# Copyright (c) 2011-2015 Advanced Micro Devices, Inc. # All rights reserved. # # For use for simulation and test purposes only # # Redistribution and use in source and binary forms, with or without # modification, are permitted provided that the following conditions are met: # # 1. Redistributions of source code must retain the above copyright notice, # this list of conditions and the following disclaimer. # # 2. Redistributions in binary form must reproduce the above copyright notice, # this list of conditions and the following disclaimer in the documentation # and/or other materials provided with the distribution. # # 3. Neither the name of the copyright holder nor the names of its # contributors may be used to endorse or promote products derived from this # software without specific prior written permission. # # THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" # AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE # IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE # ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT HOLDER OR CONTRIBUTORS BE # LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR # CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF # SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS # INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN # CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) # ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE # POSSIBILITY OF SUCH DAMAGE. # # Authors: Lisa Hsu import math import m5 from m5.objects import * from m5.defines import buildEnv from m5.util import addToPath from Ruby import create_topology from Ruby import send_evicts addToPath('../') from topologies.Cluster import Cluster from topologies.Crossbar import Crossbar class CntrlBase: _seqs = 0 @classmethod def seqCount(cls): # Use SeqCount not class since we need global count CntrlBase._seqs += 1 return CntrlBase._seqs - 1 _cntrls = 0 @classmethod def cntrlCount(cls): # Use CntlCount not class since we need global count CntrlBase._cntrls += 1 return CntrlBase._cntrls - 1 _version = 0 @classmethod def versionCount(cls): cls._version += 1 # Use count for this particular type return cls._version - 1 class L1Cache(RubyCache): resourceStalls = False dataArrayBanks = 2 tagArrayBanks = 2 dataAccessLatency = 1 tagAccessLatency = 1 def create(self, size, assoc, options): self.size = MemorySize(size) self.assoc = assoc self.replacement_policy = PseudoLRUReplacementPolicy() class L2Cache(RubyCache): resourceStalls = False assoc = 16 dataArrayBanks = 16 tagArrayBanks = 16 def create(self, size, assoc, options): self.size = MemorySize(size) self.assoc = assoc self.replacement_policy = PseudoLRUReplacementPolicy() class CPCntrl(CorePair_Controller, CntrlBase): def create(self, options, ruby_system, system): self.version = self.versionCount() self.L1Icache = L1Cache() self.L1Icache.create(options.l1i_size, options.l1i_assoc, options) self.L1D0cache = L1Cache() self.L1D0cache.create(options.l1d_size, options.l1d_assoc, options) self.L1D1cache = L1Cache() self.L1D1cache.create(options.l1d_size, options.l1d_assoc, options) self.L2cache = L2Cache() self.L2cache.create(options.l2_size, options.l2_assoc, options) self.sequencer = RubySequencer() self.sequencer.version = self.seqCount() self.sequencer.icache = self.L1Icache self.sequencer.dcache = self.L1D0cache self.sequencer.ruby_system = ruby_system self.sequencer.coreid = 0 self.sequencer.is_cpu_sequencer = True self.sequencer1 = RubySequencer() self.sequencer1.version = self.seqCount() self.sequencer1.icache = self.L1Icache self.sequencer1.dcache = self.L1D1cache self.sequencer1.ruby_system = ruby_system self.sequencer1.coreid = 1 self.sequencer1.is_cpu_sequencer = True self.issue_latency = options.cpu_to_dir_latency self.send_evictions = send_evicts(options) self.ruby_system = ruby_system if options.recycle_latency: self.recycle_latency = options.recycle_latency class TCPCache(RubyCache): size = "16kB" assoc = 16 dataArrayBanks = 16 #number of data banks tagArrayBanks = 16 #number of tag banks dataAccessLatency = 4 tagAccessLatency = 1 def create(self, options): self.size = MemorySize(options.tcp_size) self.assoc = options.tcp_assoc self.resourceStalls = options.no_tcc_resource_stalls self.replacement_policy = PseudoLRUReplacementPolicy() class TCPCntrl(TCP_Controller, CntrlBase): def