35
36import math
37import m5
38from m5.objects import *
39from m5.defines import buildEnv
40from Ruby import create_topology
41from Ruby import send_evicts
42
43from topologies.Cluster import Cluster
44from topologies.Crossbar import Crossbar
45
46class CntrlBase:
47 _seqs = 0
48 @classmethod
49 def seqCount(cls):
50 # Use SeqCount not class since we need global count
51 CntrlBase._seqs += 1
52 return CntrlBase._seqs - 1
53
54 _cntrls = 0
55 @classmethod
56 def cntrlCount(cls):
57 # Use CntlCount not class since we need global count
58 CntrlBase._cntrls += 1
59 return CntrlBase._cntrls - 1
60
61 _version = 0
62 @classmethod
63 def versionCount(cls):
64 cls._version += 1 # Use count for this particular type
65 return cls._version - 1
66
67class L1DCache(RubyCache):
68 resourceStalls = False
69 def create(self, options):
70 self.size = MemorySize(options.l1d_size)
71 self.assoc = options.l1d_assoc
72 self.replacement_policy = PseudoLRUReplacementPolicy()
73
74class L1ICache(RubyCache):
75 resourceStalls = False
76 def create(self, options):
77 self.size = MemorySize(options.l1i_size)
78 self.assoc = options.l1i_assoc
79 self.replacement_policy = PseudoLRUReplacementPolicy()
80
81class L2Cache(RubyCache):
82 resourceStalls = False
83 def create(self, options):
84 self.size = MemorySize(options.l2_size)
85 self.assoc = options.l2_assoc
86 self.replacement_policy = PseudoLRUReplacementPolicy()
87
88class CPCntrl(CorePair_Controller, CntrlBase):
89
90 def create(self, options, ruby_system, system):
91 self.version = self.versionCount()
92
93 self.L1Icache = L1ICache()
94 self.L1Icache.create(options)
95 self.L1D0cache = L1DCache()
96 self.L1D0cache.create(options)
97 self.L1D1cache = L1DCache()
98 self.L1D1cache.create(options)
99 self.L2cache = L2Cache()
100 self.L2cache.create(options)
101
102 self.sequencer = RubySequencer()
103 self.sequencer.icache_hit_latency = 2
104 self.sequencer.dcache_hit_latency = 2
105 self.sequencer.version = self.seqCount()
106 self.sequencer.icache = self.L1Icache
107 self.sequencer.dcache = self.L1D0cache
108 self.sequencer.ruby_system = ruby_system
109 self.sequencer.coreid = 0
110 self.sequencer.is_cpu_sequencer = True
111
112 self.sequencer1 = RubySequencer()
113 self.sequencer1.version = self.seqCount()
114 self.sequencer1.icache = self.L1Icache
115 self.sequencer1.dcache = self.L1D1cache
116 self.sequencer1.icache_hit_latency = 2
117 self.sequencer1.dcache_hit_latency = 2
118 self.sequencer1.ruby_system = ruby_system
119 self.sequencer1.coreid = 1
120 self.sequencer1.is_cpu_sequencer = True
121
122 self.issue_latency = options.cpu_to_dir_latency
123 self.send_evictions = send_evicts(options)
124
125 self.ruby_system = ruby_system
126
127 if options.recycle_latency:
128 self.recycle_latency = options.recycle_latency
129
130class L3Cache(RubyCache):
131 assoc = 8
132 dataArrayBanks = 256
133 tagArrayBanks = 256
134
135 def create(self, options, ruby_system, system):
136 self.size = MemorySize(options.l3_size)
137 self.size.value /= options.num_dirs
138 self.dataArrayBanks /= options.num_dirs
139 self.tagArrayBanks /= options.num_dirs
