mem-ruby,mem-garnet: Multiple networks per RubySystem
[gem5.git] / configs / ruby / MOESI_AMD_Base.py
1 # Copyright (c) 2010-2015 Advanced Micro Devices, Inc.
2 # All rights reserved.
3 #
4 # For use for simulation and test purposes only
5 #
6 # Redistribution and use in source and binary forms, with or without
7 # modification, are permitted provided that the following conditions are met:
8 #
9 # 1. Redistributions of source code must retain the above copyright notice,
10 # this list of conditions and the following disclaimer.
11 #
12 # 2. Redistributions in binary form must reproduce the above copyright notice,
13 # this list of conditions and the following disclaimer in the documentation
14 # and/or other materials provided with the distribution.
15 #
16 # 3. Neither the name of the copyright holder nor the names of its
17 # contributors may be used to endorse or promote products derived from this
18 # software without specific prior written permission.
19 #
20 # THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
21 # AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
22 # IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
23 # ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT HOLDER OR CONTRIBUTORS BE
24 # LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR
25 # CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF
26 # SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS
27 # INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN
28 # CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE)
29 # ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
30 # POSSIBILITY OF SUCH DAMAGE.
31
32 import math
33 import m5
34 from m5.objects import *
35 from m5.defines import buildEnv
36 from m5.util import addToPath
37 from Ruby import create_topology
38 from Ruby import send_evicts
39 from common import FileSystemConfig
40
41 addToPath('../')
42
43 from topologies.Cluster import Cluster
44 from topologies.Crossbar import Crossbar
45
46 class 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
67 class 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 = TreePLRURP()
73
74 class 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 = TreePLRURP()
80
81 class 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 = TreePLRURP()
87
88 class 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.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.ruby_system = ruby_system
115 self.sequencer1.coreid = 1
116 self.sequencer1.is_cpu_sequencer = True
117
118 # Defines icache/dcache hit latency
119 self.mandatory_queue_latency = 2
120
121 self.issue_latency = options.cpu_to_dir_latency
122 self.send_evictions = send_evicts(options)
123
124 self.ruby_system = ruby_system
125
126 if options.recycle_latency:
127 self.recycle_latency = options.recycle_latency
128
129 class L3Cache(RubyCache):
130 assoc = 8
131 dataArrayBanks = 256
132 tagArrayBanks = 256
133
134 def create(self, options, ruby_system, system):
135 self.size = MemorySize(options.l3_size)
136 self.size.value /= options.num_dirs
137 self.dataArrayBanks /= options.num_dirs
138 self.tagArrayBanks /= options.num_dirs
139 self.dataArrayBanks /= options.num_dirs
140 self.tagArrayBanks /= options.num_dirs
141 self.dataAccessLatency = options.l3_data_latency
142 self.tagAccessLatency = options.l3_tag_latency
143 self.resourceStalls = options.no_resource_stalls
144 self.replacement_policy = TreePLRURP()
145
146 class L3Cntrl(L3Cache_Controller, CntrlBase):
147 def create(self, options, ruby_system, system):
148 self.version = self.versionCount()
149 self.L3cache = L3Cache()
150 self.L3cache.create(options, ruby_system, system)
151
152 self.l3_response_latency = max(self.L3cache.dataAccessLatency,
153 self.L3cache.tagAccessLatency)
154 self.ruby_system = ruby_system
155
156 if options.recycle_latency:
157 self.recycle_latency = options.recycle_latency
158
159 def connectWireBuffers(self, req_to_dir, resp_to_dir, l3_unblock_to_dir,
160 req_to_l3, probe_to_l3, resp_to_l3):
161 self.reqToDir = req_to_dir
162 self.respToDir = resp_to_dir
163 self.l3UnblockToDir = l3_unblock_to_dir
164 self.reqToL3 = req_to_l3
165 self.probeToL3 = probe_to_l3
166 self.respToL3 = resp_to_l3
167
168 class DirCntrl(Directory_Controller, CntrlBase):
169 def create(self, options, dir_ranges, ruby_system, system):
170 self.version = self.versionCount()
171
