def construct(options, system, ruby_system): if (buildEnv['PROTOCOL'] != 'GPU_VIPER' or buildEnv['PROTOCOL'] != 'GPU_VIPER_Region' or buildEnv['PROTOCOL'] != 'GPU_VIPER_Baseline'): panic("This script requires VIPER based protocols \ to be built.") cpu_sequencers = [] cpuCluster = None cpuCluster = Cluster(name="CPU Cluster", extBW=8, intBW=8) # 16 GB/s for i in range((options.num_cpus + 1) // 2): cp_cntrl = CPCntrl() cp_cntrl.create(options, ruby_system, system) # Connect the CP controllers to the ruby network cp_cntrl.requestFromCore = ruby_system.network.slave cp_cntrl.responseFromCore = ruby_system.network.slave cp_cntrl.unblockFromCore = ruby_system.network.slave cp_cntrl.probeToCore = ruby_system.network.master cp_cntrl.responseToCore = ruby_system.network.master exec("system.cp_cntrl%d = cp_cntrl" % i) # # Add controllers and sequencers to the appropriate lists # cpu_sequencers.extend([cp_cntrl.sequencer, cp_cntrl.sequencer1]) cpuCluster.add(cp_cntrl) return cpu_sequencers, cpuCluster
def construct(options, system, ruby_system): if (buildEnv['PROTOCOL'] != 'GPU_VIPER' or buildEnv['PROTOCOL'] != 'GPU_VIPER_Region' or buildEnv['PROTOCOL'] != 'GPU_VIPER_Baseline'): panic("This script requires VIPER based protocols \ to be built.") cpu_sequencers = [] cpuCluster = None cpuCluster = Cluster(name="CPU Cluster", extBW = 8, intBW=8) # 16 GB/s for i in xrange((options.num_cpus + 1) / 2): cp_cntrl = CPCntrl() cp_cntrl.create(options, ruby_system, system) # Connect the CP controllers to the ruby network cp_cntrl.requestFromCore = ruby_system.network.slave cp_cntrl.responseFromCore = ruby_system.network.slave cp_cntrl.unblockFromCore = ruby_system.network.slave cp_cntrl.probeToCore = ruby_system.network.master cp_cntrl.responseToCore = ruby_system.network.master exec("system.cp_cntrl%d = cp_cntrl" % i) # # Add controllers and sequencers to the appropriate lists # cpu_sequencers.extend([cp_cntrl.sequencer, cp_cntrl.sequencer1]) cpuCluster.add(cp_cntrl) return cpu_sequencers, cpuCluster
def create_system(options, full_system, system, dma_devices, bootmem, ruby_system): if buildEnv['PROTOCOL'] != 'MOESI_AMD_Base': panic("This script requires the MOESI_AMD_Base protocol.") cpu_sequencers = [] # # The ruby network creation expects the list of nodes in the system to # be consistent with the NetDest list. Therefore the l1 controller # nodes must be listed before the directory nodes and directory nodes # before dma nodes, etc. # l1_cntrl_nodes = [] l3_cntrl_nodes = [] dir_cntrl_nodes = [] control_count = 0 # # Must create the individual controllers before the network to ensure # the controller constructors are called before the network constructor # # This is the base crossbar that connects the L3s, Dirs, and cpu # Cluster mainCluster = Cluster(extBW=512, intBW=512) # 1 TB/s if options.numa_high_bit: numa_bit = options.numa_high_bit else: # if the numa_bit is not specified, set the directory bits as the # lowest bits above the block offset bits, and the numa_bit as the # highest of those directory bits dir_bits = int(math.log(options.num_dirs, 2)) block_size_bits = int(math.log(options.cacheline_size, 2)) numa_bit = block_size_bits + dir_bits - 1 for i in range(options.num_dirs): dir_ranges = [] for r in system.mem_ranges: addr_range = m5.objects.AddrRange(r.start, size=r.size(), intlvHighBit=numa_bit, intlvBits=dir_bits, intlvMatch=i) dir_ranges.append(addr_range) dir_cntrl = DirCntrl(TCC_select_num_bits=0) dir_cntrl.create(options, dir_ranges, ruby_system, system) # Connect the Directory controller to the ruby network dir_cntrl.requestFromCores = MessageBuffer(ordered=True) dir_cntrl.requestFromCores.slave = ruby_system.network.master dir_cntrl.responseFromCores = MessageBuffer() dir_cntrl.responseFromCores.slave = ruby_system.network.master dir_cntrl.unblockFromCores = MessageBuffer() dir_cntrl.unblockFromCores.slave = ruby_system.network.master dir_cntrl.probeToCore = MessageBuffer() dir_cntrl.probeToCore.master = ruby_system.network.slave dir_cntrl.responseToCore = MessageBuffer() dir_cntrl.responseToCore.master = ruby_system.network.slave dir_cntrl.triggerQueue = MessageBuffer(ordered=True) dir_cntrl.L3triggerQueue = MessageBuffer(ordered=True) dir_cntrl.responseFromMemory = MessageBuffer() exec("system.dir_cntrl%d = dir_cntrl" % i) dir_cntrl_nodes.append(dir_cntrl) mainCluster.add(dir_cntrl) # Technically this config can support an odd number of cpus, but the top # level config files, such as the ruby_random_tester, will get confused if # the number of cpus does not equal the number of sequencers. Thus make # sure that an even number of cpus is specified. assert ((options.num_cpus % 2) == 0) # For an odd number of CPUs, still create the right number of controllers cpuCluster = Cluster(extBW=512, intBW=512) # 1 TB/s for i in range((options.num_cpus + 1) // 2): cp_cntrl = CPCntrl() cp_cntrl.create(options, ruby_system, system) exec("system.cp_cntrl%d = cp_cntrl" % i) # # Add controllers and sequencers to the appropriate lists # cpu_sequencers.extend([cp_cntrl.sequencer, cp_cntrl.sequencer1]) # Connect the CP controllers and the network cp_cntrl.requestFromCore = MessageBuffer() cp_cntrl.requestFromCore.master = ruby_system.network.slave cp_cntrl.responseFromCore = MessageBuffer() cp_cntrl.responseFromCore.master = ruby_system.network.slave cp_cntrl.unblockFromCore = MessageBuffer() cp_cntrl.unblockFromCore.master = ruby_system.network.slave cp_cntrl.probeToCore = MessageBuffer() cp_cntrl.probeToCore.slave = ruby_system.network.master cp_cntrl.responseToCore = MessageBuffer() cp_cntrl.responseToCore.slave = ruby_system.network.master cp_cntrl.mandatoryQueue = MessageBuffer() cp_cntrl.triggerQueue = MessageBuffer(ordered=True) cpuCluster.add(cp_cntrl) # Register CPUs and caches for each CorePair and directory (SE mode only) if not full_system: FileSystemConfig.config_filesystem(options) for i in xrange((options.num_cpus + 1) // 2): FileSystemConfig.register_cpu(physical_package_id=0, core_siblings=xrange( options.num_cpus), core_id=i * 2, thread_siblings=[]) FileSystemConfig.register_cpu(physical_package_id=0, core_siblings=xrange( options.num_cpus), core_id=i * 2 + 1, thread_siblings=[]) FileSystemConfig.register_cache(level=0, idu_type='Instruction', size=options.l1i_size, line_size=options.cacheline_size, assoc=options.l1i_assoc, cpus=[i * 2, i * 2 + 1]) FileSystemConfig.register_cache(level=0, idu_type='Data', size=options.l1d_size, line_size=options.cacheline_size, assoc=options.l1d_assoc, cpus=[i * 2]) FileSystemConfig.register_cache(level=0, idu_type='Data', size=options.l1d_size, line_size=options.cacheline_size, assoc=options.l1d_assoc, cpus=[i * 2 + 1]) FileSystemConfig.register_cache(level=1, idu_type='Unified', size=options.l2_size, line_size=options.cacheline_size, assoc=options.l2_assoc, cpus=[i * 2, i * 2 + 1]) for i in range(options.num_dirs): FileSystemConfig.register_cache( level=2, idu_type='Unified', size=options.l3_size, line_size=options.cacheline_size, assoc=options.l3_assoc, cpus=[n for n in xrange(options.num_cpus)]) # Assuming no DMA devices assert (len(dma_devices) == 0) # Add cpu/gpu clusters to main cluster mainCluster.add(cpuCluster) ruby_system.network.number_of_virtual_networks = 10 return (cpu_sequencers, dir_cntrl_nodes, mainCluster)
def create_system(options, full_system, system, dma_devices, bootmem, ruby_system): if buildEnv['PROTOCOL'] != 'GPU_VIPER': panic("This script requires the GPU_VIPER protocol to be built.") cpu_sequencers = [] # # The ruby network creation expects the list of nodes in the system to be # consistent with the NetDest list. Therefore the l1 controller nodes # must be listed before the directory nodes and directory nodes before # dma nodes, etc. # cp_cntrl_nodes = [] tcp_cntrl_nodes = [] sqc_cntrl_nodes = [] tcc_cntrl_nodes = [] dir_cntrl_nodes = [] l3_cntrl_nodes = [] # # Must create the individual controllers before the network to ensure the # controller constructors are called before the network constructor # # For an odd number of CPUs, still create the right number of controllers TCC_bits = int(math.log(options.num_tccs, 2)) # This is the base crossbar that connects the L3s, Dirs, and cpu/gpu # Clusters crossbar_bw = None mainCluster = None if options.numa_high_bit: numa_bit = options.numa_high_bit else: # if the numa_bit is not specified, set the directory bits as the # lowest bits above the block offset bits, and the numa_bit as the # highest of those directory bits dir_bits = int(math.log(options.num_dirs, 2)) block_size_bits = int(math.log(options.cacheline_size, 2)) numa_bit = block_size_bits + dir_bits - 1 if hasattr(options, 'bw_scalor') and options.bw_scalor > 0: #Assuming a 2GHz clock crossbar_bw = 16 * options.num_compute_units * options.bw_scalor mainCluster = Cluster(intBW=crossbar_bw) else: mainCluster = Cluster(intBW=8) # 16 GB/s for i in range(options.num_dirs): dir_ranges = [] for r in system.mem_ranges: addr_range = m5.objects.AddrRange(r.start, size=r.size(), intlvHighBit=numa_bit, intlvBits=dir_bits, intlvMatch=i) dir_ranges.append(addr_range) dir_cntrl = DirCntrl(noTCCdir=True, TCC_select_num_bits=TCC_bits) dir_cntrl.create(options, dir_ranges, ruby_system, system) dir_cntrl.number_of_TBEs = options.num_tbes dir_cntrl.useL3OnWT = options.use_L3_on_WT # the number_of_TBEs is inclusive of TBEs below # Connect the Directory controller to the ruby network dir_cntrl.requestFromCores = MessageBuffer(ordered=True) dir_cntrl.requestFromCores.slave = ruby_system.network.master dir_cntrl.responseFromCores = MessageBuffer() dir_cntrl.responseFromCores.slave = ruby_system.network.master dir_cntrl.unblockFromCores = MessageBuffer() dir_cntrl.unblockFromCores.slave = ruby_system.network.master dir_cntrl.probeToCore = MessageBuffer() dir_cntrl.probeToCore.master = ruby_system.network.slave dir_cntrl.responseToCore = MessageBuffer() dir_cntrl.responseToCore.master = ruby_system.network.slave dir_cntrl.triggerQueue = MessageBuffer(ordered=True) dir_cntrl.L3triggerQueue = MessageBuffer(ordered=True) dir_cntrl.requestToMemory = MessageBuffer() dir_cntrl.responseFromMemory = MessageBuffer() dir_cntrl.requestFromDMA = MessageBuffer(ordered=True) dir_cntrl.requestFromDMA.slave = ruby_system.network.master dir_cntrl.responseToDMA = MessageBuffer() dir_cntrl.responseToDMA.master = ruby_system.network.slave dir_cntrl.requestToMemory = MessageBuffer() dir_cntrl.responseFromMemory = MessageBuffer() exec("ruby_system.dir_cntrl%d = dir_cntrl" % i) dir_cntrl_nodes.append(dir_cntrl) mainCluster.add(dir_cntrl) cpuCluster = None if hasattr(options, 'bw_scalor') and options.bw_scalor > 0: cpuCluster = Cluster(extBW=crossbar_bw, intBW=crossbar_bw) else: cpuCluster = Cluster(extBW=8, intBW=8) # 16 GB/s for i in range((options.num_cpus + 1) // 2): cp_cntrl = CPCntrl() cp_cntrl.create(options, ruby_system, system) exec("ruby_system.cp_cntrl%d = cp_cntrl" % i) # # Add controllers and sequencers to the appropriate lists # cpu_sequencers.extend([cp_cntrl.sequencer, cp_cntrl.sequencer1]) # Connect the CP controllers and the network cp_cntrl.requestFromCore = MessageBuffer() cp_cntrl.requestFromCore.master = ruby_system.network.slave cp_cntrl.responseFromCore = MessageBuffer() cp_cntrl.responseFromCore.master = ruby_system.network.slave cp_cntrl.unblockFromCore = MessageBuffer() cp_cntrl.unblockFromCore.master = ruby_system.network.slave cp_cntrl.probeToCore = MessageBuffer() cp_cntrl.probeToCore.slave = ruby_system.network.master cp_cntrl.responseToCore = MessageBuffer() cp_cntrl.responseToCore.slave = ruby_system.network.master cp_cntrl.mandatoryQueue = MessageBuffer() cp_cntrl.triggerQueue = MessageBuffer(ordered=True) cpuCluster.add(cp_cntrl) # Register CPUs and caches for each CorePair and directory (SE mode only) if not full_system: for i in range((options.num_cpus + 1) // 2): FileSystemConfig.register_cpu(physical_package_id = 0, core_siblings = \ range(options.num_cpus), core_id = i*2, thread_siblings = []) FileSystemConfig.register_cpu(physical_package_id = 0, core_siblings = \ range(options.num_cpus), core_id = i*2+1, thread_siblings = []) FileSystemConfig.register_cache(level=0, idu_type='Instruction', size=options.l1i_size, line_size=options.cacheline_size, assoc=options.l1i_assoc, cpus=[i * 2, i * 2 + 1]) FileSystemConfig.register_cache(level=0, idu_type='Data', size=options.l1d_size, line_size=options.cacheline_size, assoc=options.l1d_assoc, cpus=[i * 2]) FileSystemConfig.register_cache(level=0, idu_type='Data', size=options.l1d_size, line_size=options.cacheline_size, assoc=options.l1d_assoc, cpus=[i * 2 + 1]) FileSystemConfig.register_cache(level=1, idu_type='Unified', size=options.l2_size, line_size=options.cacheline_size, assoc=options.l2_assoc, cpus=[i * 2, i * 2 + 1]) for i in range(options.num_dirs): FileSystemConfig.register_cache( level=2, idu_type='Unified', size=options.l3_size, line_size=options.cacheline_size, assoc=options.l3_assoc, cpus=[n