140 self.dataArrayBanks /= options.num_dirs
141 self.tagArrayBanks /= options.num_dirs
142 self.dataAccessLatency = options.l3_data_latency
143 self.tagAccessLatency = options.l3_tag_latency
144 self.resourceStalls = options.no_resource_stalls
145 self.replacement_policy = PseudoLRUReplacementPolicy()
146
147class L3Cntrl(L3Cache_Controller, CntrlBase):
148 def create(self, options, ruby_system, system):
149 self.version = self.versionCount()
150 self.L3cache = L3Cache()
151 self.L3cache.create(options, ruby_system, system)
152
153 self.l3_response_latency = max(self.L3cache.dataAccessLatency,
154 self.L3cache.tagAccessLatency)
155 self.ruby_system = ruby_system
156
157 if options.recycle_latency:
158 self.recycle_latency = options.recycle_latency
159
160 def connectWireBuffers(self, req_to_dir, resp_to_dir, l3_unblock_to_dir,
161 req_to_l3, probe_to_l3, resp_to_l3):
162 self.reqToDir = req_to_dir
163 self.respToDir = resp_to_dir
164 self.l3UnblockToDir = l3_unblock_to_dir
165 self.reqToL3 = req_to_l3
166 self.probeToL3 = probe_to_l3
167 self.respToL3 = resp_to_l3
168
169class DirCntrl(Directory_Controller, CntrlBase):
170 def create(self, options, dir_ranges, ruby_system, system):
171 self.version = self.versionCount()
172
173 self.response_latency = 30
174
175 self.addr_ranges = dir_ranges
176 self.directory = RubyDirectoryMemory()
177
178 self.L3CacheMemory = L3Cache()
179 self.L3CacheMemory.create(options, ruby_system, system)
180
181 self.l3_hit_latency = max(self.L3CacheMemory.dataAccessLatency,
182 self.L3CacheMemory.tagAccessLatency)
183
184 self.number_of_TBEs = options.num_tbes
185
186 self.ruby_system = ruby_system
187
188 if options.recycle_latency:
189 self.recycle_latency = options.recycle_latency
190
191 self.CPUonly = True
192
193 def connectWireBuffers(self, req_to_dir, resp_to_dir, l3_unblock_to_dir,
194 req_to_l3, probe_to_l3, resp_to_l3):
195 self.reqToDir = req_to_dir
196 self.respToDir = resp_to_dir
197 self.l3UnblockToDir = l3_unblock_to_dir
198 self.reqToL3 = req_to_l3
199 self.probeToL3 = probe_to_l3
200 self.respToL3 = resp_to_l3
201
202def define_options(parser):
203 parser.add_option("--num-subcaches", type="int", default=4)
204 parser.add_option("--l3-data-latency", type="int", default=20)
205 parser.add_option("--l3-tag-latency", type="int", default=15)
206 parser.add_option("--cpu-to-dir-latency", type="int", default=15)
207 parser.add_option("--no-resource-stalls", action="store_false",
208 default=True)
209 parser.add_option("--num-tbes", type="int", default=256)
210 parser.add_option("--l2-latency", type="int", default=50) # load to use
211
212def create_system(options, full_system, system, dma_devices, bootmem,
213 ruby_system):
214 if buildEnv['PROTOCOL'] != 'MOESI_AMD_Base':
215 panic("This script requires the MOESI_AMD_Base protocol.")
216
217 cpu_sequencers = []
218
219 #
220 # The ruby network creation expects the list of nodes in the system to
221 # be consistent with the NetDest list. Therefore the l1 controller
222 # nodes must be listed before the directory nodes and directory nodes
223 # before dma nodes, etc.
