MOESI_hammer.py revision 9841:69c158420c51
1# Copyright (c) 2006-2007 The Regents of The University of Michigan 2# Copyright (c) 2009 Advanced Micro Devices, Inc. 3# All rights reserved. 4# 5# Redistribution and use in source and binary forms, with or without 6# modification, are permitted provided that the following conditions are 7# met: redistributions of source code must retain the above copyright 8# notice, this list of conditions and the following disclaimer; 9# redistributions in binary form must reproduce the above copyright 10# notice, this list of conditions and the following disclaimer in the 11# documentation and/or other materials provided with the distribution; 12# neither the name of the copyright holders nor the names of its 13# contributors may be used to endorse or promote products derived from 14# this software without specific prior written permission. 15# 16# THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS 17# "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT 18# LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR 19# A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT 20# OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, 21# SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT 22# LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, 23# DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY 24# THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT 25# (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE 26# OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. 27# 28# Authors: Brad Beckmann 29 30import math 31import m5 32from m5.objects import * 33from m5.defines import buildEnv 34from Ruby import create_topology 35 36# 37# Note: the L1 Cache latency is only used by the sequencer on fast path hits 38# 39class L1Cache(RubyCache): 40 latency = 2 41 42# 43# Note: the L2 Cache latency is not currently used 44# 45class L2Cache(RubyCache): 46 latency = 10 47 48# 49# Probe filter is a cache, latency is not used 50# 51class ProbeFilter(RubyCache): 52 latency = 1 53 54def define_options(parser): 55 parser.add_option("--allow-atomic-migration", action="store_true", 56 help="allow migratory sharing for atomic only accessed blocks") 57 parser.add_option("--pf-on", action="store_true", 58 help="Hammer: enable Probe Filter") 59 parser.add_option("--dir-on", action="store_true", 60 help="Hammer: enable Full-bit Directory") 61 62def create_system(options, system, piobus, dma_ports, ruby_system): 63 64 if buildEnv['PROTOCOL'] != 'MOESI_hammer': 65 panic("This script requires the MOESI_hammer protocol to be built.") 66 67 cpu_sequencers = [] 68 69 # 70 # The ruby network creation expects the list of nodes in the system to be 71 # consistent with the NetDest list. Therefore the l1 controller nodes must be 72 # listed before the directory nodes and directory nodes before dma nodes, etc. 73 # 74 l1_cntrl_nodes = [] 75 dir_cntrl_nodes = [] 76 dma_cntrl_nodes = [] 77 78 # 79 # Must create the individual controllers before the network to ensure the 80 # controller constructors are called before the network constructor 81 # 82 block_size_bits = int(math.log(options.cacheline_size, 2)) 83 84 cntrl_count = 0 85 86 for i in xrange(options.num_cpus): 87 # 88 # First create the Ruby objects associated with this cpu 89 # 90 l1i_cache = L1Cache(size = options.l1i_size, 91 assoc = options.l1i_assoc, 92 start_index_bit = block_size_bits, 93 is_icache = True) 94 l1d_cache = L1Cache(size = options.l1d_size, 95 assoc = options.l1d_assoc, 96 start_index_bit = block_size_bits) 97 l2_cache = L2Cache(size = options.l2_size, 98 assoc = options.l2_assoc, 99 start_index_bit = block_size_bits) 100 101 l1_cntrl = L1Cache_Controller(version = i, 102 cntrl_id = cntrl_count, 103 L1Icache = l1i_cache, 104 L1Dcache = l1d_cache, 105 L2cache = l2_cache, 106 no_mig_atomic = not \ 107 options.allow_atomic_migration, 108 