for n in range(options.num_cpus)]) gpuCluster = None if hasattr(options, 'bw_scalor') and options.bw_scalor > 0: gpuCluster = Cluster(extBW=crossbar_bw, intBW=crossbar_bw) else: gpuCluster = Cluster(extBW=8, intBW=8) # 16 GB/s for i in range(options.num_compute_units): tcp_cntrl = TCPCntrl(TCC_select_num_bits=TCC_bits, issue_latency=1, number_of_TBEs=2560) # TBEs set to max outstanding requests tcp_cntrl.create(options, ruby_system, system) tcp_cntrl.WB = options.WB_L1 tcp_cntrl.disableL1 = options.noL1 tcp_cntrl.L1cache.tagAccessLatency = options.TCP_latency tcp_cntrl.L1cache.dataAccessLatency = options.TCP_latency exec("ruby_system.tcp_cntrl%d = tcp_cntrl" % i) # # Add controllers and sequencers to the appropriate lists # cpu_sequencers.append(tcp_cntrl.coalescer) tcp_cntrl_nodes.append(tcp_cntrl) # Connect the TCP controller to the ruby network tcp_cntrl.requestFromTCP = MessageBuffer(ordered=True) tcp_cntrl.requestFromTCP.master = ruby_system.network.slave tcp_cntrl.responseFromTCP = MessageBuffer(ordered=True) tcp_cntrl.responseFromTCP.master = ruby_system.network.slave tcp_cntrl.unblockFromCore = MessageBuffer() tcp_cntrl.unblockFromCore.master = ruby_system.network.slave tcp_cntrl.probeToTCP = MessageBuffer(ordered=True) tcp_cntrl.probeToTCP.slave = ruby_system.network.master tcp_cntrl.responseToTCP = MessageBuffer(ordered=True) tcp_cntrl.responseToTCP.slave = ruby_system.network.master tcp_cntrl.mandatoryQueue = MessageBuffer() gpuCluster.add(tcp_cntrl) for i in range(options.num_sqc): sqc_cntrl = SQCCntrl(TCC_select_num_bits=TCC_bits) sqc_cntrl.create(options, ruby_system, system) exec("ruby_system.sqc_cntrl%d = sqc_cntrl" % i) # # Add controllers and sequencers to the appropriate lists # cpu_sequencers.append(sqc_cntrl.sequencer) # Connect the SQC controller to the ruby network sqc_cntrl.requestFromSQC = MessageBuffer(ordered=True) sqc_cntrl.requestFromSQC.master = ruby_system.network.slave sqc_cntrl.probeToSQC = MessageBuffer(ordered=True) sqc_cntrl.probeToSQC.slave = ruby_system.network.master sqc_cntrl.responseToSQC = MessageBuffer(ordered=True) sqc_cntrl.responseToSQC.slave = ruby_system.network.master sqc_cntrl.mandatoryQueue = MessageBuffer() # SQC also in GPU cluster gpuCluster.add(sqc_cntrl) for i in xrange(options.num_scalar_cache): scalar_cntrl = SQCCntrl(TCC_select_num_bits=TCC_bits) scalar_cntrl.create(options, ruby_system, system) exec('ruby_system.scalar_cntrl%d = scalar_cntrl' % i) cpu_sequencers.append(scalar_cntrl.sequencer) scalar_cntrl.requestFromSQC = MessageBuffer(ordered=True) scalar_cntrl.requestFromSQC.master = ruby_system.network.slave scalar_cntrl.probeToSQC = MessageBuffer(ordered=True) scalar_cntrl.probeToSQC.slave = ruby_system.network.master scalar_cntrl.responseToSQC = MessageBuffer(ordered=True) scalar_cntrl.responseToSQC.slave = ruby_system.network.master scalar_cntrl.mandatoryQueue = \ MessageBuffer(buffer_size=options.scalar_buffer_size) gpuCluster.add(scalar_cntrl) for i in xrange(options.num_cp): tcp_ID = options.num_compute_units + i sqc_ID = options.num_sqc + i tcp_cntrl = TCPCntrl(TCC_select_num_bits=TCC_bits, issue_latency=1, number_of_TBEs=2560) # TBEs set to max outstanding requests tcp_cntrl.createCP(options, ruby_system, system) tcp_cntrl.WB = options.WB_L1 tcp_cntrl.disableL1 = options.noL1 tcp_cntrl.L1cache.tagAccessLatency = options.TCP_latency tcp_cntrl.L1cache.dataAccessLatency = options.TCP_latency exec("ruby_system.tcp_cntrl%d = tcp_cntrl" % tcp_ID) # # Add controllers and sequencers to the appropriate lists # cpu_sequencers.append(tcp_cntrl.sequencer) tcp_cntrl_nodes.append(tcp_cntrl) # Connect the CP (TCP) controllers to the ruby network tcp_cntrl.requestFromTCP = MessageBuffer(ordered=True) tcp_cntrl.requestFromTCP.master = ruby_system.network.slave tcp_cntrl.responseFromTCP = MessageBuffer(ordered=True) tcp_cntrl.responseFromTCP.master = ruby_system.network.slave tcp_cntrl.unblockFromCore = MessageBuffer(ordered=True) tcp_cntrl.unblockFromCore.master = ruby_system.network.slave tcp_cntrl.probeToTCP = MessageBuffer(ordered=True) tcp_cntrl.probeToTCP.slave = ruby_system.network.master tcp_cntrl.responseToTCP = MessageBuffer(ordered=True) tcp_cntrl.responseToTCP.slave = ruby_system.network.master tcp_cntrl.mandatoryQueue = MessageBuffer() gpuCluster.add(tcp_cntrl) sqc_cntrl = SQCCntrl(TCC_select_num_bits=TCC_bits) sqc_cntrl.create(options, ruby_system, system) exec("ruby_system.sqc_cntrl%d = sqc_cntrl" % sqc_ID) # # Add controllers and sequencers to the appropriate lists # cpu_sequencers.append(sqc_cntrl.sequencer) # SQC also in GPU cluster gpuCluster.add(sqc_cntrl) for i in range(options.num_tccs): tcc_cntrl = TCCCntrl(l2_response_latency=options.TCC_latency) tcc_cntrl.create(options, ruby_system, system) tcc_cntrl.l2_request_latency = options.gpu_to_dir_latency tcc_cntrl.l2_response_latency = options.TCC_latency tcc_cntrl_nodes.append(tcc_cntrl) tcc_cntrl.WB = options.WB_L2 tcc_cntrl.number_of_TBEs = 2560 * options.num_compute_units # the number_of_TBEs is inclusive of TBEs below # Connect the TCC controllers to the ruby network tcc_cntrl.requestFromTCP = MessageBuffer(ordered=True) tcc_cntrl.requestFromTCP.slave = ruby_system.network.master tcc_cntrl.responseToCore = MessageBuffer(ordered=True) tcc_cntrl.responseToCore.master = ruby_system.network.slave tcc_cntrl.probeFromNB = MessageBuffer() tcc_cntrl.probeFromNB.slave = ruby_system.network.master tcc_cntrl.responseFromNB = MessageBuffer() tcc_cntrl.responseFromNB.slave = ruby_system.network.master tcc_cntrl.requestToNB = MessageBuffer(ordered=True) tcc_cntrl.requestToNB.master = ruby_system.network.slave tcc_cntrl.responseToNB = MessageBuffer() tcc_cntrl.responseToNB.master = ruby_system.network.slave tcc_cntrl.unblockToNB = MessageBuffer() tcc_cntrl.unblockToNB.master = ruby_system.network.slave tcc_cntrl.triggerQueue = MessageBuffer(ordered=True) exec("ruby_system.tcc_cntrl%d = tcc_cntrl" % i) # connect all of the wire buffers between L3 and dirs up # TCC cntrls added to the GPU cluster gpuCluster.add(tcc_cntrl) for i, dma_device in enumerate(dma_devices): dma_seq = DMASequencer(version=i, ruby_system=ruby_system) dma_cntrl = DMA_Controller(version=i, dma_sequencer=dma_seq, ruby_system=ruby_system) exec('system.dma_cntrl%d = dma_cntrl' % i) if dma_device.type == 'MemTest': exec('system.dma_cntrl%d.dma_sequencer.slave = dma_devices.test' % i) else: exec('system.dma_cntrl%d.dma_sequencer.slave = dma_device.dma' % i) dma_cntrl.requestToDir = MessageBuffer(buffer_size=0) dma_cntrl.requestToDir.master = ruby_system.network.slave dma_cntrl.responseFromDir = MessageBuffer(buffer_size=0) dma_cntrl.responseFromDir.slave = ruby_system.network.master dma_cntrl.mandatoryQueue = MessageBuffer(buffer_size=0) gpuCluster.add(dma_cntrl) # Add cpu/gpu clusters to main cluster mainCluster.add(cpuCluster) mainCluster.add(gpuCluster) ruby_system.network.number_of_virtual_networks = 11 return (cpu_sequencers, dir_cntrl_nodes, mainCluster)
def create_system(options, full_system, system, dma_devices, ruby_system): if buildEnv['PROTOCOL'] != 'GPU_VIPER_Baseline': panic("This script requires the" \ "GPU_VIPER_Baseline protocol to be built.") cpu_sequencers = [] # # The ruby network creation expects the list of nodes in the system to be # consistent with the NetDest list. Therefore the l1 controller nodes # must be listed before the directory nodes and directory nodes before # dma nodes, etc. # cp_cntrl_nodes = [] tcp_cntrl_nodes = [] sqc_cntrl_nodes = [] tcc_cntrl_nodes = [] dir_cntrl_nodes = [] l3_cntrl_nodes = [] # # Must create the individual controllers before the network to ensure the # controller constructors are called before the network constructor # # For an odd number of CPUs, still create the right number of controllers TCC_bits = int(math.log(options.num_tccs, 2)) # This is the base crossbar that connects the L3s, Dirs, and cpu/gpu # Clusters crossbar_bw = 16 * options.num_compute_units #Assuming a 2GHz clock mainCluster = Cluster(intBW=crossbar_bw) for i in xrange(options.num_dirs): dir_cntrl = DirCntrl(noTCCdir=True, TCC_select_num_bits=TCC_bits) dir_cntrl.create(options, ruby_system, system) dir_cntrl.number_of_TBEs = options.num_tbes dir_cntrl.useL3OnWT = options.use_L3_on_WT dir_cntrl.inclusiveDir = not options.nonInclusiveDir # Connect the Directory controller to the ruby network dir_cntrl.requestFromCores = MessageBuffer(ordered=True) dir_cntrl.requestFromCores.slave = ruby_system.network.master dir_cntrl.responseFromCores = MessageBuffer() dir_cntrl.responseFromCores.slave = ruby_system.network.master dir_cntrl.unblockFromCores = MessageBuffer() dir_cntrl.unblockFromCores.slave = ruby_system.network.master dir_cntrl.probeToCore = MessageBuffer() dir_cntrl.probeToCore.master = ruby_system.network.slave dir_cntrl.responseToCore = MessageBuffer() dir_cntrl.responseToCore.master = ruby_system.network.slave dir_cntrl.triggerQueue = MessageBuffer(ordered=True) dir_cntrl.L3triggerQueue = MessageBuffer(ordered=True) dir_cntrl.responseFromMemory = MessageBuffer() exec("system.dir_cntrl%d = dir_cntrl" % i) dir_cntrl_nodes.append(dir_cntrl) mainCluster.add(dir_cntrl) cpuCluster = Cluster(extBW=crossbar_bw, intBW=crossbar_bw) for i in xrange((options.num_cpus + 1) / 2): cp_cntrl = CPCntrl() cp_cntrl.create(options, ruby_system, system) exec("system.cp_cntrl%d = cp_cntrl" % i) # # Add controllers and sequencers to the appropriate lists # cpu_sequencers.extend([cp_cntrl.sequencer, cp_cntrl.sequencer1]) # Connect the CP controllers and the network cp_cntrl.requestFromCore = MessageBuffer() cp_cntrl.requestFromCore.master = ruby_system.network.slave cp_cntrl.responseFromCore = MessageBuffer() cp_cntrl.responseFromCore.master = ruby_system.network.slave cp_cntrl.unblockFromCore = MessageBuffer() cp_cntrl.unblockFromCore.master = ruby_system.network.slave cp_cntrl.probeToCore = MessageBuffer() cp_cntrl.probeToCore.slave = ruby_system.network.master cp_cntrl.responseToCore = MessageBuffer() cp_cntrl.responseToCore.slave = ruby_system.network.master cp_cntrl.mandatoryQueue = MessageBuffer() cp_cntrl.triggerQueue = MessageBuffer(ordered=True) cpuCluster.add(cp_cntrl) gpuCluster = Cluster(extBW=crossbar_bw, intBW=crossbar_bw) for i in xrange(options.num_compute_units): tcp_cntrl = TCPCntrl(TCC_select_num_bits=TCC_bits, issue_latency=1, number_of_TBEs=2560) # TBEs set to max outstanding requests tcp_cntrl.create(options, ruby_system, system) tcp_cntrl.WB = options.WB_L1 tcp_cntrl.disableL1 = options.noL1 exec("system.tcp_cntrl%d = tcp_cntrl" % i) # # Add controllers and sequencers to the appropriate lists # cpu_sequencers.append(tcp_cntrl.coalescer) tcp_cntrl_nodes.append(tcp_cntrl) # Connect the CP (TCP) controllers to the ruby network tcp_cntrl.requestFromTCP = MessageBuffer(ordered=True) tcp_cntrl.requestFromTCP.master = ruby_system.network.slave tcp_cntrl.responseFromTCP = MessageBuffer(ordered=True) tcp_cntrl.responseFromTCP.master = ruby_system.network.slave tcp_cntrl.unblockFromCore = MessageBuffer() tcp_cntrl.unblockFromCore.master = ruby_system.network.slave tcp_cntrl.probeToTCP = MessageBuffer(ordered=True) tcp_cntrl.probeToTCP.slave = ruby_system.network.master tcp_cntrl.responseToTCP = MessageBuffer(ordered=True) tcp_cntrl.responseToTCP.slave = ruby_system.network.master tcp_cntrl.mandatoryQueue = MessageBuffer() gpuCluster.add(tcp_cntrl) for i in xrange(options.num_sqc): sqc_cntrl = SQCCntrl(TCC_select_num_bits=TCC_bits) sqc_cntrl.create(options, ruby_system, system) exec("system.sqc_cntrl%d = sqc_cntrl" % i) # # Add controllers and sequencers to the appropriate lists # cpu_sequencers.append(sqc_cntrl.sequencer) # Connect the SQC controller to the ruby network sqc_cntrl.requestFromSQC = MessageBuffer(ordered=True) sqc_cntrl.requestFromSQC.master = ruby_system.network.slave sqc_cntrl.probeToSQC = MessageBuffer(ordered=True) sqc_cntrl.probeToSQC.slave = ruby_system.network.master sqc_cntrl.responseToSQC = MessageBuffer(ordered=True) sqc_cntrl.responseToSQC.slave = ruby_system.network.master sqc_cntrl.mandatoryQueue = MessageBuffer() # SQC also in GPU cluster gpuCluster.add(sqc_cntrl) # Because of wire buffers, num_tccs must equal num_tccdirs numa_bit = 6 for i in xrange(options.num_tccs): tcc_cntrl = TCCCntrl() tcc_cntrl.create(options, ruby_system, system) tcc_cntrl.l2_request_latency = options.gpu_to_dir_latency tcc_cntrl.l2_response_latency = options.TCC_latency tcc_cntrl_nodes.append(tcc_cntrl) tcc_cntrl.WB = options.WB_L2 tcc_cntrl.number_of_TBEs = 2560 * options.num_compute_units # Connect the TCC controllers to the ruby network tcc_cntrl.requestFromTCP = MessageBuffer(ordered=True) tcc_cntrl.requestFromTCP.slave = ruby_system.network.master tcc_cntrl.responseToCore = MessageBuffer(ordered=True) tcc_cntrl.responseToCore.master = ruby_system.network.slave tcc_cntrl.probeFromNB = MessageBuffer() tcc_cntrl.probeFromNB.slave = ruby_system.network.master tcc_cntrl.responseFromNB = MessageBuffer() tcc_cntrl.responseFromNB.slave = ruby_system.network.master tcc_cntrl.requestToNB = MessageBuffer(ordered=True) tcc_cntrl.requestToNB.master = ruby_system.network.slave tcc_cntrl.responseToNB = MessageBuffer() tcc_cntrl.responseToNB.master = ruby_system.network.slave tcc_cntrl.unblockToNB = MessageBuffer() tcc_cntrl.unblockToNB.master = ruby_system.network.slave tcc_cntrl.triggerQueue = MessageBuffer(ordered=True) exec("system.tcc_cntrl%d = tcc_cntrl" % i) # connect all of the wire buffers between L3 and dirs up # TCC cntrls added to the GPU cluster gpuCluster.add(tcc_cntrl) # Assuming no DMA devices assert (len(dma_devices) == 0) # Add cpu/gpu clusters to main cluster mainCluster.add(cpuCluster) mainCluster.add(gpuCluster) ruby_system.network.number_of_virtual_networks = 10 return (cpu_sequencers, dir_cntrl_nodes, mainCluster)