create(self, options, ruby_system, system): self.version = self.versionCount() self.L1cache = TCPCache(tagAccessLatency = options.TCP_latency, dataAccessLatency = options.TCP_latency) self.L1cache.resourceStalls = options.no_resource_stalls self.L1cache.create(options) self.issue_latency = 1 self.coalescer = VIPERCoalescer() self.coalescer.version = self.seqCount() self.coalescer.icache = self.L1cache self.coalescer.dcache = self.L1cache self.coalescer.ruby_system = ruby_system self.coalescer.support_inst_reqs = False self.coalescer.is_cpu_sequencer = False self.sequencer = RubySequencer() self.sequencer.version = self.seqCount() self.sequencer.icache = self.L1cache self.sequencer.dcache = self.L1cache self.sequencer.ruby_system = ruby_system self.sequencer.is_cpu_sequencer = True self.use_seq_not_coal = False self.ruby_system = ruby_system if options.recycle_latency: self.recycle_latency = options.recycle_latency def createCP(self, options, ruby_system, system): self.version = self.versionCount() self.L1cache = TCPCache(tagAccessLatency = options.TCP_latency, dataAccessLatency = options.TCP_latency) self.L1cache.resourceStalls = options.no_resource_stalls self.L1cache.create(options) self.issue_latency = 1 self.coalescer = VIPERCoalescer() self.coalescer.version = self.seqCount() self.coalescer.icache = self.L1cache self.coalescer.dcache = self.L1cache self.coalescer.ruby_system = ruby_system self.coalescer.support_inst_reqs = False self.coalescer.is_cpu_sequencer = False self.sequencer = RubySequencer() self.sequencer.version = self.seqCount() self.sequencer.icache = self.L1cache self.sequencer.dcache = self.L1cache self.sequencer.ruby_system = ruby_system self.sequencer.is_cpu_sequencer = True self.use_seq_not_coal = True self.ruby_system = ruby_system if options.recycle_latency: self.recycle_latency = options.recycle_latency class SQCCache(RubyCache): dataArrayBanks = 8 tagArrayBanks = 8 dataAccessLatency = 1 tagAccessLatency = 1 def create(self, options): self.size = MemorySize(options.sqc_size) self.assoc = options.sqc_assoc self.replacement_policy = PseudoLRUReplacementPolicy() class SQCCntrl(SQC_Controller, CntrlBase): def create(self, options, ruby_system, system): self.version = self.versionCount() self.L1cache = SQCCache() self.L1cache.create(options) self.L1cache.resourceStalls = options.no_resource_stalls self.sequencer = RubySequencer() self.sequencer.version = self.seqCount() self.sequencer.icache = self.L1cache self.sequencer.dcache = self.L1cache self.sequencer.ruby_system = ruby_system self.sequencer.support_data_reqs = False self.sequencer.is_cpu_sequencer = False self.ruby_system = ruby_system if options.recycle_latency: self.recycle_latency = options.recycle_latency class TCC(RubyCache): size = MemorySize("256kB") assoc = 16 dataAccessLatency = 8 tagAccessLatency = 2 resourceStalls = True def create(self, options): self.assoc = options.tcc_assoc if hasattr(options, 'bw_scalor') and options.bw_scalor > 0: s = options.num_compute_units tcc_size = s * 128 tcc_size = str(tcc_size)+'kB' self.size = MemorySize(tcc_size) self.dataArrayBanks = 64 self.tagArrayBanks = 64 else: self.size = MemorySize(options.tcc_size) self.dataArrayBanks = 256 / options.num_tccs #number of data banks self.tagArrayBanks = 256 / options.num_tccs #number of tag banks self.size.value = self.size.value / options.num_tccs if ((self.size.value / long(self.assoc)) < 128): self.size.value = long(128 * self.assoc) self.start_index_bit = math.log(options.cacheline_size, 2) + \ math.log(options.num_tccs, 2) self.replacement_policy = PseudoLRUReplacementPolicy() class TCCCntrl(TCC_Controller, CntrlBase): def create(self, options, ruby_system, system): self.version = self.versionCount() self.L2cache = TCC() self.L2cache.create(options) self.L2cache.resourceStalls = options.no_tcc_resource_stalls self.ruby_system = ruby_system if options.recycle_latency: self.recycle_latency = options.recycle_latency class L3Cache(RubyCache): dataArrayBanks = 16 tagArrayBanks = 16 def create(self, options, ruby_system, system): self.size = MemorySize(options.l3_size) self.size.value /= options.num_dirs self.assoc = options.l3_assoc self.dataArrayBanks /= options.num_dirs self.tagArrayBanks /= options.num_dirs self.dataArrayBanks /= options.num_dirs self.tagArrayBanks /= options.num_dirs self.dataAccessLatency = options.l3_data_latency