224 #
225 l1_cntrl_nodes = []
226 l3_cntrl_nodes = []
227 dir_cntrl_nodes = []
228
229 control_count = 0
230
231 #
232 # Must create the individual controllers before the network to ensure
233 # the controller constructors are called before the network constructor
234 #
235
236 # This is the base crossbar that connects the L3s, Dirs, and cpu
237 # Cluster
238 mainCluster = Cluster(extBW = 512, intBW = 512) # 1 TB/s
239
240 if options.numa_high_bit:
241 numa_bit = options.numa_high_bit
242 else:
243 # if the numa_bit is not specified, set the directory bits as the
244 # lowest bits above the block offset bits, and the numa_bit as the
245 # highest of those directory bits
246 dir_bits = int(math.log(options.num_dirs, 2))
247 block_size_bits = int(math.log(options.cacheline_size, 2))
248 numa_bit = block_size_bits + dir_bits - 1
249
250 for i in xrange(options.num_dirs):
251 dir_ranges = []
252 for r in system.mem_ranges:
253 addr_range = m5.objects.AddrRange(r.start, size = r.size(),
254 intlvHighBit = numa_bit,
255 intlvBits = dir_bits,
256 intlvMatch = i)
257 dir_ranges.append(addr_range)
258
259
260 dir_cntrl = DirCntrl(TCC_select_num_bits = 0)
261 dir_cntrl.create(options, dir_ranges, ruby_system, system)
262
263 # Connect the Directory controller to the ruby network
264 dir_cntrl.requestFromCores = MessageBuffer(ordered = True)
265 dir_cntrl.requestFromCores.slave = ruby_system.network.master
266
267 dir_cntrl.responseFromCores = MessageBuffer()
268 dir_cntrl.responseFromCores.slave = ruby_system.network.master
269
270 dir_cntrl.unblockFromCores = MessageBuffer()
271 dir_cntrl.unblockFromCores.slave = ruby_system.network.master
272
273 dir_cntrl.probeToCore = MessageBuffer()
274 dir_cntrl.probeToCore.master = ruby_system.network.slave
275
276 dir_cntrl.responseToCore = MessageBuffer()
277 dir_cntrl.responseToCore.master = ruby_system.network.slave
278
279 dir_cntrl.triggerQueue = MessageBuffer(ordered = True)
280 dir_cntrl.L3triggerQueue = MessageBuffer(ordered = True)
281 dir_cntrl.responseFromMemory = MessageBuffer()
282
283 exec("system.dir_cntrl%d = dir_cntrl" % i)
284 dir_cntrl_nodes.append(dir_cntrl)
285
286 mainCluster.add(dir_cntrl)
287
288 # Technically this config can support an odd number of cpus, but the top
289 # level config files, such as the ruby_random_tester, will get confused if
290 # the number of cpus does not equal the number of sequencers. Thus make
291 # sure that an even number of cpus is specified.
292 assert((options.num_cpus % 2) == 0)
293
294 # For an odd number of CPUs, still create the right number of controllers
295 cpuCluster = Cluster(extBW = 512, intBW = 512) # 1 TB/s
296 for i in xrange((options.num_cpus + 1) / 2):
297
298 cp_cntrl = CPCntrl()
299 cp_cntrl.create(options, ruby_system, system)
300
301 exec("system.cp_cntrl%d = cp_cntrl" % i)
302 #
303 # Add controllers and sequencers to the appropriate lists
304 #
305 cpu_sequencers.extend([cp_cntrl.sequencer, cp_cntrl.sequencer1])
306
307 # Connect the CP controllers and the network
308 cp_cntrl.requestFromCore = MessageBuffer()
309 cp_cntrl.requestFromCore.master = ruby_system.network.slave
