRequest(). 这里给的虚函数,需要派生类来定义.
直觉告诉我们,sequencer 会有 makeRequest
src/mem/ruby/system/Sequencer.hh
RequestStatus makeRequest(PacketPtr pkt) override;src/mem/ruby/system/Sequencer.cc RequestStatus
Sequencer::makeRequest(PacketPtr pkt)
{std::coutdebugyzzzz Sequencer::makeRequest endl;// HTM abort signals must be allowed to reach the Sequencer// the same cycle they are issued. They cannot be retried.if ((m_outstanding_count m_max_outstanding_requests) !pkt-req-isHTMAbort()) {return RequestStatus_BufferFull;}RubyRequestType primary_type RubyRequestType_NULL;RubyRequestType secondary_type RubyRequestType_NULL;if (pkt-isLLSC()) {// LL/SC instructions need to be handled carefully by the cache// coherence protocol to ensure they follow the proper semantics. In// particular, by identifying the operations as atomic, the protocol// should understand that migratory sharing optimizations should not// be performed (i.e. a load between the LL and SC should not steal// away exclusive permission).//// The following logic works correctly with the semantics// of armV8 LDEX/STEX instructions.if (pkt-isWrite()) {DPRINTF(RubySequencer, Issuing SC\n);primary_type RubyRequestType_Store_Conditional;
#if defined (PROTOCOL_MESI_Three_Level) || defined (PROTOCOL_MESI_Three_Level_HTM)secondary_type RubyRequestType_Store_Conditional;
#elsesecondary_type RubyRequestType_ST;
#endif} else {DPRINTF(RubySequencer, Issuing LL\n);assert(pkt-isRead());primary_type RubyRequestType_Load_Linked;secondary_type RubyRequestType_LD;}} else if (pkt-req-isLockedRMW()) {//// x86 locked instructions are translated to store cache coherence// requests because these requests should always be treated as read// exclusive operations and should leverage any migratory sharing// optimization built into the protocol.//if (pkt-isWrite()) {DPRINTF(RubySequencer, Issuing Locked RMW Write\n);primary_type RubyRequestType_Locked_RMW_Write;} else {DPRINTF(RubySequencer, Issuing Locked RMW Read\n);assert(pkt-isRead());primary_type RubyRequestType_Locked_RMW_Read;}secondary_type RubyRequestType_ST;} else if (pkt-req-isTlbiCmd()) {primary_type secondary_type tlbiCmdToRubyRequestType(pkt);DPRINTF(RubySequencer, Issuing TLBI\n);} else {//// To support SwapReq, we need to check isWrite() first: a SwapReq// should always be treated like a write, but since a SwapReq implies// both isWrite() and isRead() are true, check isWrite() first here.//if (pkt-isWrite()) {//// Note: M5 packets do not differentiate ST from RMW_Write//primary_type secondary_type RubyRequestType_ST;} else if (pkt-isRead()) {// hardware transactional memory commandsif (pkt-req-isHTMCmd()) {primary_type secondary_type htmCmdToRubyRequestType(pkt);} else if (pkt-req-isInstFetch()) {primary_type secondary_type RubyRequestType_IFETCH;} else {if (pkt-req-isReadModifyWrite()) {primary_type RubyRequestType_RMW_Read;secondary_type RubyRequestType_ST;} else {primary_type secondary_type RubyRequestType_LD;}}} else if (pkt-isFlush()) {primary_type secondary_type RubyRequestType_FLUSH;} else {panic(Unsupported ruby packet type\n);}}// Check if the line is blocked for a Locked_RMWif (!pkt-req-isMemMgmt() m_controller-isBlocked(makeLineAddress(pkt-getAddr())) (primary_type ! RubyRequestType_Locked_RMW_Write)) {// Return that this requests cache line address aliases with// a prior request that locked the cache line. The request cannot// proceed until the cache line is unlocked by a Locked_RMW_Writereturn RequestStatus_Aliased;}RequestStatus status insertRequest(pkt, primary_type, secondary_type);// It is OK to receive RequestStatus_Aliased, it can be considered Issuedif (status ! RequestStatus_Ready status ! RequestStatus_Aliased)return status;// non-aliased with any existing request in the request table, just issue// to the cacheif (status ! RequestStatus_Aliased)issueRequest(pkt, secondary_type);// TODO: issue hardware prefetches herereturn RequestStatus_Issued;
}打印验证了是sequencer发出的makerequest.
核心代码是 insertRequest 把request放入requsttable 和issueRequest 发出一个msg
RequestStatus status insertRequest(pkt, primary_type, secondary_type);// Insert the request in the request table. Return RequestStatus_Aliased
// if the entry was already present.
