33 34import math 35import m5 36from m5.objects import * 37from m5.defines import buildEnv 38from Ruby import create_topology 39from Ruby import send_evicts 40 41from topologies.Cluster import Cluster 42from topologies.Crossbar import Crossbar 43 44class CntrlBase: 45 _seqs = 0 46 @classmethod 47 def seqCount(cls): 48 # Use SeqCount not class since we need global count 49 CntrlBase._seqs += 1 50 return CntrlBase._seqs - 1 51 52 _cntrls = 0 53 @classmethod 54 def cntrlCount(cls): 55 # Use CntlCount not class since we need global count 56 CntrlBase._cntrls += 1 57 return CntrlBase._cntrls - 1 58 59 _version = 0 60 @classmethod 61 def versionCount(cls): 62 cls._version += 1 # Use count for this particular type 63 return cls._version - 1 64 65class L1Cache(RubyCache): 66 resourceStalls = False 67 dataArrayBanks = 2 68 tagArrayBanks = 2 69 dataAccessLatency = 1 70 tagAccessLatency = 1 71 def create(self, size, assoc, options): 72 self.size = MemorySize(size) 73 self.assoc = assoc 74 self.replacement_policy = PseudoLRUReplacementPolicy() 75 76class L2Cache(RubyCache): 77 resourceStalls = False 78 assoc = 16 79 dataArrayBanks = 16 80 tagArrayBanks = 16 81 def create(self, size, assoc, options): 82 self.size = MemorySize(size) 83 self.assoc = assoc 84 self.replacement_policy = PseudoLRUReplacementPolicy() 85 86class CPCntrl(CorePair_Controller, CntrlBase): 87 88 def create(self, options, ruby_system, system): 89 self.version = self.versionCount() 90 91 self.L1Icache = L1Cache() 92 self.L1Icache.create(options.l1i_size, options.l1i_assoc, options) 93 self.L1D0cache = L1Cache() 94 self.L1D0cache.create(options.l1d_size, options.l1d_assoc, options) 95 self.L1D1cache = L1Cache() 96 self.L1D1cache.create(options.l1d_size, options.l1d_assoc, options) 97 self.L2cache = L2Cache() 98 self.L2cache.create(options.l2_size, options.l2_assoc, options) 99 100 self.sequencer = RubySequencer() 101 self.sequencer.version = self.seqCount() 102 self.sequencer.icache = self.L1Icache 103 self.sequencer.dcache = self.L1D0cache 104 self.sequencer.ruby_system = ruby_system 105 self.sequencer.coreid = 0 106 self.sequencer.is_cpu_sequencer = True 107 108 self.sequencer1 = RubySequencer() 109 self.sequencer1.version = self.seqCount() 110 self.sequencer1.icache = self.L1Icache 111 self.sequencer1.dcache = self.L1D1cache 112 self.sequencer1.ruby_system = ruby_system 113 self.sequencer1.coreid = 1 114 self.sequencer1.is_cpu_sequencer = True 115 116 self.issue_latency = options.cpu_to_dir_latency 117 self.send_evictions = send_evicts(options) 118 119 self.ruby_system = ruby_system 120 121 if options.recycle_latency: 122 self.recycle_latency = options.recycle_latency 123 124class TCPCache(RubyCache): 125 size = "16kB" 126 assoc = 16 127 dataArrayBanks = 16 128 tagArrayBanks = 16 129 dataAccessLatency = 4 130 tagAccessLatency = 1 131 def create(self, options): 132 self.size = MemorySize(options.tcp_size) 133 self.dataArrayBanks = 16 134 self.tagArrayBanks = 16 135 self.dataAccessLatency = 4 136 self.tagAccessLatency = 1 137 self.resourceStalls = options.no_tcc_resource_stalls 138 self.replacement_policy = PseudoLRUReplacementPolicy() 139 140class TCPCntrl(TCP_Controller, CntrlBase): 141 142 def create(self, options, ruby_system, system): 143 self.version = self.versionCount() 144 self.L1cache = TCPCache() 145 self.L1cache.create(options) 