172 self.response_latency = 30
173
174 self.addr_ranges = dir_ranges
175 self.directory = RubyDirectoryMemory()
176
177 self.L3CacheMemory = L3Cache()
178 self.L3CacheMemory.create(options, ruby_system, system)
179
180 self.l3_hit_latency = max(self.L3CacheMemory.dataAccessLatency,
181 self.L3CacheMemory.tagAccessLatency)
182
183 self.number_of_TBEs = options.num_tbes
184
185 self.ruby_system = ruby_system
186
187 if options.recycle_latency:
188 self.recycle_latency = options.recycle_latency
189
190 self.CPUonly = True
191
192 def connectWireBuffers(self, req_to_dir, resp_to_dir, l3_unblock_to_dir,
193 req_to_l3, probe_to_l3, resp_to_l3):
194 self.reqToDir = req_to_dir
195 self.respToDir = resp_to_dir
196 self.l3UnblockToDir = l3_unblock_to_dir
197 self.reqToL3 = req_to_l3
198 self.probeToL3 = probe_to_l3
199 self.respToL3 = resp_to_l3
200
201 def define_options(parser):
202 parser.add_option("--num-subcaches", type="int", default=4)
203 parser.add_option("--l3-data-latency", type="int", default=20)
204 parser.add_option("--l3-tag-latency", type="int", default=15)
205 parser.add_option("--cpu-to-dir-latency", type="int", default=15)
206 parser.add_option("--no-resource-stalls", action="store_false",
207 default=True)
208 parser.add_option("--num-tbes", type="int", default=256)
209 parser.add_option("--l2-latency", type="int", default=50) # load to use
210
211 def create_system(options, full_system, system, dma_devices, bootmem,
212 ruby_system):
213 if buildEnv['PROTOCOL'] != 'MOESI_AMD_Base':
214 panic("This script requires the MOESI_AMD_Base protocol.")
215
216 cpu_sequencers = []
217
218 #
219 # The ruby network creation expects the list of nodes in the system to
220 # be consistent with the NetDest list. Therefore the l1 controller
221 # nodes must be listed before the directory nodes and directory nodes
222 # before dma nodes, etc.
223 #
224 l1_cntrl_nodes = []
225 l3_cntrl_nodes = []
226 dir_cntrl_nodes = []
227
228 control_count = 0
229
230 #
231 # Must create the individual controllers before the network to ensure
232 # the controller constructors are called before the network constructor
233 #
234
235 # This is the base crossbar that connects the L3s, Dirs, and cpu
236 # Cluster
237 mainCluster = Cluster(extBW = 512, intBW = 512) # 1 TB/s
238
239 if options.numa_high_bit:
240 numa_bit = options.numa_high_bit
241 else:
242 # if the numa_bit is not specified, set the directory bits as the
243 # lowest bits above the block offset bits, and the numa_bit as the
244 # highest of those directory bits
245 dir_bits = int(math.log(options.num_dirs, 2))
246 block_size_bits = int(math.log(options.cacheline_size, 2))
247 numa_bit = block_size_bits + dir_bits - 1
248
249 for i in range(options.num_dirs):
250 dir_ranges = []
251 for r in system.mem_ranges:
252 addr_range = m5.objects.AddrRange(r.start, size = r.size(),
253 intlvHighBit = numa_bit,
254 intlvBits = dir_bits,
255 intlvMatch = i)
256 dir_ranges.append(addr_range)
257
258
259 dir_cntrl = DirCntrl(TCC_select_num_bits = 0)
260 dir_cntrl.create(options, dir_ranges, ruby_system, system)
261
262 # Connect the Directory controller to the ruby network
263 dir_cntrl.requestFromCores = MessageBuffer(ordered = True)
264 dir_cntrl.requestFromCores.slave = ruby_system.network.master
265
266 dir_cntrl.responseFromCores = MessageBuffer()
267 dir_cntrl.responseFromCores.slave = ruby_system.network.master
268
269 dir_cntrl.unblockFromCores = MessageBuffer()
270 dir_cntrl.unblockFromCores.slave = ruby_system.network.master
271
272 dir_cntrl.probeToCore = MessageBuffer()
273 dir_cntrl.probeToCore.master = ruby_system.network.slave
274
275 dir_cntrl.responseToCore = MessageBuffer()
276 dir_cntrl.responseToCore.master = ruby_system.network.slave
277
278 dir_cntrl.triggerQueue = MessageBuffer(ordered = True)
279 dir_cntrl.L3triggerQueue = MessageBuffer(ordered = True)
280
281 dir_cntrl.requestToMemory = MessageBuffer()
282 dir_cntrl.responseFromMemory = MessageBuffer()
283
284 exec("system.dir_cntrl%d = dir_cntrl" % i)
285 dir_cntrl_nodes.append(dir_cntrl)
286
287 mainCluster.add(dir_cntrl)
288
289 # Technically this config can support an odd number of cpus, but the top
290 # level config files, such as the ruby_random_tester, will get confused if
291 # the number of cpus does not equal the number of sequencers. Thus make
292 # sure that an even number of cpus is specified.