self.tagAccessLatency = options.l3_tag_latency self.resourceStalls = False self.replacement_policy = PseudoLRUReplacementPolicy() class L3Cntrl(L3Cache_Controller, CntrlBase): def create(self, options, ruby_system, system): self.version = self.versionCount() self.L3cache = L3Cache() self.L3cache.create(options, ruby_system, system) self.l3_response_latency = max(self.L3cache.dataAccessLatency, self.L3cache.tagAccessLatency) self.ruby_system = ruby_system if options.recycle_latency: self.recycle_latency = options.recycle_latency def connectWireBuffers(self, req_to_dir, resp_to_dir, l3_unblock_to_dir, req_to_l3, probe_to_l3, resp_to_l3): self.reqToDir = req_to_dir self.respToDir = resp_to_dir self.l3UnblockToDir = l3_unblock_to_dir self.reqToL3 = req_to_l3 self.probeToL3 = probe_to_l3 self.respToL3 = resp_to_l3 class DirMem(RubyDirectoryMemory, CntrlBase): def create(self, options, ruby_system, system): self.version = self.versionCount() phys_mem_size = AddrRange(options.mem_size).size() mem_module_size = phys_mem_size / options.num_dirs dir_size = MemorySize('0B') dir_size.value = mem_module_size self.size = dir_size class DirCntrl(Directory_Controller, CntrlBase): def create(self, options, ruby_system, system): self.version = self.versionCount() self.response_latency = 30 self.directory = DirMem() self.directory.create(options, ruby_system, system) self.L3CacheMemory = L3Cache() self.L3CacheMemory.create(options, ruby_system, system) self.l3_hit_latency = max(self.L3CacheMemory.dataAccessLatency, self.L3CacheMemory.tagAccessLatency) self.number_of_TBEs = options.num_tbes self.ruby_system = ruby_system if options.recycle_latency: self.recycle_latency = options.recycle_latency def connectWireBuffers(self, req_to_dir, resp_to_dir, l3_unblock_to_dir, req_to_l3, probe_to_l3, resp_to_l3): self.reqToDir = req_to_dir self.respToDir = resp_to_dir self.l3UnblockToDir = l3_unblock_to_dir self.reqToL3 = req_to_l3 self.probeToL3 = probe_to_l3 self.respToL3 = resp_to_l3 def define_options(parser): parser.add_option("--num-subcaches", type = "int", default = 4) parser.add_option("--l3-data-latency", type = "int", default = 20) parser.add_option("--l3-tag-latency", type = "int", default = 15) parser.add_option("--cpu-to-dir-latency", type = "int", default = 120) parser.add_option("--gpu-to-dir-latency", type = "int", default = 120) parser.add_option("--no-resource-stalls", action = "store_false", default = True) parser.add_option("--no-tcc-resource-stalls", action = "store_false", default = True) parser.add_option("--use-L3-on-WT", action = "store_true", default = False) parser.add_option("--num-tbes", type = "int", default = 256) parser.add_option("--l2-latency", type = "int", default = 50) # load to use parser.add_option("--num-tccs", type = "int", default = 1, help = "number of TCC banks in the GPU") parser.add_option("--sqc-size", type = 'string', default = '32kB', help = "SQC cache size") parser.add_option("--sqc-assoc", type = 'int', default = 8, help = "SQC cache assoc") parser.add_option("--WB_L1", action = "store_true", default = False, help = "writeback L1") parser.add_option("--WB_L2", action = "store_true", default = False, help = "writeback L2") parser.add_option("--TCP_latency", type = "int", default = 4, help = "TCP latency") parser.add_option("--TCC_latency", type = "int", default = 16, help = "TCC latency") parser.add_option("--tcc-size", type = 'string', default = '256kB', help = "agregate tcc size") parser.add_option("--tcc-assoc", type = 'int', default = 16, help = "tcc assoc") parser.add_option("--tcp-size", type = 'string', default = '16kB', help = "tcp size") parser.add_option("--tcp-assoc", type = 'int', default = 16, help = "tcp assoc") parser.add_option("--noL1", action = "store_true", default = False, help = "bypassL1") 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 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 # 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 xrange((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) 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 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 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 xrange(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_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 xrange(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)