146 self.issue_latency = 1 147 148 self.coalescer = VIPERCoalescer() 149 self.coalescer.version = self.seqCount() 150 self.coalescer.icache = self.L1cache 151 self.coalescer.dcache = self.L1cache 152 self.coalescer.ruby_system = ruby_system 153 self.coalescer.support_inst_reqs = False 154 self.coalescer.is_cpu_sequencer = False 155 156 self.sequencer = RubySequencer() 157 self.sequencer.version = self.seqCount() 158 self.sequencer.icache = self.L1cache 159 self.sequencer.dcache = self.L1cache 160 self.sequencer.ruby_system = ruby_system 161 self.sequencer.is_cpu_sequencer = True 162 163 self.use_seq_not_coal = False 164 165 self.ruby_system = ruby_system 166 if options.recycle_latency: 167 self.recycle_latency = options.recycle_latency 168 169class SQCCache(RubyCache): 170 dataArrayBanks = 8 171 tagArrayBanks = 8 172 dataAccessLatency = 1 173 tagAccessLatency = 1 174 175 def create(self, options): 176 self.size = MemorySize(options.sqc_size) 177 self.assoc = options.sqc_assoc 178 self.replacement_policy = PseudoLRUReplacementPolicy() 179 180class SQCCntrl(SQC_Controller, CntrlBase): 181 182 def create(self, options, ruby_system, system): 183 self.version = self.versionCount() 184 self.L1cache = SQCCache() 185 self.L1cache.create(options) 186 self.L1cache.resourceStalls = False 187 self.sequencer = RubySequencer() 188 self.sequencer.version = self.seqCount() 189 self.sequencer.icache = self.L1cache 190 self.sequencer.dcache = self.L1cache 191 self.sequencer.ruby_system = ruby_system 192 self.sequencer.support_data_reqs = False 193 self.sequencer.is_cpu_sequencer = False 194 self.ruby_system = ruby_system 195 if options.recycle_latency: 196 self.recycle_latency = options.recycle_latency 197 198class TCC(RubyCache): 199 size = MemorySize("256kB") 200 assoc = 16 201 dataAccessLatency = 8 202 tagAccessLatency = 2 203 resourceStalls = True 204 def create(self, options): 205 self.assoc = options.tcc_assoc 206 if hasattr(options, 'bw_scalor') and options.bw_scalor > 0: 207 s = options.num_compute_units 208 tcc_size = s * 128 209 tcc_size = str(tcc_size)+'kB' 210 self.size = MemorySize(tcc_size) 211 self.dataArrayBanks = 64 212 self.tagArrayBanks = 64 213 else: 214 self.size = MemorySize(options.tcc_size) 215 self.dataArrayBanks = 256 / options.num_tccs #number of data banks 216 self.tagArrayBanks = 256 / options.num_tccs #number of tag banks 217 self.size.value = self.size.value / options.num_tccs 218 if ((self.size.value / long(self.assoc)) < 128): 219 self.size.value = long(128 * self.assoc) 220 self.start_index_bit = math.log(options.cacheline_size, 2) + \ 221 math.log(options.num_tccs, 2) 222 self.replacement_policy = PseudoLRUReplacementPolicy() 223 224class TCCCntrl(TCC_Controller, CntrlBase): 225 def create(self, options, ruby_system, system): 226 self.version = self.versionCount() 227 self.L2cache = TCC() 228 self.L2cache.create(options) 229 self.ruby_system = ruby_system 230 self.L2cache.resourceStalls = options.no_tcc_resource_stalls 231 232 if options.recycle_latency: 233 self.recycle_latency = options.recycle_latency 234 235class L3Cache(RubyCache): 236 dataArrayBanks = 16 237 tagArrayBanks = 16 238 239 def create(self, options, ruby_system, system): 240 self.size = MemorySize(options.l3_size) 241 self.size.value /= options.num_dirs 242 self.assoc = options.l3_assoc 243 self.dataArrayBanks /= options.num_dirs 244 self.tagArrayBanks /= options.num_dirs 245 self.dataArrayBanks /= options.num_dirs 246 self.tagArrayBanks /= options.num_dirs 247 self.dataAccessLatency = options.l3_data_latency 248 self.tagAccessLatency = options.l3_tag_latency 249 self.resourceStalls = False 250 self.replacement_policy = PseudoLRUReplacementPolicy() 251 252class ProbeFilter(RubyCache): 253 size = "4MB" 254 assoc = 16 255 dataArrayBanks = 256 256 tagArrayBanks = 256 257 258 def create(self, options, ruby_system, system): 259 self.block_size = "%dB" % (64 * options.blocks_per_region) 260 self.size = options.region_dir_entries * \ 261 self.block_size * options.num_compute_units 262 self.assoc = 8 263 self.tagArrayBanks = 8 264 self.tagAccessLatency = options.dir_tag_latency 265 self.dataAccessLatency = 1 266 self.resourceStalls = options.no_resource_stalls 267 self.start_index_bit = 6 + int(math.log(options.blocks_per_region, 2)) 268 self.replacement_policy = PseudoLRUReplacementPolicy() 269 270class L3Cntrl(L3Cache_Controller, CntrlBase): 271 def create(self, options, ruby_system, system): 272 self.version = self.versionCount() 273 self.L3cache = L3Cache() 274 self.L3cache.create(options, ruby_system, system) 275 self.l3_response_latency = \ 276 max(self.L3cache.dataAccessLatency, self.L3cache.tagAccessLatency) 277 self.ruby_system = ruby_system 278 if options.recycle_latency: 279 self.recycle_latency = options.recycle_latency 280 281 def connectWireBuffers(self, req_to_dir, resp_to_dir, l3_unblock_to_dir, 282 req_to_l3, probe_to_l3, resp_to_l3): 283 self.reqToDir = req_to_dir 284 self.respToDir = resp_to_dir 285 self.l3UnblockToDir = l3_unblock_to_dir 286 self.reqToL3 = req_to_l3 287 self.probeToL3 = probe_to_l3 288 self.respToL3 = resp_to_l3 289 290class DirMem(RubyDirectoryMemory, CntrlBase): 291 def create(self, options, ruby_system, system): 292 self.version = self.versionCount() 293 294 phys_mem_size = AddrRange(options.mem_size).size() 295 mem_module_size = phys_mem_size / options.num_dirs 296 dir_size = MemorySize('0B') 297 dir_size.value = mem_module_size 298 self.size = dir_size 299 300class DirCntrl(Directory_Controller, CntrlBase): 301 def create(self, options, ruby_system, system): 302 self.version = self.versionCount() 303 self.response_latency = 30 304 self.directory = DirMem() 305 self.directory.create(options, ruby_system, system) 306 self.L3CacheMemory = L3Cache() 307 self.L3CacheMemory.create(options, ruby_system, system) 308 self.ProbeFilterMemory = ProbeFilter() 309 self.ProbeFilterMemory.create(options, ruby_system, system) 310 self.l3_hit_latency = \ 311 max(self.L3CacheMemory.dataAccessLatency, 312 self.L3CacheMemory.tagAccessLatency) 313 314 self.ruby_system = ruby_system 315 if options.recycle_latency: 316 self.recycle_latency = options.recycle_latency 317 318 def connectWireBuffers(self, req_to_dir, resp_to_dir, l3_unblock_to_dir, 319 req_to_l3, probe_to_l3, resp_to_l3): 320 self.reqToDir = req_to_dir 321 self.respToDir = resp_to_dir 322 self.l3UnblockToDir = l3_unblock_to_dir 323 self.reqToL3 = req_to_l3 324 self.probeToL3 = probe_to_l3 325 self.respToL3 = resp_to_l3 326 327def define_options(parser): 328 parser.add_option("--num-subcaches", type = "int", default = 4) 329 parser.add_option("--l3-data-latency", type = "int", default = 20) 330 parser.add_option("--l3-tag-latency", type = "int", default = 15) 331 parser.add_option("--cpu-to-dir-latency", type = "int", default = 120) 332 parser.add_option("--gpu-to-dir-latency", type = "int", default = 120) 333 parser.add_option("--no-resource-stalls", action = "store_false", 334 default = True) 335 parser.add_option("--no-tcc-resource-stalls", action = "store_false", 336 default = True) 337 parser.add_option("--num-tbes", type = "int", default = 2560) 338 parser.add_option("--l2-latency", type = "int", default = 50) # load to use 339 parser.add_option("--num-tccs", type = "int", default = 1, 340 help = "number of TCC banks in the GPU") 341 parser.add_option("--sqc-size", type = 'string', default = '32kB', 342 help = "SQC cache size") 343 parser.add_option("--sqc-assoc", type = 'int', default = 8, 344 help = "SQC cache assoc") 345 parser.add_option("--region-dir-entries", type = "int", default = 8192) 346 parser.add_option("--dir-tag-latency", type = "int", default = 8) 347 parser.add_option("--dir-tag-banks", type = "int", default = 4) 348 parser.add_option("--blocks-per-region", type = "int", default = 1) 349 parser.add_option("--use-L3-on-WT", action = "store_true", default = False) 350 parser.add_option("--nonInclusiveDir", action = "store_true", 351 default = False) 352 parser.add_option("--WB_L1", action = "store_true", 353 default = False, help = "writeback L2") 354 parser.add_option("--WB_L2", action = "store_true", 355 default = False, help = "writeback L2") 356 parser.add_option("--TCP_latency", type = "int", 357 default = 4, help = "TCP latency") 358 parser.add_option("--TCC_latency", type = "int", 359 default = 16, help = "TCC latency") 360 parser.add_option("--tcc-size", type = 'string', default = '2MB', 361 help = "agregate tcc size") 362 parser.add_option("--tcc-assoc", type = 'int', default = 16, 363 help = "tcc assoc") 364 parser.add_option("--tcp-size", type = 'string', default = '16kB', 365 help = "tcp size") 366 parser.add_option("--sampler-sets", type = "int", default = 1024) 367 parser.add_option("--sampler-assoc", type = "int", default = 16) 368 parser.add_option("--sampler-counter", type = "int", default = 512) 369 parser.add_option("--noL1", action = "store_true", default = False, 370 help = "bypassL1") 371 parser.add_option("--noL2", action = "store_true", default = False, 372 help = "bypassL2") 373 374def create_system(options, full_system, system, dma_devices, bootmem, 375 ruby_system): 376 if buildEnv['PROTOCOL'] != 'GPU_VIPER_Baseline': 377 panic("This script requires the" \ 378 "GPU_VIPER_Baseline protocol to be built.") 379 380 cpu_sequencers = [] 381 382 # 383 # The ruby network creation expects the list of nodes in the system to be 384 # consistent with the NetDest list. Therefore the l1 controller nodes 385 # must be listed before the directory nodes and directory nodes before 386 # dma nodes, etc. 387 # 388 cp_cntrl_nodes = [] 389 tcp_cntrl_nodes = [] 390 sqc_cntrl_nodes = [] 391 tcc_cntrl_nodes = [] 392 dir_cntrl_nodes = [] 393 l3_cntrl_nodes = [] 394 395 # 396 # Must create the individual controllers before the network to ensure the 397 # controller constructors are called before the network constructor 398 # 399 400 # For an odd number of CPUs, still create the right number of controllers 401 TCC_bits = int(math.log(options.num_tccs, 2)) 402 403 # This is the base crossbar that connects