def create_system(options, full_system, system, dma_devices, bootmem, ruby_system): if buildEnv['PROTOCOL'] != 'GPU_RfO': panic("This script requires the GPU_RfO protocol to be built.") cpu_sequencers = [] # # The ruby network creation expects the list of nodes in the system to be # consistent with the NetDest list. Therefore the l1 controller nodes # must be listed before the directory nodes and directory nodes before # dma nodes, etc. # cp_cntrl_nodes = [] tcp_cntrl_nodes = [] sqc_cntrl_nodes = [] tcc_cntrl_nodes = [] tccdir_cntrl_nodes = [] dir_cntrl_nodes = [] l3_cntrl_nodes = [] # # Must create the individual controllers before the network to ensure the # controller constructors are called before the network constructor # TCC_bits = int(math.log(options.num_tccs, 2)) # This is the base crossbar that connects the L3s, Dirs, and cpu/gpu # Clusters mainCluster = Cluster(extBW=512, intBW=512) # 1 TB/s if options.numa_high_bit: numa_bit = options.numa_high_bit else: # if the numa_bit is not specified, set the directory bits as the # lowest bits above the block offset bits, and the numa_bit as the # highest of those directory bits dir_bits = int(math.log(options.num_dirs, 2)) block_size_bits = int(math.log(options.cacheline_size, 2)) numa_bit = block_size_bits + dir_bits - 1 for i in range(options.num_dirs): dir_ranges = [] for r in system.mem_ranges: addr_range = m5.objects.AddrRange(r.start, size=r.size(), intlvHighBit=numa_bit, intlvBits=dir_bits, intlvMatch=i) dir_ranges.append(addr_range) dir_cntrl = DirCntrl(TCC_select_num_bits=TCC_bits) dir_cntrl.create(options, dir_ranges, ruby_system, system) dir_cntrl.number_of_TBEs = 2560 * options.num_compute_units #Enough TBEs for all TCP TBEs # Connect the Directory controller to the ruby network dir_cntrl.requestFromCores = MessageBuffer(ordered=True) dir_cntrl.requestFromCores.slave = ruby_system.network.master dir_cntrl.responseFromCores = MessageBuffer() dir_cntrl.responseFromCores.slave = ruby_system.network.master dir_cntrl.unblockFromCores = MessageBuffer() dir_cntrl.unblockFromCores.slave = ruby_system.network.master dir_cntrl.probeToCore = MessageBuffer() dir_cntrl.probeToCore.master = ruby_system.network.slave dir_cntrl.responseToCore = MessageBuffer() dir_cntrl.responseToCore.master = ruby_system.network.slave dir_cntrl.triggerQueue = MessageBuffer(ordered=True) dir_cntrl.L3triggerQueue = MessageBuffer(ordered=True) dir_cntrl.responseFromMemory = MessageBuffer() exec("system.dir_cntrl%d = dir_cntrl" % i) dir_cntrl_nodes.append(dir_cntrl) mainCluster.add(dir_cntrl) # For an odd number of CPUs, still create the right number of controllers cpuCluster = Cluster(extBW=512, intBW=512) # 1 TB/s for i in range((options.num_cpus + 1) // 2): cp_cntrl = CPCntrl() cp_cntrl.create(options, ruby_system, system) exec("system.cp_cntrl%d = cp_cntrl" % i) # # Add controllers and sequencers to the appropriate lists # cpu_sequencers.extend([cp_cntrl.sequencer, cp_cntrl.sequencer1]) # Connect the CP controllers and the network cp_cntrl.requestFromCore = MessageBuffer() cp_cntrl.requestFromCore.master = ruby_system.network.slave cp_cntrl.responseFromCore = MessageBuffer() cp_cntrl.responseFromCore.master = ruby_system.network.slave cp_cntrl.unblockFromCore = MessageBuffer() cp_cntrl.unblockFromCore.master = ruby_system.network.slave cp_cntrl.probeToCore = MessageBuffer() cp_cntrl.probeToCore.slave = ruby_system.network.master cp_cntrl.responseToCore = MessageBuffer() cp_cntrl.responseToCore.slave = ruby_system.network.master cp_cntrl.mandatoryQueue = MessageBuffer() cp_cntrl.triggerQueue = MessageBuffer(ordered=True) cpuCluster.add(cp_cntrl) gpuCluster = Cluster(extBW=512, intBW=512) # 1 TB/s for i in range(options.num_compute_units): tcp_cntrl = TCPCntrl(TCC_select_num_bits=TCC_bits, number_of_TBEs=2560) # max outstanding requests tcp_cntrl.create(options, ruby_system, system) exec("system.tcp_cntrl%d = tcp_cntrl" % i) # # Add controllers and sequencers to the appropriate lists # cpu_sequencers.append(tcp_cntrl.coalescer) tcp_cntrl_nodes.append(tcp_cntrl) # Connect the TCP controller to the ruby network tcp_cntrl.requestFromTCP = MessageBuffer(ordered=True) tcp_cntrl.requestFromTCP.master = ruby_system.network.slave tcp_cntrl.responseFromTCP = MessageBuffer(ordered=True) tcp_cntrl.responseFromTCP.master = ruby_system.network.slave tcp_cntrl.unblockFromCore = MessageBuffer(ordered=True) tcp_cntrl.unblockFromCore.master = ruby_system.network.slave tcp_cntrl.probeToTCP = MessageBuffer(ordered=True) tcp_cntrl.probeToTCP.slave = ruby_system.network.master tcp_cntrl.responseToTCP = MessageBuffer(ordered=True) tcp_cntrl.responseToTCP.slave = ruby_system.network.master tcp_cntrl.mandatoryQueue = MessageBuffer() gpuCluster.add(tcp_cntrl) for i in range(options.num_sqc): sqc_cntrl = SQCCntrl(TCC_select_num_bits=TCC_bits) sqc_cntrl.create(options, ruby_system, system) exec("system.sqc_cntrl%d = sqc_cntrl" % i) # # Add controllers and sequencers to the appropriate lists # cpu_sequencers.append(sqc_cntrl.sequencer) # Connect the SQC controller to the ruby network sqc_cntrl.requestFromSQC = MessageBuffer(ordered=True) sqc_cntrl.requestFromSQC.master = ruby_system.network.slave sqc_cntrl.responseFromSQC = MessageBuffer(ordered=True) sqc_cntrl.responseFromSQC.master = ruby_system.network.slave sqc_cntrl.unblockFromCore = MessageBuffer(ordered=True) sqc_cntrl.unblockFromCore.master = ruby_system.network.slave sqc_cntrl.probeToSQC = MessageBuffer(ordered=True) sqc_cntrl.probeToSQC.slave = ruby_system.network.master sqc_cntrl.responseToSQC = MessageBuffer(ordered=True) sqc_cntrl.responseToSQC.slave = ruby_system.network.master sqc_cntrl.mandatoryQueue = MessageBuffer() # SQC also in GPU cluster gpuCluster.add(sqc_cntrl) for i in range(options.num_cp): tcp_cntrl = TCPCntrl(TCC_select_num_bits=TCC_bits, number_of_TBEs=2560) # max outstanding requests tcp_cntrl.createCP(options, ruby_system, system) exec("system.tcp_cntrl%d = tcp_cntrl" % (options.num_compute_units + i)) # # Add controllers and sequencers to the appropriate lists # cpu_sequencers.append(tcp_cntrl.sequencer) tcp_cntrl_nodes.append(tcp_cntrl) # Connect the TCP controller to the ruby network tcp_cntrl.requestFromTCP = MessageBuffer(ordered=True) tcp_cntrl.requestFromTCP.master = ruby_system.network.slave tcp_cntrl.responseFromTCP = MessageBuffer(ordered=True) tcp_cntrl.responseFromTCP.master = ruby_system.network.slave tcp_cntrl.unblockFromCore = MessageBuffer(ordered=True) tcp_cntrl.unblockFromCore.master = ruby_system.network.slave tcp_cntrl.probeToTCP = MessageBuffer(ordered=True) tcp_cntrl.probeToTCP.slave = ruby_system.network.master tcp_cntrl.responseToTCP = MessageBuffer(ordered=True) tcp_cntrl.responseToTCP.slave = ruby_system.network.master tcp_cntrl.mandatoryQueue = MessageBuffer() gpuCluster.add(tcp_cntrl) sqc_cntrl = SQCCntrl(TCC_select_num_bits=TCC_bits) sqc_cntrl.createCP(options, ruby_system, system) exec("system.sqc_cntrl%d = sqc_cntrl" % (options.num_compute_units + i)) # # Add controllers and sequencers to the appropriate lists # cpu_sequencers.append(sqc_cntrl.sequencer) # Connect the SQC controller to the ruby network sqc_cntrl.requestFromSQC = MessageBuffer(ordered=True) sqc_cntrl.requestFromSQC.master = ruby_system.network.slave sqc_cntrl.responseFromSQC = MessageBuffer(ordered=True) sqc_cntrl.responseFromSQC.master = ruby_system.network.slave sqc_cntrl.unblockFromCore = MessageBuffer(ordered=True) sqc_cntrl.unblockFromCore.master = ruby_system.network.slave sqc_cntrl.probeToSQC = MessageBuffer(ordered=True) sqc_cntrl.probeToSQC.slave = ruby_system.network.master sqc_cntrl.responseToSQC = MessageBuffer(ordered=True) sqc_cntrl.responseToSQC.slave = ruby_system.network.master sqc_cntrl.mandatoryQueue = MessageBuffer() # SQC also in GPU cluster gpuCluster.add(sqc_cntrl) for i in range(options.num_tccs): tcc_cntrl = TCCCntrl(TCC_select_num_bits=TCC_bits, number_of_TBEs=options.num_compute_units * 2560) #Enough TBEs for all TCP TBEs tcc_cntrl.create(options, ruby_system, system) tcc_cntrl_nodes.append(tcc_cntrl) tccdir_cntrl = TCCDirCntrl(TCC_select_num_bits=TCC_bits, number_of_TBEs=options.num_compute_units * 2560) #Enough TBEs for all TCP TBEs tccdir_cntrl.create(options, ruby_system, system) tccdir_cntrl_nodes.append(tccdir_cntrl) exec("system.tcc_cntrl%d = tcc_cntrl" % i) exec("system.tccdir_cntrl%d = tccdir_cntrl" % i) # connect all of the wire buffers between L3 and dirs up req_to_tccdir = RubyWireBuffer() resp_to_tccdir = RubyWireBuffer() tcc_unblock_to_tccdir = RubyWireBuffer() req_to_tcc = RubyWireBuffer() probe_to_tcc = RubyWireBuffer() resp_to_tcc = RubyWireBuffer() tcc_cntrl.connectWireBuffers(req_to_tccdir, resp_to_tccdir, tcc_unblock_to_tccdir, req_to_tcc, probe_to_tcc, resp_to_tcc) tccdir_cntrl.connectWireBuffers(req_to_tccdir, resp_to_tccdir, tcc_unblock_to_tccdir, req_to_tcc, probe_to_tcc, resp_to_tcc) # Connect the TCC controller to the ruby network tcc_cntrl.responseFromTCC = MessageBuffer(ordered=True) tcc_cntrl.responseFromTCC.master = ruby_system.network.slave tcc_cntrl.responseToTCC = MessageBuffer(ordered=True) tcc_cntrl.responseToTCC.slave = ruby_system.network.master # Connect the TCC Dir controller to the ruby network tccdir_cntrl.requestFromTCP = MessageBuffer(ordered=True) tccdir_cntrl.requestFromTCP.slave = ruby_system.network.master tccdir_cntrl.responseFromTCP = MessageBuffer(ordered=True) tccdir_cntrl.responseFromTCP.slave = ruby_system.network.master tccdir_cntrl.unblockFromTCP = MessageBuffer(ordered=True) tccdir_cntrl.unblockFromTCP.slave = ruby_system.network.master tccdir_cntrl.probeToCore = MessageBuffer(ordered=True) tccdir_cntrl.probeToCore.master = ruby_system.network.slave tccdir_cntrl.responseToCore = MessageBuffer(ordered=True) tccdir_cntrl.responseToCore.master = ruby_system.network.slave tccdir_cntrl.probeFromNB = MessageBuffer() tccdir_cntrl.probeFromNB.slave = ruby_system.network.master tccdir_cntrl.responseFromNB = MessageBuffer() tccdir_cntrl.responseFromNB.slave = ruby_system.network.master tccdir_cntrl.requestToNB = MessageBuffer() tccdir_cntrl.requestToNB.master = ruby_system.network.slave tccdir_cntrl.responseToNB = MessageBuffer() tccdir_cntrl.responseToNB.master = ruby_system.network.slave tccdir_cntrl.unblockToNB = MessageBuffer() tccdir_cntrl.unblockToNB.master = ruby_system.network.slave tccdir_cntrl.triggerQueue = MessageBuffer(ordered=True) # TCC cntrls added to the GPU cluster gpuCluster.add(tcc_cntrl) gpuCluster.add(tccdir_cntrl) # Assuming no DMA devices assert (len(dma_devices) == 0) # Add cpu/gpu clusters to main cluster mainCluster.add(cpuCluster) mainCluster.add(gpuCluster) ruby_system.network.number_of_virtual_networks = 10 return (cpu_sequencers, dir_cntrl_nodes, mainCluster)