RequestStatus
Sequencer::insertRequest(PacketPtr pkt, RubyRequestType primary_type,RubyRequestType secondary_type)
...
//下面是核心代码,把这个request插入到m_RequestTable里.
Addr line_addr makeLineAddress(pkt-getAddr());// Check if there is any outstanding request for the same cache line.auto seq_req_list m_RequestTable[line_addr];// Create a default entryseq_req_list.emplace_back(pkt, primary_type,secondary_type, curCycle());
...src/mem/ruby/system/Sequencer.cc issueRequest
void
Sequencer::issueRequest(PacketPtr pkt, RubyRequestType secondary_type)
{assert(pkt ! NULL);ContextID proc_id pkt-req-hasContextId() ?pkt-req-contextId() : InvalidContextID;ContextID core_id coreId();// If valid, copy the pc to the ruby requestAddr pc 0;if (pkt-req-hasPC()) {pc pkt-req-getPC();}// check if the packet has data as for example prefetch and flush// requests do notstd::shared_ptrRubyRequest msg;if (pkt-req-isMemMgmt()) {msg std::make_sharedRubyRequest(clockEdge(),pc, secondary_type,RubyAccessMode_Supervisor, pkt,proc_id, core_id);DPRINTFR(ProtocolTrace, %15s %3s %10s%20s %6s%-6s %s\n,curTick(), m_version, Seq, Begin, , ,RubyRequestType_to_string(secondary_type));if (pkt-req-isTlbiCmd()) {msg-m_isTlbi true;switch (secondary_type) {case RubyRequestType_TLBI_EXT_SYNC_COMP:msg-m_tlbiTransactionUid pkt-req-getExtraData();break;case RubyRequestType_TLBI:case RubyRequestType_TLBI_SYNC:msg-m_tlbiTransactionUid \getCurrentUnaddressedTransactionID();break;default:panic(Unexpected TLBI RubyRequestType);}DPRINTF(RubySequencer, Issuing TLBI %016x\n,msg-m_tlbiTransactionUid);}} else {msg std::make_sharedRubyRequest(clockEdge(), pkt-getAddr(),pkt-getSize(), pc, secondary_type,RubyAccessMode_Supervisor, pkt,PrefetchBit_No, proc_id, core_id);DPRINTFR(ProtocolTrace, %15s %3s %10s%20s %6s%-6s %#x %s\n,curTick(), m_version, Seq, Begin, , ,printAddress(msg-getPhysicalAddress()),RubyRequestType_to_string(secondary_type));}// hardware transactional memory// If the request originates in a transaction,// then mark the Ruby message as such.if (pkt-isHtmTransactional()) {msg-m_htmFromTransaction true;msg-m_htmTransactionUid pkt-getHtmTransactionUid();}Tick latency cyclesToTicks(m_controller-mandatoryQueueLatency(secondary_type));assert(latency 0);assert(m_mandatory_q_ptr ! NULL);m_mandatory_q_ptr-enqueue(msg, clockEdge(), latency);
}
issueRequst的关键是 m_mandatory_q_ptr-enqueue(msg, clockEdge(), latency);. m_mandatory_q_ptr 是在父类 src/mem/ruby/system/RubyPort.hh 中定义的 MessageBuffer* m_mandatory_q_ptr;
父类 src/mem/ruby/system/RubyPort.cc 中 RubyPort::init() m_mandatory_q_ptr m_controller-getMandatoryQueue();
就这样,自己的sequencer的request pkt,变成msg进入了rubyport 自己的 m_mandatory_q_ptr, 并且与m_controller-getMandatoryQueue()画上了等号.
因为我们查看 m_mandatory_q_ptr的操作很少,我们下面看怎么对msg操作的时候,需要看 getMandatoryQueue()
msg 如何从mandatoryq进入NetworkInterface暂定.
这两个代码也许是线索. src/mem/slicc/symbols/StateMachine.py 中
MessageBuffer*
$c_ident::getMandatoryQueue() const
{return $mq_ident;
}mq_ident NULLfor port in self.in_ports:if port.code.find(mandatoryQueue_ptr) 0:mq_ident m_mandatoryQueue_ptrNI将msg变成flit
核心是 if (flitisizeMessage(msg_ptr, vnet)) ,会把msg变成flit,然后在NoC了里传递.