the L3s, Dirs, and cpu/gpu 404 # Clusters 405 crossbar_bw = 16 * options.num_compute_units #Assuming a 2GHz clock 406 mainCluster = Cluster(intBW = crossbar_bw) 407 for i in xrange(options.num_dirs): 408 409 dir_cntrl = DirCntrl(noTCCdir=True,TCC_select_num_bits = TCC_bits) 410 dir_cntrl.create(options, ruby_system, system) 411 dir_cntrl.number_of_TBEs = options.num_tbes 412 dir_cntrl.useL3OnWT = options.use_L3_on_WT 413 dir_cntrl.inclusiveDir = not options.nonInclusiveDir 414 415 # Connect the Directory controller to the ruby network 416 dir_cntrl.requestFromCores = MessageBuffer(ordered = True) 417 dir_cntrl.requestFromCores.slave = ruby_system.network.master 418 419 dir_cntrl.responseFromCores = MessageBuffer() 420 dir_cntrl.responseFromCores.slave = ruby_system.network.master 421 422 dir_cntrl.unblockFromCores = MessageBuffer() 423 dir_cntrl.unblockFromCores.slave = ruby_system.network.master 424 425 dir_cntrl.probeToCore = MessageBuffer() 426 dir_cntrl.probeToCore.master = ruby_system.network.slave 427 428 dir_cntrl.responseToCore = MessageBuffer() 429 dir_cntrl.responseToCore.master = ruby_system.network.slave 430 431 dir_cntrl.triggerQueue = MessageBuffer(ordered = True) 432 dir_cntrl.L3triggerQueue = MessageBuffer(ordered = True) 433 dir_cntrl.responseFromMemory = MessageBuffer() 434 435 exec("system.dir_cntrl%d = dir_cntrl" % i) 436 dir_cntrl_nodes.append(dir_cntrl) 437 mainCluster.add(dir_cntrl) 438 439 cpuCluster = Cluster(extBW = crossbar_bw, intBW=crossbar_bw) 440 for i in xrange((options.num_cpus + 1) / 2): 441 442 cp_cntrl = CPCntrl() 443 cp_cntrl.create(options, ruby_system, system) 444 445 exec("system.cp_cntrl%d = cp_cntrl" % i) 446 # 447 # Add controllers and sequencers to the appropriate lists 448 # 449 cpu_sequencers.extend([cp_cntrl.sequencer, cp_cntrl.sequencer1]) 450 451 # Connect the CP controllers and the network 452 cp_cntrl.requestFromCore = MessageBuffer() 453 cp_cntrl.requestFromCore.master = ruby_system.network.slave 454 455 cp_cntrl.responseFromCore = MessageBuffer() 456 cp_cntrl.responseFromCore.master = ruby_system.network.slave 457 458 cp_cntrl.unblockFromCore = MessageBuffer() 459 cp_cntrl.unblockFromCore.master = ruby_system.network.slave 460 461 cp_cntrl.probeToCore = MessageBuffer() 462 cp_cntrl.probeToCore.slave = ruby_system.network.master 463 464 cp_cntrl.responseToCore = MessageBuffer() 465 cp_cntrl.responseToCore.slave = ruby_system.network.master 466 467 cp_cntrl.mandatoryQueue = MessageBuffer() 468 cp_cntrl.triggerQueue = MessageBuffer(ordered = True) 469 470 cpuCluster.add(cp_cntrl) 471 472 gpuCluster = Cluster(extBW = crossbar_bw, intBW = crossbar_bw) 473 for i in xrange(options.num_compute_units): 474 475 tcp_cntrl = TCPCntrl(TCC_select_num_bits = TCC_bits, 476 issue_latency = 1, 477 number_of_TBEs = 2560) 478 # TBEs set to max outstanding requests 479 tcp_cntrl.create(options, ruby_system, system) 480 tcp_cntrl.WB = options.WB_L1 481 tcp_cntrl.disableL1 = options.noL1 482 483 exec("system.tcp_cntrl%d = tcp_cntrl" % i) 484 # 485 # Add controllers and sequencers to the appropriate lists 486 # 487 cpu_sequencers.append(tcp_cntrl.coalescer) 488 tcp_cntrl_nodes.append(tcp_cntrl) 489 490 # Connect the CP (TCP) controllers to the ruby network 491 tcp_cntrl.requestFromTCP = MessageBuffer(ordered = True) 492 tcp_cntrl.requestFromTCP.master = ruby_system.network.slave 493 494 tcp_cntrl.responseFromTCP = MessageBuffer(ordered = True) 495 tcp_cntrl.responseFromTCP.master = ruby_system.network.slave 496 497 tcp_cntrl.unblockFromCore = MessageBuffer() 498 tcp_cntrl.unblockFromCore.master = ruby_system.network.slave 499 500 tcp_cntrl.probeToTCP = MessageBuffer(ordered = True) 501 tcp_cntrl.probeToTCP.slave = ruby_system.network.master 502 503 tcp_cntrl.responseToTCP = MessageBuffer(ordered = True) 504 tcp_cntrl.responseToTCP.slave = ruby_system.network.master 505 506 tcp_cntrl.mandatoryQueue = MessageBuffer() 507 508 gpuCluster.add(tcp_cntrl) 509 510 for i in xrange(options.num_sqc): 511 512 sqc_cntrl = SQCCntrl(TCC_select_num_bits = TCC_bits) 513 sqc_cntrl.create(options, ruby_system, system) 514 515 exec("system.sqc_cntrl%d = sqc_cntrl" % i) 516 # 517 # Add controllers and sequencers to the appropriate lists 518 # 519 cpu_sequencers.append(sqc_cntrl.sequencer) 520 521 # Connect the SQC controller to the ruby network 522 sqc_cntrl.requestFromSQC = MessageBuffer(ordered = True) 523 sqc_cntrl.requestFromSQC.master = ruby_system.network.slave 524 525 sqc_cntrl.probeToSQC = MessageBuffer(ordered = True) 526 sqc_cntrl.probeToSQC.slave = ruby_system.network.master 527 528 sqc_cntrl.responseToSQC = MessageBuffer(ordered = True) 529 sqc_cntrl.responseToSQC.slave = ruby_system.network.master 530 531 sqc_cntrl.mandatoryQueue = MessageBuffer() 532 533 # SQC also in GPU cluster 534 gpuCluster.add(sqc_cntrl) 535 536 # Because of wire buffers, num_tccs must equal num_tccdirs 537 numa_bit = 6 538 539 for i in xrange(options.num_tccs): 540 541 tcc_cntrl = TCCCntrl() 542 tcc_cntrl.create(options, ruby_system, system) 543 tcc_cntrl.l2_request_latency = options.gpu_to_dir_latency 544 tcc_cntrl.l2_response_latency = options.TCC_latency 545 tcc_cntrl_nodes.append(tcc_cntrl) 546 tcc_cntrl.WB = options.WB_L2 547 tcc_cntrl.number_of_TBEs = 2560 * options.num_compute_units 548 549 # Connect the TCC controllers to the ruby network 550 tcc_cntrl.requestFromTCP = MessageBuffer(ordered = True) 551 tcc_cntrl.requestFromTCP.slave = ruby_system.network.master 552 553 tcc_cntrl.responseToCore = MessageBuffer(ordered = True) 554 tcc_cntrl.responseToCore.master = ruby_system.network.slave 555 556 tcc_cntrl.probeFromNB = MessageBuffer() 557 tcc_cntrl.probeFromNB.slave = ruby_system.network.master 558 559 tcc_cntrl.responseFromNB = MessageBuffer() 560 tcc_cntrl.responseFromNB.slave = ruby_system.network.master 561 562 tcc_cntrl.requestToNB = MessageBuffer(ordered = True) 563 tcc_cntrl.requestToNB.master = ruby_system.network.slave 564 565 tcc_cntrl.responseToNB = MessageBuffer() 566 tcc_cntrl.responseToNB.master = ruby_system.network.slave 567 568 tcc_cntrl.unblockToNB = MessageBuffer() 569 tcc_cntrl.unblockToNB.master = ruby_system.network.slave 570 571 tcc_cntrl.triggerQueue = MessageBuffer(ordered = True) 572 573 exec("system.tcc_cntrl%d = tcc_cntrl" % i) 574 # connect all of the wire buffers between L3 and dirs up 575 # TCC cntrls added to the GPU cluster 576 gpuCluster.add(tcc_cntrl) 577 578 # Assuming no DMA devices 579 assert(len(dma_devices) == 0) 580 581 # Add cpu/gpu clusters to main cluster 582 mainCluster.add(cpuCluster) 583 mainCluster.add(gpuCluster) 584 585 ruby_system.network.number_of_virtual_networks = 10 586 587 return (cpu_sequencers, dir_cntrl_nodes, mainCluster)
|