def create_system(options, full_system, system, dma_devices, bootmem, ruby_system): if buildEnv['PROTOCOL'] != 'GPU_VIPER': panic("This script requires the GPU_VIPER protocol to be built.") cpu_sequencers = [] # # The ruby network creation expects the list of nodes in the system to be # consistent with the NetDest list. Therefore the l1 controller nodes # must be listed before the directory nodes and directory nodes before # dma nodes, etc. # cp_cntrl_nodes = [] tcp_cntrl_nodes = [] sqc_cntrl_nodes = [] tcc_cntrl_nodes = [] dir_cntrl_nodes = [] l3_cntrl_nodes = [] # # Must create the individual controllers before the network to ensure the # controller constructors are called before the network constructor # # For an odd number of CPUs, still create the right number of controllers TCC_bits = int(math.log(options.num_tccs, 2)) # This is the base crossbar that connects the L3s, Dirs, and cpu/gpu # Clusters crossbar_bw = None mainCluster = None if hasattr(options, 'bw_scalor') and options.bw_scalor > 0: #Assuming a 2GHz clock crossbar_bw = 16 * options.num_compute_units * options.bw_scalor mainCluster = Cluster(intBW=crossbar_bw) else: mainCluster = Cluster(intBW=8) # 16 GB/s for i in range(options.num_dirs): dir_cntrl = DirCntrl(noTCCdir = True, TCC_select_num_bits = TCC_bits) dir_cntrl.create(options, ruby_system, system) dir_cntrl.number_of_TBEs = options.num_tbes dir_cntrl.useL3OnWT = options.use_L3_on_WT # the number_of_TBEs is inclusive of TBEs below # Connect the Directory controller to the ruby network dir_cntrl.requestFromCores = MessageBuffer(ordered = True) dir_cntrl.requestFromCores.slave = ruby_system.network.master dir_cntrl.responseFromCores = MessageBuffer() dir_cntrl.responseFromCores.slave = ruby_system.network.master dir_cntrl.unblockFromCores = MessageBuffer() dir_cntrl.unblockFromCores.slave = ruby_system.network.master dir_cntrl.probeToCore = MessageBuffer() dir_cntrl.probeToCore.master = ruby_system.network.slave dir_cntrl.responseToCore = MessageBuffer() dir_cntrl.responseToCore.master = ruby_system.network.slave dir_cntrl.triggerQueue = MessageBuffer(ordered = True) dir_cntrl.L3triggerQueue = MessageBuffer(ordered = True) dir_cntrl.responseFromMemory = MessageBuffer() exec("ruby_system.dir_cntrl%d = dir_cntrl" % i) dir_cntrl_nodes.append(dir_cntrl) mainCluster.add(dir_cntrl) cpuCluster = None if hasattr(options, 'bw_scalor') and options.bw_scalor > 0: cpuCluster = Cluster(extBW = crossbar_bw, intBW = crossbar_bw) else: cpuCluster = Cluster(extBW = 8, intBW = 8) # 16 GB/s for i in range((options.num_cpus + 1) // 2): cp_cntrl = CPCntrl() cp_cntrl.create(options, ruby_system, system) exec("ruby_system.cp_cntrl%d = cp_cntrl" % i) # # Add controllers and sequencers to the appropriate lists # cpu_sequencers.extend([cp_cntrl.sequencer, cp_cntrl.sequencer1]) # Connect the CP controllers and the network cp_cntrl.requestFromCore = MessageBuffer() cp_cntrl.requestFromCore.master = ruby_system.network.slave cp_cntrl.responseFromCore = MessageBuffer() cp_cntrl.responseFromCore.master = ruby_system.network.slave cp_cntrl.unblockFromCore = MessageBuffer() cp_cntrl.unblockFromCore.master = ruby_system.network.slave cp_cntrl.probeToCore = MessageBuffer() cp_cntrl.probeToCore.slave = ruby_system.network.master cp_cntrl.responseToCore = MessageBuffer() cp_cntrl.responseToCore.slave = ruby_system.network.master cp_cntrl.mandatoryQueue = MessageBuffer() cp_cntrl.triggerQueue = MessageBuffer(ordered = True) cpuCluster.add(cp_cntrl) # Register CPUs and caches for each CorePair and directory (SE mode only) if not full_system: for i in xrange((options.num_cpus + 1) // 2): FileSystemConfig.register_cpu(physical_package_id = 0, core_siblings = \ xrange(options.num_cpus), core_id = i*2, thread_siblings = []) FileSystemConfig.register_cpu(physical_package_id = 0, core_siblings = \ xrange(options.num_cpus), core_id = i*2+1, thread_siblings = []) FileSystemConfig.register_cache(level = 0, idu_type = 'Instruction', size = options.l1i_size, line_size = options.cacheline_size, assoc = options.l1i_assoc, cpus = [i*2, i*2+1]) FileSystemConfig.register_cache(level = 0, idu_type = 'Data', size = options.l1d_size, line_size = options.cacheline_size, assoc = options.l1d_assoc, cpus = [i*2]) FileSystemConfig.register_cache(level = 0, idu_type = 'Data', size = options.l1d_size, line_size = options.cacheline_size, assoc = options.l1d_assoc, cpus = [i*2+1]) FileSystemConfig.register_cache(level = 1, idu_type = 'Unified', size = options.l2_size, line_size = options.cacheline_size, assoc = options.l2_assoc, cpus = [i*2, i*2+1]) for i in range(options.num_dirs): FileSystemConfig.register_cache(level = 2, idu_type = 'Unified', size = options.l3_size, line_size = options.cacheline_size, assoc = options.l3_assoc, cpus = [n for n in xrange(options.num_cpus)]) gpuCluster = None if hasattr(options, 'bw_scalor') and options.bw_scalor > 0: gpuCluster = Cluster(extBW = crossbar_bw, intBW = crossbar_bw) else: gpuCluster = Cluster(extBW = 8, intBW = 8) # 16 GB/s for i in range(options.num_compute_units): tcp_cntrl = TCPCntrl(TCC_select_num_bits = TCC_bits, issue_latency = 1, number_of_TBEs = 2560) # TBEs set to max outstanding requests tcp_cntrl.create(options, ruby_system, system) tcp_cntrl.WB = options.WB_L1 tcp_cntrl.disableL1 = options.noL1 tcp_cntrl.L1cache.tagAccessLatency = options.TCP_latency tcp_cntrl.L1cache.dataAccessLatency = options.TCP_latency exec("ruby_system.tcp_cntrl%d = tcp_cntrl" % i) # # Add controllers and sequencers to the appropriate lists # cpu_sequencers.append(tcp_cntrl.coalescer) tcp_cntrl_nodes.append(tcp_cntrl) # Connect the TCP controller to the ruby network tcp_cntrl.requestFromTCP = MessageBuffer(ordered = True) tcp_cntrl.requestFromTCP.master = ruby_system.network.slave tcp_cntrl.responseFromTCP = MessageBuffer(ordered = True) tcp_cntrl.responseFromTCP.master = ruby_system.network.slave tcp_cntrl.unblockFromCore = MessageBuffer() tcp_cntrl.unblockFromCore.master = ruby_system.network.slave tcp_cntrl.probeToTCP = MessageBuffer(ordered = True) tcp_cntrl.probeToTCP.slave = ruby_system.network.master tcp_cntrl.responseToTCP = MessageBuffer(ordered = True) tcp_cntrl.responseToTCP.slave = ruby_system.network.master tcp_cntrl.mandatoryQueue = MessageBuffer() gpuCluster.add(tcp_cntrl) for i in range(options.num_sqc): sqc_cntrl = SQCCntrl(TCC_select_num_bits = TCC_bits) sqc_cntrl.create(options, ruby_system, system) exec("ruby_system.sqc_cntrl%d = sqc_cntrl" % i) # # Add controllers and sequencers to the appropriate lists # cpu_sequencers.append(sqc_cntrl.sequencer) # Connect the SQC controller to the ruby network sqc_cntrl.requestFromSQC = MessageBuffer(ordered = True) sqc_cntrl.requestFromSQC.master = ruby_system.network.slave sqc_cntrl.probeToSQC = MessageBuffer(ordered = True) sqc_cntrl.probeToSQC.slave = ruby_system.network.master sqc_cntrl.responseToSQC = MessageBuffer(ordered = True) sqc_cntrl.responseToSQC.slave = ruby_system.network.master sqc_cntrl.mandatoryQueue = MessageBuffer() # SQC also in GPU cluster gpuCluster.add(sqc_cntrl) for i in range(options.num_cp): tcp_ID = options.num_compute_units + i sqc_ID = options.num_sqc + i tcp_cntrl = TCPCntrl(TCC_select_num_bits = TCC_bits, issue_latency = 1, number_of_TBEs = 2560) # TBEs set to max outstanding requests tcp_cntrl.createCP(options, ruby_system, system) tcp_cntrl.WB = options.WB_L1 tcp_cntrl.disableL1 = options.noL1 tcp_cntrl.L1cache.tagAccessLatency = options.TCP_latency tcp_cntrl.L1cache.dataAccessLatency = options.TCP_latency exec("ruby_system.tcp_cntrl%d = tcp_cntrl" % tcp_ID) # # Add controllers and sequencers to the appropriate lists # cpu_sequencers.append(tcp_cntrl.sequencer) tcp_cntrl_nodes.append(tcp_cntrl) # Connect the CP (TCP) controllers to the ruby network tcp_cntrl.requestFromTCP = MessageBuffer(ordered = True) tcp_cntrl.requestFromTCP.master = ruby_system.network.slave tcp_cntrl.responseFromTCP = MessageBuffer(ordered = True) tcp_cntrl.responseFromTCP.master = ruby_system.network.slave tcp_cntrl.unblockFromCore = MessageBuffer(ordered = True) tcp_cntrl.unblockFromCore.master = ruby_system.network.slave tcp_cntrl.probeToTCP = MessageBuffer(ordered = True) tcp_cntrl.probeToTCP.slave = ruby_system.network.master tcp_cntrl.responseToTCP = MessageBuffer(ordered = True) tcp_cntrl.responseToTCP.slave = ruby_system.network.master tcp_cntrl.mandatoryQueue = MessageBuffer() gpuCluster.add(tcp_cntrl) sqc_cntrl = SQCCntrl(TCC_select_num_bits = TCC_bits) sqc_cntrl.create(options, ruby_system, system) exec("ruby_system.sqc_cntrl%d = sqc_cntrl" % sqc_ID) # # Add controllers and sequencers to the appropriate lists # cpu_sequencers.append(sqc_cntrl.sequencer) # SQC also in GPU cluster gpuCluster.add(sqc_cntrl) for i in range(options.num_tccs): tcc_cntrl = TCCCntrl(l2_response_latency = options.TCC_latency) tcc_cntrl.create(options, ruby_system, system) tcc_cntrl.l2_request_latency = options.gpu_to_dir_latency tcc_cntrl.l2_response_latency = options.TCC_latency tcc_cntrl_nodes.append(tcc_cntrl) tcc_cntrl.WB = options.WB_L2 tcc_cntrl.number_of_TBEs = 2560 * options.num_compute_units # the number_of_TBEs is inclusive of TBEs below # Connect the TCC controllers to the ruby network tcc_cntrl.requestFromTCP = MessageBuffer(ordered = True) tcc_cntrl.requestFromTCP.slave = ruby_system.network.master tcc_cntrl.responseToCore = MessageBuffer(ordered = True) tcc_cntrl.responseToCore.master = ruby_system.network.slave tcc_cntrl.probeFromNB = MessageBuffer() tcc_cntrl.probeFromNB.slave = ruby_system.network.master tcc_cntrl.responseFromNB = MessageBuffer() tcc_cntrl.responseFromNB.slave = ruby_system.network.master tcc_cntrl.requestToNB = MessageBuffer(ordered = True) tcc_cntrl.requestToNB.master = ruby_system.network.slave tcc_cntrl.responseToNB = MessageBuffer() tcc_cntrl.responseToNB.master = ruby_system.network.slave tcc_cntrl.unblockToNB = MessageBuffer() tcc_cntrl.unblockToNB.master = ruby_system.network.slave tcc_cntrl.triggerQueue = MessageBuffer(ordered = True) exec("ruby_system.tcc_cntrl%d = tcc_cntrl" % i) # connect all of the wire buffers between L3 and dirs up # TCC cntrls added to the GPU cluster gpuCluster.add(tcc_cntrl) # Assuming no DMA devices assert(len(dma_devices) == 0) # Add cpu/gpu clusters to main cluster mainCluster.add(cpuCluster) mainCluster.add(gpuCluster) ruby_system.network.number_of_virtual_networks = 10 return (cpu_sequencers, dir_cntrl_nodes, mainCluster)
def create_system(options, full_system, system, dma_devices, bootmem, ruby_system, cpus): if buildEnv['PROTOCOL'] != 'GPU_VIPER': panic("This script requires the GPU_VIPER protocol to be built.") cpu_sequencers = [] # # Must create the individual controllers before the network to ensure the # controller constructors are called before the network constructor # # This is the base crossbar that connects the L3s, Dirs, and cpu/gpu # Clusters crossbar_bw = None mainCluster = None cpuCluster = None gpuCluster = None if hasattr(options, 'bw_scalor') and options.bw_scalor > 0: #Assuming a 2GHz clock crossbar_bw = 16 * options.num_compute_units * options.bw_scalor mainCluster = Cluster(intBW = crossbar_bw) cpuCluster = Cluster(extBW = crossbar_bw, intBW = crossbar_bw) gpuCluster = Cluster(extBW = crossbar_bw, intBW = crossbar_bw) else: mainCluster = Cluster(intBW = 8) # 16 GB/s cpuCluster = Cluster(extBW = 8, intBW = 8) # 16 GB/s gpuCluster = Cluster(extBW = 8, intBW = 8) # 16 GB/s # Create CPU directory controllers dir_cntrl_nodes = \ construct_dirs(options, system, ruby_system, ruby_system.network) for dir_cntrl in dir_cntrl_nodes: mainCluster.add(dir_cntrl) # Create CPU core pairs (cp_sequencers, cp_cntrl_nodes) = \ construct_corepairs(options, system, ruby_system, ruby_system.network) cpu_sequencers.extend(cp_sequencers) for cp_cntrl in cp_cntrl_nodes: cpuCluster.add(cp_cntrl) # Register CPUs and caches for each CorePair and directory (SE mode only) if not full_system: for i in range((options.num_cpus + 1) // 2): FileSystemConfig.register_cpu(physical_package_id = 0, core_siblings = \ range(options.num_cpus), core_id = i*2, thread_siblings = []) FileSystemConfig.register_cpu(physical_package_id = 0, core_siblings = \ range(options.num_cpus), core_id = i*2+1, thread_siblings = []) FileSystemConfig.register_cache(level = 0, idu_type = 'Instruction', size = options.l1i_size, line_size = options.cacheline_size, assoc = options.l1i_assoc, cpus = [i*2, i*2+1]) FileSystemConfig.register_cache(level = 0, idu_type = 'Data', size = options.l1d_size, line_size = options.cacheline_size, assoc = options.l1d_assoc, cpus = [i*2]) FileSystemConfig.register_cache(level = 0, idu_type = 'Data', size = options.l1d_size, line_size = options.cacheline_size, assoc = options.l1d_assoc, cpus = [i*2+1]) FileSystemConfig.register_cache(level = 1, idu_type = 'Unified', size = options.l2_size, line_size = options.cacheline_size, assoc = options.l2_assoc, cpus = [i*2, i*2+1]) for i in range(options.num_dirs): FileSystemConfig.register_cache(level = 2, idu_type = 'Unified', size = options.l3_size, line_size = options.cacheline_size, assoc = options.l3_assoc, cpus = [n for n in range(options.num_cpus)]) # Create TCPs (tcp_sequencers, tcp_cntrl_nodes) = \ construct_tcps(options, system, ruby_system, ruby_system.network) cpu_sequencers.extend(tcp_sequencers) for tcp_cntrl in tcp_cntrl_nodes: gpuCluster.add(tcp_cntrl) # Create SQCs (sqc_sequencers, sqc_cntrl_nodes) = \ construct_sqcs(options, system, ruby_system, ruby_system.network) cpu_sequencers.extend(sqc_sequencers) for sqc_cntrl in sqc_cntrl_nodes: gpuCluster.add(sqc_cntrl) # Create Scalars (scalar_sequencers, scalar_cntrl_nodes) = \ construct_scalars(options, system, ruby_system, ruby_system.network) cpu_sequencers.extend(scalar_sequencers) for scalar_cntrl in scalar_cntrl_nodes: gpuCluster.add(scalar_cntrl) # Create command processors (cmdproc_sequencers, cmdproc_cntrl_nodes) = \ construct_cmdprocs(options, system, ruby_system, ruby_system.network) cpu_sequencers.extend(cmdproc_sequencers) for cmdproc_cntrl in cmdproc_cntrl_nodes: gpuCluster.add(cmdproc_cntrl) # Create TCCs tcc_cntrl_nodes = \ construct_tccs(options, system, ruby_system, ruby_system.network) for tcc_cntrl in tcc_cntrl_nodes: gpuCluster.add(tcc_cntrl) for i, dma_device in enumerate(dma_devices): dma_seq = DMASequencer(version=i, ruby_system=ruby_system) dma_cntrl = DMA_Controller(version=i, dma_sequencer=dma_seq, ruby_system=ruby_system) exec('system.dma_cntrl%d = dma_cntrl' % i) # IDE doesn't have a .type but seems like everything else does. if not hasattr(dma_device, 'type'): exec('system.dma_cntrl%d.dma_sequencer.in_ports = dma_device' % i) elif dma_device.type == 'MemTest': exec('system.dma_cntrl%d.dma_sequencer.in_ports = dma_devices.test' % i) else: exec('system.dma_cntrl%d.dma_sequencer.in_ports = dma_device.dma' % i) dma_cntrl.requestToDir = MessageBuffer(buffer_size=0) dma_cntrl.requestToDir.out_port = ruby_system.network.in_port dma_cntrl.responseFromDir = MessageBuffer(buffer_size=0) dma_cntrl.responseFromDir.in_port = ruby_system.network.out_port dma_cntrl.mandatoryQueue = MessageBuffer(buffer_size = 0) gpuCluster.add(dma_cntrl) # Add cpu/gpu clusters to main cluster mainCluster.add(cpuCluster) mainCluster.add(gpuCluster) ruby_system.network.number_of_virtual_networks = 11 return (cpu_sequencers, dir_cntrl_nodes, mainCluster)