void
NetworkInterface::wakeup()
{std::ostringstream oss;for (auto oPort: outPorts) {oss oPort-routerID() [ oPort-printVnets() ] ;}DPRINTF(RubyNetwork, Network Interface %d connected to router:%s woke up. Period: %ld\n, m_id, oss.str(), clockPeriod());std::coutcoutdebugyzzzz NetworkInterface::wakeup() m_id connected to router oss.str() clockPeriod()is clockPeriod() curTick()is curTick()std::endl;assert(curTick() clockEdge());MsgPtr msg_ptr;Tick curTime clockEdge();// Checking for messages coming from the protocol// can pick up a message/cycle for each virtual netfor (int vnet 0; vnet inNode_ptr.size(); vnet) {MessageBuffer *b inNode_ptr[vnet];if (b nullptr) {continue;}if (b-isReady(curTime)) { // Is there a message waitingmsg_ptr b-peekMsgPtr();std::coutcoutdebugyzzzzNI::wakeup()_msg_ptr msg_ptr.get() curTick()is curTick()std::endl;if (flitisizeMessage(msg_ptr, vnet)) {b-dequeue(curTime);}}}小结
这个博客总结了GEM5里,一个PYTHON文件如何生成pkt,这个pkt如何变成msg的. 以及一个msg如何变成flit的. msg如何从sequencer生成,到被Networkinterface操作有待下一篇完善细节…
下面别看,只是草稿
下面别看,只是草稿
下面别看,只是草稿
下面别看,只是草稿
附录
TimingRequestProtocol 和 TimingResponseProtocol的相应
RequestPort::sendTimingReq 方法尝试通过 TimingRequestProtocol 发送数据包并处理可能出现的异常。TimingRequestProtocol::sendReq 方法则负责确保请求的有效性并将请求转发给相应的响应协议TimingResponseProtocol进行处理。
流程和继承关系:
consumer.hh 定义了 virtual void wakeup() 0; src/mem/ruby/network/garnet/Router.hh 定义了 class Router : public BasicRouter, public Consumer继承了 父类 BasicRouter和 Consumer. src/mem/ruby/network/garnet/GarnetNetwork.cc (注意,不是.hh) 引用了router.hh #include “mem/ruby/network/garnet/Router.hh”.
consumer.hh
表明 wakeup 是一个必须在派生类中实现的接口函数。 0 语法: 这个部分将 wakeup 函数声明为纯虚拟pure virtual函数。在 C 中纯虚拟函数是一种特殊类型的虚拟函数它在基类中没有具体的实现并且要求任何非抽象的派生类必须提供该函数的实现。
flitize msg
分配vc
首先是要找空闲的vc,有一个封装好的函数会返回:
// Looking for a free output vc
int
NetworkInterface::calculateVC(int vnet)
{for (int i 0; i m_vc_per_vnet; i) {int delta m_vc_allocator[vnet];m_vc_allocator[vnet];if (m_vc_allocator[vnet] m_vc_per_vnet)m_vc_allocator[vnet] 0;if (outVcState[(vnet*m_vc_per_vnet) delta].isInState(IDLE_, curTick())) {vc_busy_counter[vnet] 0;return ((vnet*m_vc_per_vnet) delta);}}vc_busy_counter[vnet] 1;panic_if(vc_busy_counter[vnet] m_deadlock_threshold,%s: Possible network deadlock in vnet: %d at time: %llu \n,name(), vnet, curTick());return -1;
}下面是解读: 函数签名: int NetworkInterface::calculateVC(int vnet): 这个函数属于 NetworkInterface 类并返回一个整型值。它接受一个整型参数 vnet通常代表虚拟网络的标识。
遍历虚拟通道: for 循环遍历与给定虚拟网络 (vnet) 相关的所有虚拟通道。m_vc_per_vnet 是每个虚拟网络的虚拟通道数。 虚拟通道分配: 循环中的 delta 变量根据 m_vc_allocator[vnet] 的值设置表示当前虚拟通道的索引偏移。 m_vc_allocator[vnet] 更新虚拟通道分配器的值用于下一次调用此函数时选择不同的虚拟通道。 如果 m_vc_allocator[vnet] 达到 m_vc_per_vnet 的值它会重置为 0以循环方式遍历所有虚拟通道。 检查虚拟通道状态:
使用 outVcState[(vnet*m_vc_per_vnet) delta].isInState(IDLE_, curTick()) 检查当前虚拟通道是否处于空闲IDLE状态。如果是空闲状态函数返回该虚拟通道的索引。 虚拟通道忙碌计数器:
如果所有虚拟通道都不处于空闲状态vc_busy_counter[vnet] 加一表示此次调用没有找到空闲的虚拟通道。 如果 vc_busy_counter[vnet] 超过 m_deadlock_threshold 阈值函数会触发 panic意味着可能出现网络死锁并输出错误信息。 返回值:
如果找到空闲的虚拟通道则返回该通道的索引。 如果没有找到空闲的虚拟通道则返回 -1表示当前没有可用的虚拟通道。