310
311 cp_cntrl.responseFromCore = MessageBuffer()
312 cp_cntrl.responseFromCore.master = ruby_system.network.slave
313
314 cp_cntrl.unblockFromCore = MessageBuffer()
315 cp_cntrl.unblockFromCore.master = ruby_system.network.slave
316
317 cp_cntrl.probeToCore = MessageBuffer()
318 cp_cntrl.probeToCore.slave = ruby_system.network.master
319
320 cp_cntrl.responseToCore = MessageBuffer()
321 cp_cntrl.responseToCore.slave = ruby_system.network.master
322
323 cp_cntrl.mandatoryQueue = MessageBuffer()
324 cp_cntrl.triggerQueue = MessageBuffer(ordered = True)
325
326 cpuCluster.add(cp_cntrl)
327
328 # Assuming no DMA devices
329 assert(len(dma_devices) == 0)
330
331 # Add cpu/gpu clusters to main cluster
332 mainCluster.add(cpuCluster)
333
334 ruby_system.network.number_of_virtual_networks = 10
335
336 return (cpu_sequencers, dir_cntrl_nodes, mainCluster)
| 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 L1DCache(RubyCache):
66 resourceStalls = False
67 def create(self, options):
68 self.size = MemorySize(options.l1d_size)
69 self.assoc = options.l1d_assoc
70 self.replacement_policy = PseudoLRUReplacementPolicy()
71
72class L1ICache(RubyCache):
73 resourceStalls = False
74 def create(self, options):
75 self.size = MemorySize(options.l1i_size)
76 self.assoc = options.l1i_assoc
77 self.replacement_policy = PseudoLRUReplacementPolicy()
78
79class L2Cache(RubyCache):
80 resourceStalls = False
81 def create(self, options):
82 self.size = MemorySize(options.l2_size)
83 self.assoc = options.l2_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 = L1ICache()
92 self.L1Icache.create(options)
93 self.L1D0cache = L1DCache()
94 self.L1D0cache.create(options)
95 self.L1D1cache = L1DCache()
96 self.L1D1cache.create(options)
97 self.L2cache = L2Cache()
98 self.L2cache.create(options)
99
100 self.sequencer = RubySequencer()
101 self.sequencer.icache_hit_latency = 2
102 self.sequencer.dcache_hit_latency = 2
103 self.sequencer.version = self.seqCount()
104 self.sequencer.icache = self.L1Icache
105 self.sequencer.dcache = self.L1D0cache
106 self.sequencer.ruby_system = ruby_system
107 self.sequencer.coreid = 0
108 self.sequencer.is_cpu_sequencer = True
109
110 self.sequencer1 = RubySequencer()
111 self.sequencer1.version = self.seqCount()
112 self.sequencer1.icache = self.L1Icache
113 self.sequencer1.dcache = self.L1D1cache
114 self.sequencer1.icache_hit_latency = 2
115 self.sequencer1.dcache_hit_latency = 2
116 self.sequencer1.ruby_system = ruby_system
117 self.sequencer1.coreid = 1
118 self.sequencer1.is_cpu_sequencer = True
119
120 self.issue_latency = options.cpu_to_dir_latency
121 self.send_evictions = send_evicts(options)
122
123 self.ruby_system = ruby_system
124
125 if options.recycle_latency:
126 self.recycle_latency = options.recycle_latency
127
128class L3Cache(RubyCache):
129 assoc = 8
130 dataArrayBanks = 256
131 tagArrayBanks = 256
132
133 def create(self, options, ruby_system, system):
134 self.size = MemorySize(options.l3_size)
135 self.size.value /= options.num_dirs
136 self.dataArrayBanks /= options.num_dirs
137 self.tagArrayBanks /= options.num_dirs