def create_system(options, full_system, system, dma_devices, bootmem, ruby_system): if buildEnv['PROTOCOL'] != 'GPU_VIPER_Baseline': panic("This script requires the" \ "GPU_VIPER_Baseline protocol to be built.") cpu_sequencers = [] # # The ruby network creation expects the list of nodes in the system to be # consistent with the NetDest list. Therefore the l1 controller nodes # must be listed before the directory nodes and directory nodes before # dma nodes, etc. # cp_cntrl_nodes = [] tcp_cntrl_nodes = [] sqc_cntrl_nodes = [] tcc_cntrl_nodes = [] dir_cntrl_nodes = [] l3_cntrl_nodes = [] # # Must create the individual controllers before the network to ensure the # controller constructors are called before the network constructor # # For an odd number of CPUs, still create the right number of controllers TCC_bits = int(math.log(options.num_tccs, 2)) # This is the base crossbar that connects the L3s, Dirs, and cpu/gpu # Clusters crossbar_bw = 16 * options.num_compute_units #Assuming a 2GHz clock mainCluster = Cluster(intBW = crossbar_bw) for i in range(options.num_dirs): dir_cntrl = DirCntrl(noTCCdir=True,TCC_select_num_bits = TCC_bits) dir_cntrl.create(options, ruby_system, system) dir_cntrl.number_of_TBEs = options.num_tbes dir_cntrl.useL3OnWT = options.use_L3_on_WT dir_cntrl.inclusiveDir = not options.nonInclusiveDir # Connect the Directory controller to the ruby network dir_cntrl.requestFromCores = MessageBuffer(ordered = True) dir_cntrl.requestFromCores.slave = ruby_system.network.master dir_cntrl.responseFromCores = MessageBuffer() dir_cntrl.responseFromCores.slave = ruby_system.network.master dir_cntrl.unblockFromCores = MessageBuffer() dir_cntrl.unblockFromCores.slave = ruby_system.network.master dir_cntrl.probeToCore = MessageBuffer() dir_cntrl.probeToCore.master = ruby_system.network.slave dir_cntrl.responseToCore = MessageBuffer() dir_cntrl.responseToCore.master = ruby_system.network.slave dir_cntrl.triggerQueue = MessageBuffer(ordered = True) dir_cntrl.L3triggerQueue = MessageBuffer(ordered = True) dir_cntrl.responseFromMemory = MessageBuffer() exec("system.dir_cntrl%d = dir_cntrl" % i) dir_cntrl_nodes.append(dir_cntrl) mainCluster.add(dir_cntrl) cpuCluster = Cluster(extBW = crossbar_bw, intBW=crossbar_bw) for i in range((options.num_cpus + 1) // 2): cp_cntrl = CPCntrl() cp_cntrl.create(options, ruby_system, system) exec("system.cp_cntrl%d = cp_cntrl" % i) # # Add controllers and sequencers to the appropriate lists # cpu_sequencers.extend([cp_cntrl.sequencer, cp_cntrl.sequencer1]) # Connect the CP controllers and the network cp_cntrl.requestFromCore = MessageBuffer() cp_cntrl.requestFromCore.master = ruby_system.network.slave cp_cntrl.responseFromCore = MessageBuffer() cp_cntrl.responseFromCore.master = ruby_system.network.slave cp_cntrl.unblockFromCore = MessageBuffer() cp_cntrl.unblockFromCore.master = ruby_system.network.slave cp_cntrl.probeToCore = MessageBuffer() cp_cntrl.probeToCore.slave = ruby_system.network.master cp_cntrl.responseToCore = MessageBuffer() cp_cntrl.responseToCore.slave = ruby_system.network.master cp_cntrl.mandatoryQueue = MessageBuffer() cp_cntrl.triggerQueue = MessageBuffer(ordered = True) cpuCluster.add(cp_cntrl) gpuCluster = Cluster(extBW = crossbar_bw, intBW = crossbar_bw) for i in range(options.num_compute_units): tcp_cntrl = TCPCntrl(TCC_select_num_bits = TCC_bits, issue_latency = 1, number_of_TBEs = 2560) # TBEs set to max outstanding requests tcp_cntrl.create(options, ruby_system, system) tcp_cntrl.WB = options.WB_L1 tcp_cntrl.disableL1 = options.noL1 exec("system.tcp_cntrl%d = tcp_cntrl" % i) # # Add controllers and sequencers to the appropriate lists # cpu_sequencers.append(tcp_cntrl.coalescer) tcp_cntrl_nodes.append(tcp_cntrl) # Connect the CP (TCP) controllers to the ruby network tcp_cntrl.requestFromTCP = MessageBuffer(ordered = True) tcp_cntrl.requestFromTCP.master = ruby_system.network.slave tcp_cntrl.responseFromTCP = MessageBuffer(ordered = True) tcp_cntrl.responseFromTCP.master = ruby_system.network.slave tcp_cntrl.unblockFromCore = MessageBuffer() tcp_cntrl.unblockFromCore.master = ruby_system.network.slave tcp_cntrl.probeToTCP = MessageBuffer(ordered = True) tcp_cntrl.probeToTCP.slave = ruby_system.network.master tcp_cntrl.responseToTCP = MessageBuffer(ordered = True) tcp_cntrl.responseToTCP.slave = ruby_system.network.master tcp_cntrl.mandatoryQueue = MessageBuffer() gpuCluster.add(tcp_cntrl) for i in range(options.num_sqc): sqc_cntrl = SQCCntrl(TCC_select_num_bits = TCC_bits) sqc_cntrl.create(options, ruby_system, system) exec("system.sqc_cntrl%d = sqc_cntrl" % i) # # Add controllers and sequencers to the appropriate lists # cpu_sequencers.append(sqc_cntrl.sequencer) # Connect the SQC controller to the ruby network sqc_cntrl.requestFromSQC = MessageBuffer(ordered = True) sqc_cntrl.requestFromSQC.master = ruby_system.network.slave sqc_cntrl.probeToSQC = MessageBuffer(ordered = True) sqc_cntrl.probeToSQC.slave = ruby_system.network.master sqc_cntrl.responseToSQC = MessageBuffer(ordered = True) sqc_cntrl.responseToSQC.slave = ruby_system.network.master sqc_cntrl.mandatoryQueue = MessageBuffer() # SQC also in GPU cluster gpuCluster.add(sqc_cntrl) # Because of wire buffers, num_tccs must equal num_tccdirs numa_bit = 6 for i in range(options.num_tccs): tcc_cntrl = TCCCntrl() tcc_cntrl.create(options, ruby_system, system) tcc_cntrl.l2_request_latency = options.gpu_to_dir_latency tcc_cntrl.l2_response_latency = options.TCC_latency tcc_cntrl_nodes.append(tcc_cntrl) tcc_cntrl.WB = options.WB_L2 tcc_cntrl.number_of_TBEs = 2560 * options.num_compute_units # Connect the TCC controllers to the ruby network tcc_cntrl.requestFromTCP = MessageBuffer(ordered = True) tcc_cntrl.requestFromTCP.slave = ruby_system.network.master tcc_cntrl.responseToCore = MessageBuffer(ordered = True) tcc_cntrl.responseToCore.master = ruby_system.network.slave tcc_cntrl.probeFromNB = MessageBuffer() tcc_cntrl.probeFromNB.slave = ruby_system.network.master tcc_cntrl.responseFromNB = MessageBuffer() tcc_cntrl.responseFromNB.slave = ruby_system.network.master tcc_cntrl.requestToNB = MessageBuffer(ordered = True) tcc_cntrl.requestToNB.master = ruby_system.network.slave tcc_cntrl.responseToNB = MessageBuffer() tcc_cntrl.responseToNB.master = ruby_system.network.slave tcc_cntrl.unblockToNB = MessageBuffer() tcc_cntrl.unblockToNB.master = ruby_system.network.slave tcc_cntrl.triggerQueue = MessageBuffer(ordered = True) exec("system.tcc_cntrl%d = tcc_cntrl" % i) # connect all of the wire buffers between L3 and dirs up # TCC cntrls added to the GPU cluster gpuCluster.add(tcc_cntrl) # Assuming no DMA devices assert(len(dma_devices) == 0) # Add cpu/gpu clusters to main cluster mainCluster.add(cpuCluster) mainCluster.add(gpuCluster) ruby_system.network.number_of_virtual_networks = 10 return (cpu_sequencers, dir_cntrl_nodes, mainCluster)
def create_system(options, full_system, system, dma_devices, bootmem, ruby_system): if buildEnv['PROTOCOL'] != 'GPU_VIPER_Region': panic( "This script requires the GPU_VIPER_Region protocol to be built.") cpu_sequencers = [] # # The ruby network creation expects the list of nodes in the system to be # consistent with the NetDest list. Therefore the l1 controller nodes # must be listed before the directory nodes and directory nodes before # dma nodes, etc. # dir_cntrl_nodes = [] # For an odd number of CPUs, still create the right number of controllers TCC_bits = int(math.log(options.num_tccs, 2)) # # Must create the individual controllers before the network to ensure the # controller constructors are called before the network constructor # # For an odd number of CPUs, still create the right number of controllers crossbar_bw = 16 * options.num_compute_units #Assuming a 2GHz clock cpuCluster = Cluster(extBW=(crossbar_bw), intBW=crossbar_bw) for i in xrange((options.num_cpus + 1) / 2): cp_cntrl = CPCntrl() cp_cntrl.create(options, ruby_system, system) rb_cntrl = RBCntrl() rb_cntrl.create(options, ruby_system, system) rb_cntrl.number_of_TBEs = 256 rb_cntrl.isOnCPU = True cp_cntrl.regionBufferNum = rb_cntrl.version exec("system.cp_cntrl%d = cp_cntrl" % i) exec("system.rb_cntrl%d = rb_cntrl" % i) # # Add controllers and sequencers to the appropriate lists # cpu_sequencers.extend([cp_cntrl.sequencer, cp_cntrl.sequencer1]) # Connect the CP controllers and the network cp_cntrl.requestFromCore = MessageBuffer() cp_cntrl.requestFromCore.master = ruby_system.network.slave cp_cntrl.responseFromCore = MessageBuffer() cp_cntrl.responseFromCore.master = ruby_system.network.slave cp_cntrl.unblockFromCore = MessageBuffer() cp_cntrl.unblockFromCore.master = ruby_system.network.slave cp_cntrl.probeToCore = MessageBuffer() cp_cntrl.probeToCore.slave = ruby_system.network.master cp_cntrl.responseToCore = MessageBuffer() cp_cntrl.responseToCore.slave = ruby_system.network.master cp_cntrl.mandatoryQueue = MessageBuffer() cp_cntrl.triggerQueue = MessageBuffer(ordered=True) # Connect the RB controllers to the ruby network rb_cntrl.requestFromCore = MessageBuffer(ordered=True) rb_cntrl.requestFromCore.slave = ruby_system.network.master rb_cntrl.responseFromCore = MessageBuffer() rb_cntrl.responseFromCore.slave = ruby_system.network.master rb_cntrl.requestToNetwork = MessageBuffer() rb_cntrl.requestToNetwork.master = ruby_system.network.slave rb_cntrl.notifyFromRegionDir = MessageBuffer() rb_cntrl.notifyFromRegionDir.slave = ruby_system.network.master rb_cntrl.probeFromRegionDir = MessageBuffer() rb_cntrl.probeFromRegionDir.slave = ruby_system.network.master rb_cntrl.unblockFromDir = MessageBuffer() rb_cntrl.unblockFromDir.slave = ruby_system.network.master rb_cntrl.responseToRegDir = MessageBuffer() rb_cntrl.responseToRegDir.master = ruby_system.network.slave rb_cntrl.triggerQueue = MessageBuffer(ordered=True) cpuCluster.add(cp_cntrl) cpuCluster.add(rb_cntrl) gpuCluster = Cluster(extBW=(crossbar_bw), intBW=crossbar_bw) for i in xrange(options.num_compute_units): tcp_cntrl = TCPCntrl(TCC_select_num_bits=TCC_bits, issue_latency=1, number_of_TBEs=2560) # TBEs set to max outstanding requests tcp_cntrl.create(options, ruby_system, system) tcp_cntrl.WB = options.WB_L1 tcp_cntrl.disableL1 = False exec("system.tcp_cntrl%d = tcp_cntrl" % i) # # Add controllers and sequencers to the appropriate lists # cpu_sequencers.append(tcp_cntrl.coalescer) # Connect the CP (TCP) controllers to the ruby network tcp_cntrl.requestFromTCP = MessageBuffer(ordered=True) tcp_cntrl.requestFromTCP.master = ruby_system.network.slave tcp_cntrl.responseFromTCP = MessageBuffer(ordered=True) tcp_cntrl.responseFromTCP.master = ruby_system.network.slave tcp_cntrl.unblockFromCore = MessageBuffer() tcp_cntrl.unblockFromCore.master = ruby_system.network.slave tcp_cntrl.probeToTCP = MessageBuffer(ordered=True) tcp_cntrl.probeToTCP.slave = ruby_system.network.master tcp_cntrl.responseToTCP = MessageBuffer(ordered=True) tcp_cntrl.responseToTCP.slave = ruby_system.network.master tcp_cntrl.mandatoryQueue = MessageBuffer() gpuCluster.add(tcp_cntrl) for i in xrange(options.num_sqc): sqc_cntrl = SQCCntrl(TCC_select_num_bits=TCC_bits) sqc_cntrl.create(options, ruby_system, system) exec("system.sqc_cntrl%d = sqc_cntrl" % i) # # Add controllers and sequencers to the appropriate lists # cpu_sequencers.append(sqc_cntrl.sequencer) # Connect the SQC controller to the ruby network sqc_cntrl.requestFromSQC = MessageBuffer(ordered=True) sqc_cntrl.requestFromSQC.master = ruby_system.network.slave sqc_cntrl.probeToSQC = MessageBuffer(ordered=True) sqc_cntrl.probeToSQC.slave = ruby_system.network.master sqc_cntrl.responseToSQC = MessageBuffer(ordered=True) sqc_cntrl.responseToSQC.slave = ruby_system.network.master