send_evictions = ( 109 options.cpu_type == "detailed"), 110 transitions_per_cycle = options.ports, 111 ruby_system = ruby_system) 112 113 cpu_seq = RubySequencer(version = i, 114 icache = l1i_cache, 115 dcache = l1d_cache, 116 ruby_system = ruby_system) 117 118 l1_cntrl.sequencer = cpu_seq 119 120 if piobus != None: 121 cpu_seq.pio_port = piobus.slave 122 123 if options.recycle_latency: 124 l1_cntrl.recycle_latency = options.recycle_latency 125 126 exec("ruby_system.l1_cntrl%d = l1_cntrl" % i) 127 # 128 # Add controllers and sequencers to the appropriate lists 129 # 130 cpu_sequencers.append(cpu_seq) 131 l1_cntrl_nodes.append(l1_cntrl) 132 133 cntrl_count += 1 134 135 phys_mem_size = sum(map(lambda r: r.size(), system.mem_ranges)) 136 assert(phys_mem_size % options.num_dirs == 0) 137 mem_module_size = phys_mem_size / options.num_dirs 138 139 # 140 # determine size and index bits for probe filter 141 # By default, the probe filter size is configured to be twice the 142 # size of the L2 cache. 143 # 144 pf_size = MemorySize(options.l2_size) 145 pf_size.value = pf_size.value * 2 146 dir_bits = int(math.log(options.num_dirs, 2)) 147 pf_bits = int(math.log(pf_size.value, 2)) 148 if options.numa_high_bit: 149 if options.pf_on or options.dir_on: 150 # if numa high bit explicitly set, make sure it does not overlap 151 # with the probe filter index 152 assert(options.numa_high_bit - dir_bits > pf_bits) 153 154 # set the probe filter start bit to just above the block offset 155 pf_start_bit = block_size_bits 156 else: 157 if dir_bits > 0: 158 pf_start_bit = dir_bits + block_size_bits - 1 159 else: 160 pf_start_bit = block_size_bits 161 162 # Run each of the ruby memory controllers at a ratio of the frequency of 163 # the ruby system 164 # clk_divider value is a fix to pass regression. 165 ruby_system.memctrl_clk_domain = DerivedClockDomain( 166 clk_domain=ruby_system.clk_domain, 167 clk_divider=3) 168 169 for i in xrange(options.num_dirs): 170 # 171 # Create the Ruby objects associated with the directory controller 172 # 173 174 mem_cntrl = RubyMemoryControl( 175 clk_domain = ruby_system.memctrl_clk_domain, 176 version = i, 177 ruby_system = ruby_system) 178 179 dir_size = MemorySize('0B') 180 dir_size.value = mem_module_size 181 182 pf = ProbeFilter(size = pf_size, assoc = 4, 183 start_index_bit = pf_start_bit) 184 185 dir_cntrl = Directory_Controller(version = i, 186 cntrl_id = cntrl_count, 187 directory = \ 188 RubyDirectoryMemory( \ 189 version = i, 190 size = dir_size, 191 use_map = options.use_map, 192 map_levels = \ 193 options.map_levels, 194 numa_high_bit = \ 195 options.numa_high_bit), 196 probeFilter = pf, 197 memBuffer = mem_cntrl, 198 probe_filter_enabled = options.pf_on, 199 full_bit_dir_enabled = options.dir_on, 200 transitions_per_cycle = options.ports, 201 ruby_system = ruby_system) 202 203 if options.recycle_latency: 204 dir_cntrl.recycle_latency = options.recycle_latency 205 206 exec("ruby_system.dir_cntrl%d = dir_cntrl" % i) 207 dir_cntrl_nodes.append(dir_cntrl) 208 209 cntrl_count += 1 210 211 for i, dma_port in enumerate(dma_ports): 212 # 213 # Create the Ruby objects associated with the dma controller 214 # 215 dma_seq = DMASequencer(version = i, 216 ruby_system = ruby_system) 217 218 dma_cntrl = DMA_Controller(version = i, 219 cntrl_id = cntrl_count, 220 dma_sequencer = dma_seq, 221 transitions_per_cycle = options.ports, 222 ruby_system = ruby_system) 223 224 exec("ruby_system.dma_cntrl%d = dma_cntrl" % i) 225 exec("ruby_system.dma_cntrl%d.dma_sequencer.slave = dma_port" % i) 226 dma_cntrl_nodes.append(dma_cntrl) 227 228 if options.recycle_latency: 229 dma_cntrl.recycle_latency = options.recycle_latency 230 231 cntrl_count += 1 232 233 all_cntrls = l1_cntrl_nodes + dir_cntrl_nodes + dma_cntrl_nodes 234 235 topology = create_topology(all_cntrls, options) 236 237 return (cpu_sequencers, dir_cntrl_nodes, topology) 238