138 self.dataArrayBanks /= options.num_dirs
139 self.tagArrayBanks /= options.num_dirs
140 self.dataAccessLatency = options.l3_data_latency
141 self.tagAccessLatency = options.l3_tag_latency
142 self.resourceStalls = options.no_resource_stalls
143 self.replacement_policy = PseudoLRUReplacementPolicy()
144
145class L3Cntrl(L3Cache_Controller, CntrlBase):
146 def create(self, options, ruby_system, system):
147 self.version = self.versionCount()
148 self.L3cache = L3Cache()
149 self.L3cache.create(options, ruby_system, system)
150
151 self.l3_response_latency = max(self.L3cache.dataAccessLatency,
152 self.L3cache.tagAccessLatency)
153 self.ruby_system = ruby_system
154
155 if options.recycle_latency:
156 self.recycle_latency = options.recycle_latency
157
158 def connectWireBuffers(self, req_to_dir, resp_to_dir, l3_unblock_to_dir,
159 req_to_l3, probe_to_l3, resp_to_l3):
160 self.reqToDir = req_to_dir
161 self.respToDir = resp_to_dir
162 self.l3UnblockToDir = l3_unblock_to_dir
163 self.reqToL3 = req_to_l3
164 self.probeToL3 = probe_to_l3
165 self.respToL3 = resp_to_l3
166
167class DirCntrl(Directory_Controller, CntrlBase):
168 def create(self, options, dir_ranges, ruby_system, system):
169 self.version = self.versionCount()
170
171 self.response_latency = 30
172
173 self.addr_ranges = dir_ranges
174 self.directory = RubyDirectoryMemory()
175
176 self.L3CacheMemory = L3Cache()
177 self.L3CacheMemory.create(options, ruby_system, system)
178
179 self.l3_hit_latency = max(self.L3CacheMemory.dataAccessLatency,
180 self.L3CacheMemory.tagAccessLatency)
181
182 self.number_of_TBEs = options.num_tbes
183
184 self.ruby_system = ruby_system
185
186 if options.recycle_latency:
187 self.recycle_latency = options.recycle_latency
188
189 self.CPUonly = True
190
191 def connectWireBuffers(self, req_to_dir, resp_to_dir, l3_unblock_to_dir,
192 req_to_l3, probe_to_l3, resp_to_l3):
193 self.reqToDir = req_to_dir
194 self.respToDir = resp_to_dir
195 self.l3UnblockToDir = l3_unblock_to_dir
196 self.reqToL3 = req_to_l3
197 self.probeToL3 = probe_to_l3
198 self.respToL3 = resp_to_l3
199
200def define_options(parser):
201 parser.add_option("--num-subcaches", type="int", default=4)
202 parser.add_option("--l3-data-latency", type="int", default=20)
203 parser.add_option("--l3-tag-latency", type="int", default=15)
204 parser.add_option("--cpu-to-dir-latency", type="int", default=15)
205 parser.add_option("--no-resource-stalls", action="store_false",
206 default=True)
207 parser.add_option("--num-tbes", type="int", default=256)
208 parser.add_option("--l2-latency", type="int", default=50) # load to use
209
210def create_system(options, full_system, system, dma_devices, bootmem,
211 ruby_system):
212 if buildEnv['PROTOCOL'] != 'MOESI_AMD_Base':
213 panic("This script requires the MOESI_AMD_Base protocol.")
214
215 cpu_sequencers = []
216
217 #
218 # The ruby network creation expects the list of nodes in the system to
219 # be consistent with the NetDest list. Therefore the l1 controller
220 # nodes must be listed before the directory nodes and directory nodes
221 # before dma nodes, etc.