sqc_cntrl.mandatoryQueue = MessageBuffer() # SQC also in GPU cluster gpuCluster.add(sqc_cntrl) numa_bit = 6 for i in xrange(options.num_tccs): tcc_cntrl = TCCCntrl() tcc_cntrl.create(options, ruby_system, system) tcc_cntrl.l2_request_latency = 1 tcc_cntrl.l2_response_latency = options.TCC_latency tcc_cntrl.WB = options.WB_L2 tcc_cntrl.number_of_TBEs = 2560 * options.num_compute_units # Connect the TCC controllers to the ruby network tcc_cntrl.requestFromTCP = MessageBuffer(ordered=True) tcc_cntrl.requestFromTCP.slave = ruby_system.network.master tcc_cntrl.responseToCore = MessageBuffer(ordered=True) tcc_cntrl.responseToCore.master = ruby_system.network.slave tcc_cntrl.probeFromNB = MessageBuffer() tcc_cntrl.probeFromNB.slave = ruby_system.network.master tcc_cntrl.responseFromNB = MessageBuffer() tcc_cntrl.responseFromNB.slave = ruby_system.network.master tcc_cntrl.requestToNB = MessageBuffer(ordered=True) tcc_cntrl.requestToNB.master = ruby_system.network.slave tcc_cntrl.responseToNB = MessageBuffer() tcc_cntrl.responseToNB.master = ruby_system.network.slave tcc_cntrl.unblockToNB = MessageBuffer() tcc_cntrl.unblockToNB.master = ruby_system.network.slave tcc_cntrl.triggerQueue = MessageBuffer(ordered=True) rb_cntrl = RBCntrl() rb_cntrl.create(options, ruby_system, system) rb_cntrl.number_of_TBEs = 2560 * options.num_compute_units rb_cntrl.isOnCPU = False # Connect the RB controllers to the ruby network rb_cntrl.requestFromCore = MessageBuffer(ordered=True) rb_cntrl.requestFromCore.slave = ruby_system.network.master rb_cntrl.responseFromCore = MessageBuffer() rb_cntrl.responseFromCore.slave = ruby_system.network.master rb_cntrl.requestToNetwork = MessageBuffer() rb_cntrl.requestToNetwork.master = ruby_system.network.slave rb_cntrl.notifyFromRegionDir = MessageBuffer() rb_cntrl.notifyFromRegionDir.slave = ruby_system.network.master rb_cntrl.probeFromRegionDir = MessageBuffer() rb_cntrl.probeFromRegionDir.slave = ruby_system.network.master rb_cntrl.unblockFromDir = MessageBuffer() rb_cntrl.unblockFromDir.slave = ruby_system.network.master rb_cntrl.responseToRegDir = MessageBuffer() rb_cntrl.responseToRegDir.master = ruby_system.network.slave rb_cntrl.triggerQueue = MessageBuffer(ordered=True) tcc_cntrl.regionBufferNum = rb_cntrl.version exec("system.tcc_cntrl%d = tcc_cntrl" % i) exec("system.tcc_rb_cntrl%d = rb_cntrl" % i) # TCC cntrls added to the GPU cluster gpuCluster.add(tcc_cntrl) gpuCluster.add(rb_cntrl) # Because of wire buffers, num_l3caches must equal num_dirs # Region coherence only works with 1 dir assert (options.num_l3caches == options.num_dirs == 1) # This is the base crossbar that connects the L3s, Dirs, and cpu/gpu # Clusters mainCluster = Cluster(intBW=crossbar_bw) dir_cntrl = DirCntrl() dir_cntrl.create(options, ruby_system, system) dir_cntrl.number_of_TBEs = 2560 * options.num_compute_units dir_cntrl.useL3OnWT = options.use_L3_on_WT # Connect the Directory controller to the ruby network dir_cntrl.requestFromCores = MessageBuffer() dir_cntrl.requestFromCores.slave = ruby_system.network.master dir_cntrl.responseFromCores = MessageBuffer() dir_cntrl.responseFromCores.slave = ruby_system.network.master dir_cntrl.unblockFromCores = MessageBuffer() dir_cntrl.unblockFromCores.slave = ruby_system.network.master dir_cntrl.probeToCore = MessageBuffer() dir_cntrl.probeToCore.master = ruby_system.network.slave dir_cntrl.responseToCore = MessageBuffer() dir_cntrl.responseToCore.master = ruby_system.network.slave dir_cntrl.reqFromRegBuf = MessageBuffer() dir_cntrl.reqFromRegBuf.slave = ruby_system.network.master dir_cntrl.reqToRegDir = MessageBuffer(ordered=True) dir_cntrl.reqToRegDir.master = ruby_system.network.slave dir_cntrl.reqFromRegDir = MessageBuffer(ordered=True) dir_cntrl.reqFromRegDir.slave = ruby_system.network.master dir_cntrl.unblockToRegDir = MessageBuffer() dir_cntrl.unblockToRegDir.master = ruby_system.network.slave dir_cntrl.triggerQueue = MessageBuffer(ordered=True) dir_cntrl.L3triggerQueue = MessageBuffer(ordered=True) dir_cntrl.responseFromMemory = MessageBuffer() exec("system.dir_cntrl%d = dir_cntrl" % i) dir_cntrl_nodes.append(dir_cntrl) mainCluster.add(dir_cntrl) reg_cntrl = RegionCntrl(noTCCdir=True, TCC_select_num_bits=TCC_bits) reg_cntrl.create(options, ruby_system, system) reg_cntrl.number_of_TBEs = options.num_tbes reg_cntrl.cpuRegionBufferNum = system.rb_cntrl0.version reg_cntrl.gpuRegionBufferNum = system.tcc_rb_cntrl0.version # Connect the Region Dir controllers to the ruby network reg_cntrl.requestToDir = MessageBuffer(ordered=True) reg_cntrl.requestToDir.master = ruby_system.network.slave reg_cntrl.notifyToRBuffer = MessageBuffer() reg_cntrl.notifyToRBuffer.master = ruby_system.network.slave reg_cntrl.probeToRBuffer = MessageBuffer() reg_cntrl.probeToRBuffer.master = ruby_system.network.slave reg_cntrl.responseFromRBuffer = MessageBuffer() reg_cntrl.responseFromRBuffer.slave = ruby_system.network.master reg_cntrl.requestFromRegBuf = MessageBuffer() reg_cntrl.requestFromRegBuf.slave = ruby_system.network.master reg_cntrl.triggerQueue = MessageBuffer(ordered=True) exec("system.reg_cntrl%d = reg_cntrl" % i) mainCluster.add(reg_cntrl) # Assuming no DMA devices assert (len(dma_devices) == 0) # Add cpu/gpu clusters to main cluster mainCluster.add(cpuCluster) mainCluster.add(gpuCluster) ruby_system.network.number_of_virtual_networks = 10 return (cpu_sequencers, dir_cntrl_nodes, mainCluster)
def create_system(options, full_system, system, dma_devices, bootmem, ruby_system): if buildEnv['PROTOCOL'] != 'MOESI_AMD_Base': panic("This script requires the MOESI_AMD_Base protocol.") cpu_sequencers = [] # # The ruby network creation expects the list of nodes in the system to # be consistent with the NetDest list. Therefore the l1 controller # nodes must be listed before the directory nodes and directory nodes # before dma nodes, etc. # l1_cntrl_nodes = [] l3_cntrl_nodes = [] dir_cntrl_nodes = [] control_count = 0 # # Must create the individual controllers before the network to ensure # the controller constructors are called before the network constructor # # This is the base crossbar that connects the L3s, Dirs, and cpu # Cluster mainCluster = Cluster(extBW = 512, intBW = 512) # 1 TB/s if options.numa_high_bit: numa_bit = options.numa_high_bit else: # if the numa_bit is not specified, set the directory bits as the # lowest bits above the block offset bits, and the numa_bit as the # highest of those directory bits dir_bits = int(math.log(options.num_dirs, 2)) block_size_bits = int(math.log(options.cacheline_size, 2)) numa_bit = block_size_bits + dir_bits - 1 for i in range(options.num_dirs): dir_ranges = [] for r in system.mem_ranges: addr_range = m5.objects.AddrRange(r.start, size = r.size(), intlvHighBit = numa_bit, intlvBits = dir_bits, intlvMatch = i) dir_ranges.append(addr_range) dir_cntrl = DirCntrl(TCC_select_num_bits = 0) dir_cntrl.create(options, dir_ranges, ruby_system, system) # Connect the Directory controller to the ruby network dir_cntrl.requestFromCores = MessageBuffer(ordered = True) dir_cntrl.requestFromCores.slave = ruby_system.network.master dir_cntrl.responseFromCores = MessageBuffer() dir_cntrl.responseFromCores.slave = ruby_system.network.master dir_cntrl.unblockFromCores = MessageBuffer() dir_cntrl.unblockFromCores.slave = ruby_system.network.master dir_cntrl.probeToCore = MessageBuffer() dir_cntrl.probeToCore.master = ruby_system.network.slave dir_cntrl.responseToCore = MessageBuffer() dir_cntrl.responseToCore.master = ruby_system.network.slave dir_cntrl.triggerQueue = MessageBuffer(ordered = True) dir_cntrl.L3triggerQueue = MessageBuffer(ordered = True) dir_cntrl.responseFromMemory = MessageBuffer() exec("system.dir_cntrl%d = dir_cntrl" % i) dir_cntrl_nodes.append(dir_cntrl) mainCluster.add(dir_cntrl) # Technically this config can support an odd number of cpus, but the top # level config files, such as the ruby_random_tester, will get confused if # the number of cpus does not equal the number of sequencers. Thus make # sure that an even number of cpus is specified. assert((options.num_cpus % 2) == 0) # For an odd number of CPUs, still create the right number of controllers cpuCluster = Cluster(extBW = 512, intBW = 512) # 1 TB/s for i in range((options.num_cpus + 1) // 2): cp_cntrl = CPCntrl() cp_cntrl.create(options, ruby_system, system) exec("system.cp_cntrl%d = cp_cntrl" % i) # # Add controllers and sequencers to the appropriate lists # cpu_sequencers.extend([cp_cntrl.sequencer, cp_cntrl.sequencer1]) # Connect the CP controllers and the network cp_cntrl.requestFromCore = MessageBuffer() cp_cntrl.requestFromCore.master = ruby_system.network.slave cp_cntrl.responseFromCore = MessageBuffer() cp_cntrl.responseFromCore.master = ruby_system.network.slave cp_cntrl.unblockFromCore = MessageBuffer() cp_cntrl.unblockFromCore.master = ruby_system.network.slave cp_cntrl.probeToCore = MessageBuffer() cp_cntrl.probeToCore.slave = ruby_system.network.master cp_cntrl.responseToCore = MessageBuffer() cp_cntrl.responseToCore.slave = ruby_system.network.master cp_cntrl.mandatoryQueue = MessageBuffer() cp_cntrl.triggerQueue = MessageBuffer(ordered = True) cpuCluster.add(cp_cntrl) # Register CPUs and caches for each CorePair and directory (SE mode only) if not full_system: FileSystemConfig.config_filesystem(options) for i in xrange((options.num_cpus + 1) // 2): FileSystemConfig.register_cpu(physical_package_id = 0, core_siblings = xrange(options.num_cpus), core_id = i*2, thread_siblings = []) FileSystemConfig.register_cpu(physical_package_id = 0, core_siblings = xrange(options.num_cpus), core_id = i*2+1, thread_siblings = []) FileSystemConfig.register_cache(level = 0, idu_type = 'Instruction', size = options.l1i_size, line_size = options.cacheline_size, assoc = options.l1i_assoc, cpus = [i*2, i*2+1]) FileSystemConfig.register_cache(level = 0, idu_type = 'Data', size = options.l1d_size, line_size = options.cacheline_size, assoc = options.l1d_assoc, cpus = [i*2]) FileSystemConfig.register_cache(level = 0, idu_type = 'Data', size = options.l1d_size, line_size = options.cacheline_size, assoc = options.l1d_assoc, cpus = [i*2+1]) FileSystemConfig.register_cache(level = 1, idu_type = 'Unified', size = options.l2_size, line_size = options.cacheline_size, assoc = options.l2_assoc, cpus = [i*2, i*2+1]) for i in range(options.num_dirs): FileSystemConfig.register_cache(level = 2, idu_type = 'Unified', size = options.l3_size, line_size = options.cacheline_size, assoc = options.l3_assoc, cpus = [n for n in xrange(options.num_cpus)]) # Assuming no DMA devices assert(len(dma_devices) == 0) # Add cpu/gpu clusters to main cluster mainCluster.add(cpuCluster) ruby_system.network.number_of_virtual_networks = 10 return (cpu_sequencers, dir_cntrl_nodes, mainCluster)
def create_system(options, full_system, system, dma_devices, bootmem, ruby_system): if buildEnv['PROTOCOL'] != 'GPU_RfO': panic("This script requires the GPU_RfO protocol to be built.") cpu_sequencers = [] # # The ruby network creation expects the list of nodes in the system to be # consistent with the NetDest list. Therefore the l1 controller nodes # must be listed before the directory nodes and directory nodes before # dma nodes, etc. # cp_cntrl_nodes = [] tcp_cntrl_nodes = [] sqc_cntrl_nodes = [] tcc_cntrl_nodes = [] tccdir_cntrl_nodes = [] dir_cntrl_nodes = [] l3_cntrl_nodes = [] # # Must create the individual controllers before the network to ensure the # controller constructors are called before the network constructor # TCC_bits = int(math.log(options.num_tccs, 2)) # This is the base crossbar that connects the L3s, Dirs, and cpu/gpu # Clusters mainCluster = Cluster(extBW = 512, intBW = 512) # 1 TB/s if options.numa_high_bit: numa_bit = options.numa_high_bit else: # if the numa_bit is not specified, set the directory bits as the # lowest bits above the block offset bits, and the numa_bit as the # highest of those directory bits dir_bits = int(math.log(options.num_dirs, 2)) block_size_bits = int(math.log(options.cacheline_size, 2)) numa_bit = block_size_bits + dir_bits - 1 for i in xrange(options.num_dirs): dir_ranges = [] for r in system.mem_ranges: addr_range = m5.objects.AddrRange(r.start, size = r.size(), intlvHighBit = numa_bit, intlvBits = dir_bits, intlvMatch = i) dir_ranges.append(addr_range) dir_cntrl = DirCntrl(TCC_select_num_bits = TCC_bits) dir_cntrl.create(options, dir_ranges, ruby_system, system) dir_cntrl.number_of_TBEs = 2560 * options.num_compute_units #Enough TBEs for all TCP TBEs # Connect the Directory controller to the ruby network dir_cntrl.requestFromCores = MessageBuffer(ordered = True) dir_cntrl.requestFromCores.slave = ruby_system.network.master dir_cntrl.responseFromCores = MessageBuffer() dir_cntrl.responseFromCores.slave = ruby_system.network.master dir_cntrl.unblockFromCores = MessageBuffer() dir_cntrl.unblockFromCores.slave = ruby_system.network.master