293 assert((options.num_cpus % 2) == 0)
294
295 # For an odd number of CPUs, still create the right number of controllers
296 cpuCluster = Cluster(extBW = 512, intBW = 512) # 1 TB/s
297 for i in range((options.num_cpus + 1) // 2):
298
299 cp_cntrl = CPCntrl()
300 cp_cntrl.create(options, ruby_system, system)
301
302 exec("system.cp_cntrl%d = cp_cntrl" % i)
303 #
304 # Add controllers and sequencers to the appropriate lists
305 #
306 cpu_sequencers.extend([cp_cntrl.sequencer, cp_cntrl.sequencer1])
307
308 # Connect the CP controllers and the network
309 cp_cntrl.requestFromCore = MessageBuffer()
310 cp_cntrl.requestFromCore.master = ruby_system.network.slave
311
312 cp_cntrl.responseFromCore = MessageBuffer()
313 cp_cntrl.responseFromCore.master = ruby_system.network.slave
314
315 cp_cntrl.unblockFromCore = MessageBuffer()
316 cp_cntrl.unblockFromCore.master = ruby_system.network.slave
317
318 cp_cntrl.probeToCore = MessageBuffer()
319 cp_cntrl.probeToCore.slave = ruby_system.network.master
320
321 cp_cntrl.responseToCore = MessageBuffer()
322 cp_cntrl.responseToCore.slave = ruby_system.network.master
323
324 cp_cntrl.mandatoryQueue = MessageBuffer()
325 cp_cntrl.triggerQueue = MessageBuffer(ordered = True)
326
327 cpuCluster.add(cp_cntrl)
328
329 # Register CPUs and caches for each CorePair and directory (SE mode only)
330 if not full_system:
331 for i in xrange((options.num_cpus + 1) // 2):
332 FileSystemConfig.register_cpu(physical_package_id = 0,
333 core_siblings =
334 xrange(options.num_cpus),
335 core_id = i*2,
336 thread_siblings = [])
337
338 FileSystemConfig.register_cpu(physical_package_id = 0,
339 core_siblings =
340 xrange(options.num_cpus),
341 core_id = i*2+1,
342 thread_siblings = [])
343
344 FileSystemConfig.register_cache(level = 0,
345 idu_type = 'Instruction',
346 size = options.l1i_size,
347 line_size = options.cacheline_size,
348 assoc = options.l1i_assoc,
349 cpus = [i*2, i*2+1])
350
351 FileSystemConfig.register_cache(level = 0,
352 idu_type = 'Data',
353 size = options.l1d_size,
354 line_size = options.cacheline_size,
355 assoc = options.l1d_assoc,
356 cpus = [i*2])
357
358 FileSystemConfig.register_cache(level = 0,
359 idu_type = 'Data',
360 size = options.l1d_size,
361 line_size = options.cacheline_size,
362 assoc = options.l1d_assoc,
363 cpus = [i*2+1])
364
365 FileSystemConfig.register_cache(level = 1,
366 idu_type = 'Unified',
367 size = options.l2_size,
368 line_size = options.cacheline_size,
369 assoc = options.l2_assoc,
370 cpus = [i*2, i*2+1])
371
372 for i in range(options.num_dirs):
373 FileSystemConfig.register_cache(level = 2,
374 idu_type = 'Unified',
375 size = options.l3_size,
376 line_size = options.cacheline_size,
377 assoc = options.l3_assoc,
378 cpus = [n for n in
379 xrange(options.num_cpus)])
380
381 # Assuming no DMA devices
382 assert(len(dma_devices) == 0)
383
384 # Add cpu/gpu clusters to main cluster
385 mainCluster.add(cpuCluster)
386
387 ruby_system.network.number_of_virtual_networks = 10
388
389 return (cpu_sequencers, dir_cntrl_nodes, mainCluster)