222 #
223 l1_cntrl_nodes = []
224 l3_cntrl_nodes = []
225 dir_cntrl_nodes = []
226
227 control_count = 0
228
229 #
230 # Must create the individual controllers before the network to ensure
231 # the controller constructors are called before the network constructor
232 #
233
234 # This is the base crossbar that connects the L3s, Dirs, and cpu
235 # Cluster
236 mainCluster = Cluster(extBW = 512, intBW = 512) # 1 TB/s
237
238 if options.numa_high_bit:
239 numa_bit = options.numa_high_bit
240 else:
241 # if the numa_bit is not specified, set the directory bits as the
242 # lowest bits above the block offset bits, and the numa_bit as the
243 # highest of those directory bits
244 dir_bits = int(math.log(options.num_dirs, 2))
245 block_size_bits = int(math.log(options.cacheline_size, 2))
246 numa_bit = block_size_bits + dir_bits - 1
247
248 for i in xrange(options.num_dirs):
249 dir_ranges = []
250 for r in system.mem_ranges:
251 addr_range = m5.objects.AddrRange(r.start, size = r.size(),
252 intlvHighBit = numa_bit,
253 intlvBits = dir_bits,
254 intlvMatch = i)
255 dir_ranges.append(addr_range)
256
257
258 dir_cntrl = DirCntrl(TCC_select_num_bits = 0)
259 dir_cntrl.create(options, dir_ranges, ruby_system, system)
260
261 # Connect the Directory controller to the ruby network
262 dir_cntrl.requestFromCores = MessageBuffer(ordered = True)
263 dir_cntrl.requestFromCores.slave = ruby_system.network.master
264
265 dir_cntrl.responseFromCores = MessageBuffer()
266 dir_cntrl.responseFromCores.slave = ruby_system.network.master
267
268 dir_cntrl.unblockFromCores = MessageBuffer()
269 dir_cntrl.unblockFromCores.slave = ruby_system.network.master
270
271 dir_cntrl.probeToCore = MessageBuffer()
272 dir_cntrl.probeToCore.master = ruby_system.network.slave
273
274 dir_cntrl.responseToCore = MessageBuffer()
275 dir_cntrl.responseToCore.master = ruby_system.network.slave
276
277 dir_cntrl.triggerQueue = MessageBuffer(ordered = True)
278 dir_cntrl.L3triggerQueue = MessageBuffer(ordered = True)
279 dir_cntrl.responseFromMemory = MessageBuffer()
280
281 exec("system.dir_cntrl%d = dir_cntrl" % i)
282 dir_cntrl_nodes.append(dir_cntrl)
283
284 mainCluster.add(dir_cntrl)
285
286 # Technically this config can support an odd number of cpus, but the top
287 # level config files, such as the ruby_random_tester, will get confused if
288 # the number of cpus does not equal the number of sequencers. Thus make
289 # sure that an even number of cpus is specified.
290 assert((options.num_cpus % 2) == 0)
291
292 # For an odd number of CPUs, still create the right number of controllers
293 cpuCluster = Cluster(extBW = 512, intBW = 512) # 1 TB/s
294 for i in xrange((options.num_cpus + 1) / 2):
295
296 cp_cntrl = CPCntrl()
297 cp_cntrl.create(options, ruby_system, system)
298
299 exec("system.cp_cntrl%d = cp_cntrl" % i)
300 #
301 # Add controllers and sequencers to the appropriate lists
302 #
303 cpu_sequencers.extend([cp_cntrl.sequencer, cp_cntrl.sequencer1])
304
305 # Connect the CP controllers and the network
306 cp_cntrl.requestFromCore = MessageBuffer()
307 cp_cntrl.requestFromCore.master = ruby_system.network.slave
308
309 cp_cntrl.responseFromCore = MessageBuffer()
310 cp_cntrl.responseFromCore.master = ruby_system.network.slave
311
312 cp_cntrl.unblockFromCore = MessageBuffer()
313 cp_cntrl.unblockFromCore.master = ruby_system.network.slave
314
315 cp_cntrl.probeToCore = MessageBuffer()
316 cp_cntrl.probeToCore.slave = ruby_system.network.master
317
318 cp_cntrl.responseToCore = MessageBuffer()
319 cp_cntrl.responseToCore.slave = ruby_system.network.master
320
321 cp_cntrl.mandatoryQueue = MessageBuffer()
322 cp_cntrl.triggerQueue = MessageBuffer(ordered = True)
323
324 cpuCluster.add(cp_cntrl)
325
326 # Assuming no DMA devices
327 assert(len(dma_devices) == 0)
328
329 # Add cpu/gpu clusters to main cluster
330 mainCluster.add(cpuCluster)
331
332 ruby_system.network.number_of_virtual_networks = 10
333
334 return (cpu_sequencers, dir_cntrl_nodes, mainCluster)
|