dir_cntrl.probeToCore = MessageBuffer() dir_cntrl.probeToCore.master = ruby_system.network.slave dir_cntrl.responseToCore = MessageBuffer() dir_cntrl.responseToCore.master = ruby_system.network.slave dir_cntrl.triggerQueue = MessageBuffer(ordered = True) dir_cntrl.L3triggerQueue = MessageBuffer(ordered = True) dir_cntrl.responseFromMemory = MessageBuffer() exec("system.dir_cntrl%d = dir_cntrl" % i) dir_cntrl_nodes.append(dir_cntrl) mainCluster.add(dir_cntrl) # For an odd number of CPUs, still create the right number of controllers cpuCluster = Cluster(extBW = 512, intBW = 512) # 1 TB/s for i in xrange((options.num_cpus + 1) / 2): cp_cntrl = CPCntrl() cp_cntrl.create(options, ruby_system, system) exec("system.cp_cntrl%d = cp_cntrl" % i) # # Add controllers and sequencers to the appropriate lists # cpu_sequencers.extend([cp_cntrl.sequencer, cp_cntrl.sequencer1]) # Connect the CP controllers and the network cp_cntrl.requestFromCore = MessageBuffer() cp_cntrl.requestFromCore.master = ruby_system.network.slave cp_cntrl.responseFromCore = MessageBuffer() cp_cntrl.responseFromCore.master = ruby_system.network.slave cp_cntrl.unblockFromCore = MessageBuffer() cp_cntrl.unblockFromCore.master = ruby_system.network.slave cp_cntrl.probeToCore = MessageBuffer() cp_cntrl.probeToCore.slave = ruby_system.network.master cp_cntrl.responseToCore = MessageBuffer() cp_cntrl.responseToCore.slave = ruby_system.network.master cp_cntrl.mandatoryQueue = MessageBuffer() cp_cntrl.triggerQueue = MessageBuffer(ordered = True) cpuCluster.add(cp_cntrl) gpuCluster = Cluster(extBW = 512, intBW = 512) # 1 TB/s for i in xrange(options.num_compute_units): tcp_cntrl = TCPCntrl(TCC_select_num_bits = TCC_bits, number_of_TBEs = 2560) # max outstanding requests tcp_cntrl.create(options, ruby_system, system) exec("system.tcp_cntrl%d = tcp_cntrl" % i) # # Add controllers and sequencers to the appropriate lists # cpu_sequencers.append(tcp_cntrl.coalescer) tcp_cntrl_nodes.append(tcp_cntrl) # Connect the TCP controller to the ruby network tcp_cntrl.requestFromTCP = MessageBuffer(ordered = True) tcp_cntrl.requestFromTCP.master = ruby_system.network.slave tcp_cntrl.responseFromTCP = MessageBuffer(ordered = True) tcp_cntrl.responseFromTCP.master = ruby_system.network.slave tcp_cntrl.unblockFromCore = MessageBuffer(ordered = True) tcp_cntrl.unblockFromCore.master = ruby_system.network.slave tcp_cntrl.probeToTCP = MessageBuffer(ordered = True) tcp_cntrl.probeToTCP.slave = ruby_system.network.master tcp_cntrl.responseToTCP = MessageBuffer(ordered = True) tcp_cntrl.responseToTCP.slave = ruby_system.network.master tcp_cntrl.mandatoryQueue = MessageBuffer() gpuCluster.add(tcp_cntrl) for i in xrange(options.num_sqc): sqc_cntrl = SQCCntrl(TCC_select_num_bits = TCC_bits) sqc_cntrl.create(options, ruby_system, system) exec("system.sqc_cntrl%d = sqc_cntrl" % i) # # Add controllers and sequencers to the appropriate lists # cpu_sequencers.append(sqc_cntrl.sequencer) # Connect the SQC controller to the ruby network sqc_cntrl.requestFromSQC = MessageBuffer(ordered = True) sqc_cntrl.requestFromSQC.master = ruby_system.network.slave sqc_cntrl.responseFromSQC = MessageBuffer(ordered = True) sqc_cntrl.responseFromSQC.master = ruby_system.network.slave sqc_cntrl.unblockFromCore = MessageBuffer(ordered = True) sqc_cntrl.unblockFromCore.master = ruby_system.network.slave sqc_cntrl.probeToSQC = MessageBuffer(ordered = True) sqc_cntrl.probeToSQC.slave = ruby_system.network.master sqc_cntrl.responseToSQC = MessageBuffer(ordered = True) sqc_cntrl.responseToSQC.slave = ruby_system.network.master sqc_cntrl.mandatoryQueue = MessageBuffer() # SQC also in GPU cluster gpuCluster.add(sqc_cntrl) for i in xrange(options.num_cp): tcp_cntrl = TCPCntrl(TCC_select_num_bits = TCC_bits, number_of_TBEs = 2560) # max outstanding requests tcp_cntrl.createCP(options, ruby_system, system) exec("system.tcp_cntrl%d = tcp_cntrl" % (options.num_compute_units + i)) # # Add controllers and sequencers to the appropriate lists # cpu_sequencers.append(tcp_cntrl.sequencer) tcp_cntrl_nodes.append(tcp_cntrl) # Connect the TCP controller to the ruby network tcp_cntrl.requestFromTCP = MessageBuffer(ordered = True) tcp_cntrl.requestFromTCP.master = ruby_system.network.slave tcp_cntrl.responseFromTCP = MessageBuffer(ordered = True) tcp_cntrl.responseFromTCP.master = ruby_system.network.slave tcp_cntrl.unblockFromCore = MessageBuffer(ordered = True) tcp_cntrl.unblockFromCore.master = ruby_system.network.slave tcp_cntrl.probeToTCP = MessageBuffer(ordered = True) tcp_cntrl.probeToTCP.slave = ruby_system.network.master tcp_cntrl.responseToTCP = MessageBuffer(ordered = True) tcp_cntrl.responseToTCP.slave = ruby_system.network.master tcp_cntrl.mandatoryQueue = MessageBuffer() gpuCluster.add(tcp_cntrl) sqc_cntrl = SQCCntrl(TCC_select_num_bits = TCC_bits) sqc_cntrl.createCP(options, ruby_system, system) exec("system.sqc_cntrl%d = sqc_cntrl" % (options.num_compute_units + i)) # # Add controllers and sequencers to the appropriate lists # cpu_sequencers.append(sqc_cntrl.sequencer) # Connect the SQC controller to the ruby network sqc_cntrl.requestFromSQC = MessageBuffer(ordered = True) sqc_cntrl.requestFromSQC.master = ruby_system.network.slave sqc_cntrl.responseFromSQC = MessageBuffer(ordered = True) sqc_cntrl.responseFromSQC.master = ruby_system.network.slave sqc_cntrl.unblockFromCore = MessageBuffer(ordered = True) sqc_cntrl.unblockFromCore.master = ruby_system.network.slave sqc_cntrl.probeToSQC = MessageBuffer(ordered = True) sqc_cntrl.probeToSQC.slave = ruby_system.network.master sqc_cntrl.responseToSQC = MessageBuffer(ordered = True) sqc_cntrl.responseToSQC.slave = ruby_system.network.master sqc_cntrl.mandatoryQueue = MessageBuffer() # SQC also in GPU cluster gpuCluster.add(sqc_cntrl) for i in xrange(options.num_tccs): tcc_cntrl = TCCCntrl(TCC_select_num_bits = TCC_bits, number_of_TBEs = options.num_compute_units * 2560) #Enough TBEs for all TCP TBEs tcc_cntrl.create(options, ruby_system, system) tcc_cntrl_nodes.append(tcc_cntrl) tccdir_cntrl = TCCDirCntrl(TCC_select_num_bits = TCC_bits, number_of_TBEs = options.num_compute_units * 2560) #Enough TBEs for all TCP TBEs tccdir_cntrl.create(options, ruby_system, system) tccdir_cntrl_nodes.append(tccdir_cntrl) exec("system.tcc_cntrl%d = tcc_cntrl" % i) exec("system.tccdir_cntrl%d = tccdir_cntrl" % i) # connect all of the wire buffers between L3 and dirs up req_to_tccdir = RubyWireBuffer() resp_to_tccdir = RubyWireBuffer() tcc_unblock_to_tccdir = RubyWireBuffer() req_to_tcc = RubyWireBuffer() probe_to_tcc = RubyWireBuffer() resp_to_tcc = RubyWireBuffer() tcc_cntrl.connectWireBuffers(req_to_tccdir, resp_to_tccdir, tcc_unblock_to_tccdir, req_to_tcc, probe_to_tcc, resp_to_tcc) tccdir_cntrl.connectWireBuffers(req_to_tccdir, resp_to_tccdir, tcc_unblock_to_tccdir, req_to_tcc, probe_to_tcc, resp_to_tcc) # Connect the TCC controller to the ruby network tcc_cntrl.responseFromTCC = MessageBuffer(ordered = True) tcc_cntrl.responseFromTCC.master = ruby_system.network.slave tcc_cntrl.responseToTCC = MessageBuffer(ordered = True) tcc_cntrl.responseToTCC.slave = ruby_system.network.master # Connect the TCC Dir controller to the ruby network tccdir_cntrl.requestFromTCP = MessageBuffer(ordered = True) tccdir_cntrl.requestFromTCP.slave = ruby_system.network.master tccdir_cntrl.responseFromTCP = MessageBuffer(ordered = True) tccdir_cntrl.responseFromTCP.slave = ruby_system.network.master tccdir_cntrl.unblockFromTCP = MessageBuffer(ordered = True) tccdir_cntrl.unblockFromTCP.slave = ruby_system.network.master tccdir_cntrl.probeToCore = MessageBuffer(ordered = True) tccdir_cntrl.probeToCore.master = ruby_system.network.slave tccdir_cntrl.responseToCore = MessageBuffer(ordered = True) tccdir_cntrl.responseToCore.master = ruby_system.network.slave tccdir_cntrl.probeFromNB = MessageBuffer() tccdir_cntrl.probeFromNB.slave = ruby_system.network.master tccdir_cntrl.responseFromNB = MessageBuffer() tccdir_cntrl.responseFromNB.slave = ruby_system.network.master tccdir_cntrl.requestToNB = MessageBuffer() tccdir_cntrl.requestToNB.master = ruby_system.network.slave tccdir_cntrl.responseToNB = MessageBuffer() tccdir_cntrl.responseToNB.master = ruby_system.network.slave tccdir_cntrl.unblockToNB = MessageBuffer() tccdir_cntrl.unblockToNB.master = ruby_system.network.slave tccdir_cntrl.triggerQueue = MessageBuffer(ordered = True) # TCC cntrls added to the GPU cluster gpuCluster.add(tcc_cntrl) gpuCluster.add(tccdir_cntrl) # Assuming no DMA devices assert(len(dma_devices) == 0) # Add cpu/gpu clusters to main cluster mainCluster.add(cpuCluster) mainCluster.add(gpuCluster) ruby_system.network.number_of_virtual_networks = 10 return (cpu_sequencers, dir_cntrl_nodes, mainCluster)
def create_system(options, full_system, system, dma_devices, ruby_system): if buildEnv['PROTOCOL'] != 'MOESI_AMD_Base': panic("This script requires the MOESI_AMD_Base protocol.") cpu_sequencers = [] # # The ruby network creation expects the list of nodes in the system to # be consistent with the NetDest list. Therefore the l1 controller # nodes must be listed before the directory nodes and directory nodes # before dma nodes, etc. # l1_cntrl_nodes = [] l3_cntrl_nodes = [] dir_cntrl_nodes = [] control_count = 0 # # Must create the individual controllers before the network to ensure # the controller constructors are called before the network constructor # # This is the base crossbar that connects the L3s, Dirs, and cpu # Cluster mainCluster = Cluster(extBW=512, intBW=512) # 1 TB/s for i in xrange(options.num_dirs): dir_cntrl = DirCntrl(TCC_select_num_bits=0) dir_cntrl.create(options, ruby_system, system) # Connect the Directory controller to the ruby network dir_cntrl.requestFromCores = MessageBuffer(ordered=True) dir_cntrl.requestFromCores.slave = ruby_system.network.master dir_cntrl.responseFromCores = MessageBuffer() dir_cntrl.responseFromCores.slave = ruby_system.network.master dir_cntrl.unblockFromCores = MessageBuffer() dir_cntrl.unblockFromCores.slave = ruby_system.network.master dir_cntrl.probeToCore = MessageBuffer() dir_cntrl.probeToCore.master = ruby_system.network.slave dir_cntrl.responseToCore = MessageBuffer() dir_cntrl.responseToCore.master = ruby_system.network.slave dir_cntrl.triggerQueue = MessageBuffer(ordered=True) dir_cntrl.L3triggerQueue = MessageBuffer(ordered=True) dir_cntrl.responseFromMemory = MessageBuffer() exec("system.dir_cntrl%d = dir_cntrl" % i) dir_cntrl_nodes.append(dir_cntrl) mainCluster.add(dir_cntrl) # Technically this config can support an odd number of cpus, but the top # level config files, such as the ruby_random_tester, will get confused if # the number of cpus does not equal the number of sequencers. Thus make # sure that an even number of cpus is specified. assert ((options.num_cpus % 2) == 0) # For an odd number of CPUs, still create the right number of controllers cpuCluster = Cluster(extBW=512, intBW=512) # 1 TB/s for i in xrange((options.num_cpus + 1) / 2): cp_cntrl = CPCntrl() cp_cntrl.create(options, ruby_system, system) exec("system.cp_cntrl%d = cp_cntrl" % i) # # Add controllers and sequencers to the appropriate lists # cpu_sequencers.extend([cp_cntrl.sequencer, cp_cntrl.sequencer1]) # Connect the CP controllers and the network cp_cntrl.requestFromCore = MessageBuffer() cp_cntrl.requestFromCore.master = ruby_system.network.slave cp_cntrl.responseFromCore = MessageBuffer() cp_cntrl.responseFromCore.master = ruby_system.network.slave cp_cntrl.unblockFromCore = MessageBuffer() cp_cntrl.unblockFromCore.master = ruby_system.network.slave cp_cntrl.probeToCore = MessageBuffer() cp_cntrl.probeToCore.slave = ruby_system.network.master cp_cntrl.responseToCore = MessageBuffer() cp_cntrl.responseToCore.slave = ruby_system.network.master cp_cntrl.mandatoryQueue = MessageBuffer() cp_cntrl.triggerQueue = MessageBuffer(ordered=True) cpuCluster.add(cp_cntrl) # Assuming no DMA devices assert (len(dma_devices) == 0) # Add cpu/gpu clusters to main cluster mainCluster.add(cpuCluster) ruby_system.network.number_of_virtual_networks = 10 return (cpu_sequencers, dir_cntrl_nodes, mainCluster)
def create_system(options, full_system, system, dma_devices, bootmem, ruby_system): if buildEnv['PROTOCOL'] != 'GPU_VIPER_Region': panic("This script requires the GPU_VIPER_Region protocol to be built.") cpu_sequencers = [] # # The ruby network creation expects the list of nodes in the system to be # consistent with the NetDest list. Therefore the l1 controller nodes # must be listed before the directory nodes and directory nodes before # dma nodes, etc. # dir_cntrl_nodes = [] # For an odd number of CPUs, still create the right number of controllers TCC_bits = int(math.log(options.num_tccs, 2)) # # Must create the individual controllers before the network to ensure the # controller constructors are called before the network constructor # # For an odd number of CPUs, still create the right number of controllers crossbar_bw = 16 * options.num_compute_units #Assuming a 2GHz clock cpuCluster = Cluster(extBW = (crossbar_bw), intBW=crossbar_bw) for i in xrange((options.num_cpus + 1) / 2): cp_cntrl = CPCntrl() cp_cntrl.create(options, ruby_system, system) rb_cntrl = RBCntrl() rb_cntrl.create(options, ruby_system, system) rb_cntrl.number_of_TBEs = 256 rb_cntrl.isOnCPU = True cp_cntrl.regionBufferNum = rb_cntrl.version exec("system.cp_cntrl%d = cp_cntrl" % i) exec("system.rb_cntrl%d = rb_cntrl" % i) # # Add controllers and sequencers to the appropriate lists # cpu_sequencers.extend([cp_cntrl.sequencer, cp_cntrl.sequencer1]) # Connect the CP controllers and the network cp_cntrl.requestFromCore = MessageBuffer() cp_cntrl.requestFromCore.master = ruby_system.network.slave cp_cntrl.responseFromCore = MessageBuffer() cp_cntrl.responseFromCore.master = ruby_system.network.slave cp_cntrl.unblockFromCore = MessageBuffer() cp_cntrl.unblockFromCore.master = ruby_system.network.slave cp_cntrl.probeToCore = MessageBuffer() cp_cntrl.probeToCore.slave = ruby_system.network.master cp_cntrl.responseToCore = MessageBuffer() cp_cntrl.responseToCore.slave = ruby_system.network.master cp_cntrl.mandatoryQueue = MessageBuffer() cp_cntrl.triggerQueue = MessageBuffer(ordered = True) # Connect the RB controllers to the ruby network rb_cntrl.requestFromCore = MessageBuffer(ordered = True) rb_cntrl.requestFromCore.slave = ruby_system.network.master rb_cntrl.responseFromCore = MessageBuffer() rb_cntrl.responseFromCore.slave = ruby_system.network.master rb_cntrl.requestToNetwork = MessageBuffer() rb_cntrl.requestToNetwork.master = ruby_system.network.slave rb_cntrl.notifyFromRegionDir = MessageBuffer() rb_cntrl.notifyFromRegionDir.slave = ruby_system.network.master rb_cntrl.probeFromRegionDir = MessageBuffer() rb_cntrl.probeFromRegionDir.slave = ruby_system.network.master rb_cntrl.unblockFromDir = MessageBuffer() rb_cntrl.unblockFromDir.slave = ruby_system.network.master rb_cntrl.responseToRegDir = MessageBuffer() rb_cntrl.responseToRegDir.master = ruby_system.network.slave rb_cntrl.triggerQueue = MessageBuffer(ordered = True) cpuCluster.add(cp_cntrl) cpuCluster.add(rb_cntrl) gpuCluster = Cluster(extBW = (crossbar_bw), intBW = crossbar_bw) for i in xrange(options.num_compute_units): tcp_cntrl = TCPCntrl(TCC_select_num_bits = TCC_bits, issue_latency = 1, number_of_TBEs = 2560) # TBEs set to max outstanding requests tcp_cntrl.create(options, ruby_system, system) tcp_cntrl.WB = options.WB_L1 tcp_cntrl.disableL1 = False exec("system.tcp_cntrl%d = tcp_cntrl" % i) # # Add controllers and sequencers to the appropriate lists # cpu_sequencers.append(tcp_cntrl.coalescer) # Connect the CP (TCP) controllers to the ruby network tcp_cntrl.requestFromTCP = MessageBuffer(ordered = True) tcp_cntrl.requestFromTCP.master = ruby_system.network.slave tcp_cntrl.responseFromTCP = MessageBuffer(ordered = True) tcp_cntrl.responseFromTCP.master = ruby_system.network.slave tcp_cntrl.unblockFromCore = MessageBuffer() tcp_cntrl.unblockFromCore.master = ruby_system.network.slave tcp_cntrl.probeToTCP = MessageBuffer(ordered = True) tcp_cntrl.probeToTCP.slave = ruby_system.network.master tcp_cntrl.responseToTCP = MessageBuffer(ordered = True) tcp_cntrl.responseToTCP.slave = ruby_system.network.master tcp_cntrl.mandatoryQueue = MessageBuffer() gpuCluster.add(tcp_cntrl) for i in xrange(options.num_sqc): sqc_cntrl = SQCCntrl(TCC_select_num_bits = TCC_bits) sqc_cntrl.create(options, ruby_system, system) exec("system.sqc_cntrl%d = sqc_cntrl" % i) # # Add controllers and sequencers to the appropriate lists # cpu_sequencers.append(sqc_cntrl.sequencer) # Connect the SQC controller to the ruby network sqc_cntrl.requestFromSQC = MessageBuffer(ordered = True) sqc_cntrl.requestFromSQC.master = ruby_system.network.slave sqc_cntrl.probeToSQC = MessageBuffer(ordered = True) sqc_cntrl.probeToSQC.slave = ruby_system.network.master sqc_cntrl.responseToSQC = MessageBuffer(ordered = True) sqc_cntrl.responseToSQC.slave = ruby_system.network.master sqc_cntrl.mandatoryQueue = MessageBuffer() # SQC also in GPU cluster gpuCluster.add(sqc_cntrl) numa_bit = 6 for i in xrange(options.num_tccs): tcc_cntrl = TCCCntrl() tcc_cntrl.create(options, ruby_system, system) tcc_cntrl.l2_request_latency = 1 tcc_cntrl.l2_response_latency = options.TCC_latency tcc_cntrl.WB = options.WB_L2 tcc_cntrl.number_of_TBEs = 2560 * options.num_compute_units # Connect the TCC controllers to the ruby network tcc_cntrl.requestFromTCP = MessageBuffer(ordered = True) tcc_cntrl.requestFromTCP.slave = ruby_system.network.master tcc_cntrl.responseToCore = MessageBuffer(ordered = True) tcc_cntrl.responseToCore.master = ruby_system.network.slave tcc_cntrl.probeFromNB = MessageBuffer() tcc_cntrl.probeFromNB.slave = ruby_system.network.master tcc_cntrl.responseFromNB = MessageBuffer() tcc_cntrl.responseFromNB.slave = ruby_system.network.master tcc_cntrl.requestToNB = MessageBuffer(ordered = True) tcc_cntrl.requestToNB.master = ruby_system.network.slave tcc_cntrl.responseToNB = MessageBuffer() tcc_cntrl.responseToNB.master = ruby_system.network.slave tcc_cntrl.unblockToNB = MessageBuffer() tcc_cntrl.unblockToNB.master = ruby_system.network.slave tcc_cntrl.triggerQueue = MessageBuffer(ordered = True) rb_cntrl = RBCntrl() rb_cntrl.create(options, ruby_system, system) rb_cntrl.number_of_TBEs = 2560 * options.num_compute_units rb_cntrl.isOnCPU = False # Connect the RB controllers to the ruby network rb_cntrl.requestFromCore = MessageBuffer(ordered = True) rb_cntrl.requestFromCore.slave = ruby_system.network.master rb_cntrl.responseFromCore = MessageBuffer() rb_cntrl.responseFromCore.slave = ruby_system.network.master rb_cntrl.requestToNetwork = MessageBuffer() rb_cntrl.requestToNetwork.master = ruby_system.network.slave rb_cntrl.notifyFromRegionDir = MessageBuffer() rb_cntrl.notifyFromRegionDir.slave = ruby_system.network.master rb_cntrl.probeFromRegionDir = MessageBuffer() rb_cntrl.probeFromRegionDir.slave = ruby_system.network.master rb_cntrl.unblockFromDir = MessageBuffer() rb_cntrl.unblockFromDir.slave = ruby_system.network.master rb_cntrl.responseToRegDir = MessageBuffer() rb_cntrl.responseToRegDir.master = ruby_system.network.slave rb_cntrl.triggerQueue = MessageBuffer(ordered = True) tcc_cntrl.regionBufferNum = rb_cntrl.version exec("system.tcc_cntrl%d = tcc_cntrl" % i) exec("system.tcc_rb_cntrl%d = rb_cntrl" % i) # TCC cntrls added to the GPU cluster gpuCluster.add(tcc_cntrl) gpuCluster.add(rb_cntrl) # Because of wire buffers, num_l3caches must equal num_dirs # Region coherence only works with 1 dir assert(options.num_l3caches == options.num_dirs == 1) # This is the base crossbar that connects the L3s, Dirs, and cpu/gpu # Clusters mainCluster = Cluster(intBW = crossbar_bw) dir_cntrl = DirCntrl() dir_cntrl.create(options, ruby_system, system) dir_cntrl.number_of_TBEs = 2560 * options.num_compute_units dir_cntrl.useL3OnWT = options.use_L3_on_WT # Connect the Directory controller to the ruby network dir_cntrl.requestFromCores = MessageBuffer() dir_cntrl.requestFromCores.slave = ruby_system.network.master dir_cntrl.responseFromCores = MessageBuffer() dir_cntrl.responseFromCores.slave = ruby_system.network.master dir_cntrl.unblockFromCores = MessageBuffer() dir_cntrl.unblockFromCores.slave = ruby_system.network.master dir_cntrl.probeToCore = MessageBuffer() dir_cntrl.probeToCore.master = ruby_system.network.slave dir_cntrl.responseToCore = MessageBuffer() dir_cntrl.responseToCore.master = ruby_system.network.slave dir_cntrl.reqFromRegBuf = MessageBuffer() dir_cntrl.reqFromRegBuf.slave = ruby_system.network.master dir_cntrl.reqToRegDir = MessageBuffer(ordered = True) dir_cntrl.reqToRegDir.master = ruby_system.network.slave dir_cntrl.reqFromRegDir = MessageBuffer(ordered = True) dir_cntrl.reqFromRegDir.slave = ruby_system.network.master dir_cntrl.unblockToRegDir = MessageBuffer() dir_cntrl.unblockToRegDir.master = ruby_system.network.slave dir_cntrl.triggerQueue = MessageBuffer(ordered = True) dir_cntrl.L3triggerQueue = MessageBuffer(ordered = True) dir_cntrl.responseFromMemory = MessageBuffer() exec("system.dir_cntrl%d = dir_cntrl" % i) dir_cntrl_nodes.append(dir_cntrl) mainCluster.add(dir_cntrl) reg_cntrl = RegionCntrl(noTCCdir=True,TCC_select_num_bits = TCC_bits) reg_cntrl.create(options, ruby_system, system) reg_cntrl.number_of_TBEs = options.num_tbes reg_cntrl.cpuRegionBufferNum = system.rb_cntrl0.version reg_cntrl.gpuRegionBufferNum = system.tcc_rb_cntrl0.version # Connect the Region Dir controllers to the ruby network reg_cntrl.requestToDir = MessageBuffer(ordered = True) reg_cntrl.requestToDir.master = ruby_system.network.slave reg_cntrl.notifyToRBuffer = MessageBuffer() reg_cntrl.notifyToRBuffer.master = ruby_system.network.slave reg_cntrl.probeToRBuffer = MessageBuffer() reg_cntrl.probeToRBuffer.master = ruby_system.network.slave reg_cntrl.responseFromRBuffer = MessageBuffer() reg_cntrl.responseFromRBuffer.slave = ruby_system.network.master reg_cntrl.requestFromRegBuf = MessageBuffer() reg_cntrl.requestFromRegBuf.slave = ruby_system.network.master reg_cntrl.triggerQueue = MessageBuffer(ordered = True) exec("system.reg_cntrl%d = reg_cntrl" % i) mainCluster.add(reg_cntrl) # Assuming no DMA devices assert(len(dma_devices) == 0) # Add cpu/gpu clusters to main cluster mainCluster.add(cpuCluster) mainCluster.add(gpuCluster) ruby_system.network.number_of_virtual_networks = 10 return (cpu_sequencers, dir_cntrl_nodes, mainCluster)
def create_system(options, full_system, system, dma_devices, ruby_system): if buildEnv['PROTOCOL'] != 'MOESI_AMD_Base': panic("This script requires the MOESI_AMD_Base protocol.") cpu_sequencers = [] # # The ruby network creation expects the list of nodes in the system to # be consistent with the NetDest list. Therefore the l1 controller # nodes must be listed before the directory nodes and directory nodes # before dma nodes, etc. # l1_cntrl_nodes = [] l3_cntrl_nodes = [] dir_cntrl_nodes = [] control_count = 0 # # Must create the individual controllers before the network to ensure # the controller constructors are called before the network constructor # # This is the base crossbar that connects the L3s, Dirs, and cpu # Cluster mainCluster = Cluster(extBW = 512, intBW = 512) # 1 TB/s for i in xrange(options.num_dirs): dir_cntrl = DirCntrl(TCC_select_num_bits = 0) dir_cntrl.create(options, ruby_system, system) # Connect the Directory controller to the ruby network dir_cntrl.requestFromCores = MessageBuffer(ordered = True) dir_cntrl.requestFromCores.slave = ruby_system.network.master dir_cntrl.responseFromCores = MessageBuffer() dir_cntrl.responseFromCores.slave = ruby_system.network.master dir_cntrl.unblockFromCores = MessageBuffer() dir_cntrl.unblockFromCores.slave = ruby_system.network.master dir_cntrl.probeToCore = MessageBuffer() dir_cntrl.probeToCore.master = ruby_system.network.slave dir_cntrl.responseToCore = MessageBuffer() dir_cntrl.responseToCore.master = ruby_system.network.slave dir_cntrl.triggerQueue = MessageBuffer(ordered = True) dir_cntrl.L3triggerQueue = MessageBuffer(ordered = True) dir_cntrl.responseFromMemory = MessageBuffer() exec("system.dir_cntrl%d = dir_cntrl" % i) dir_cntrl_nodes.append(dir_cntrl) mainCluster.add(dir_cntrl) # Technically this config can support an odd number of cpus, but the top # level config files, such as the ruby_random_tester, will get confused if # the number of cpus does not equal the number of sequencers. Thus make # sure that an even number of cpus is specified. assert((options.num_cpus % 2) == 0) # For an odd number of CPUs, still create the right number of controllers cpuCluster = Cluster(extBW = 512, intBW = 512) # 1 TB/s for i in xrange((options.num_cpus + 1) / 2): cp_cntrl = CPCntrl() cp_cntrl.create(options, ruby_system, system) exec("system.cp_cntrl%d = cp_cntrl" % i) # # Add controllers and sequencers to the appropriate lists # cpu_sequencers.extend([cp_cntrl.sequencer, cp_cntrl.sequencer1]) # Connect the CP controllers and the network cp_cntrl.requestFromCore = MessageBuffer() cp_cntrl.requestFromCore.master = ruby_system.network.slave cp_cntrl.responseFromCore = MessageBuffer() cp_cntrl.responseFromCore.master = ruby_system.network.slave cp_cntrl.unblockFromCore = MessageBuffer() cp_cntrl.unblockFromCore.master = ruby_system.network.slave cp_cntrl.probeToCore = MessageBuffer() cp_cntrl.probeToCore.slave = ruby_system.network.master cp_cntrl.responseToCore = MessageBuffer() cp_cntrl.responseToCore.slave = ruby_system.network.master cp_cntrl.mandatoryQueue = MessageBuffer() cp_cntrl.triggerQueue = MessageBuffer(ordered = True) cpuCluster.add(cp_cntrl) # Assuming no DMA devices assert(len(dma_devices) == 0) # Add cpu/gpu clusters to main cluster mainCluster.add(cpuCluster) ruby_system.network.number_of_virtual_networks = 10 return (cpu_sequencers, dir_cntrl_nodes, mainCluster)