38import optparse, os, re 39import math 40import glob 41import inspect 42 43import m5 44from m5.objects import * 45from m5.util import addToPath 46 47addToPath('../') 48 49from ruby import Ruby 50 51from common import Options 52from common import Simulation 53from common import GPUTLBOptions, GPUTLBConfig 54 55########################## Script Options ######################## 56def setOption(parser, opt_str, value = 1): 57 # check to make sure the option actually exists 58 if not parser.has_option(opt_str): 59 raise Exception("cannot find %s in list of possible options" % opt_str) 60 61 opt = parser.get_option(opt_str) 62 # set the value 63 exec("parser.values.%s = %s" % (opt.dest, value)) 64 65def getOption(parser, opt_str): 66 # check to make sure the option actually exists 67 if not parser.has_option(opt_str): 68 raise Exception("cannot find %s in list of possible options" % opt_str) 69 70 opt = parser.get_option(opt_str) 71 # get the value 72 exec("return_value = parser.values.%s" % opt.dest) 73 return return_value 74 75# Adding script options 76parser = optparse.OptionParser() 77Options.addCommonOptions(parser) 78Options.addSEOptions(parser) 79 80parser.add_option("--cpu-only-mode", action="store_true", default=False, 81 help="APU mode. Used to take care of problems in "\ 82 "Ruby.py while running APU protocols") 83parser.add_option("-k", "--kernel-files", 84 help="file(s) containing GPU kernel code (colon separated)") 85parser.add_option("-u", "--num-compute-units", type="int", default=1, 86 help="number of GPU compute units"), 87parser.add_option("--num-cp", type="int", default=0, 88 help="Number of GPU Command Processors (CP)") 89parser.add_option("--benchmark-root", help="Root of benchmark directory tree") 90 91# not super important now, but to avoid putting the number 4 everywhere, make 92# it an option/knob 93parser.add_option("--cu-per-sqc", type="int", default=4, help="number of CUs" \ 94 "sharing an SQC (icache, and thus icache TLB)") 95parser.add_option("--simds-per-cu", type="int", default=4, help="SIMD units" \ 96 "per CU") 97parser.add_option("--wf-size", type="int", default=64, 98 help="Wavefront size(in workitems)") 99parser.add_option("--sp-bypass-path-length", type="int", default=4, \ 100 help="Number of stages of bypass path in vector ALU for Single Precision ops") 101parser.add_option("--dp-bypass-path-length", type="int", default=4, \ 102 help="Number of stages of bypass path in vector ALU for Double Precision ops") 103# issue period per SIMD unit: number of cycles before issuing another vector 104parser.add_option("--issue-period", type="int", default=4, \ 105 help="Number of cycles per vector instruction issue period") 106parser.add_option("--glbmem-wr-bus-width", type="int", default=32, \ 107 help="VGPR to Coalescer (Global Memory) data bus width in bytes") 108parser.add_option("--glbmem-rd-bus-width", type="int", default=32, \ 109 help="Coalescer to VGPR (Global Memory) data bus width in bytes") 110# Currently we only support 1 local memory pipe 111parser.add_option("--shr-mem-pipes-per-cu", type="int", default=1, \ 112 help="Number of Shared Memory pipelines per CU") 113# Currently we only support 1 global memory pipe 114parser.add_option("--glb-mem-pipes-per-cu", type="int", default=1, \ 115 help="Number of Global Memory pipelines per CU") 116parser.add_option("--wfs-per-simd", type="int", default=10, help="Number of " \ 117 "WF slots per SIMD") 118 119parser.add_option("--vreg-file-size", type="int", default=2048, 120 help="number of physical vector registers per SIMD") 121parser.add_option("--bw-scalor", type="int", default=0, 122 help="bandwidth scalor for scalability analysis") 123parser.add_option("--CPUClock", type="string", default="2GHz", 124 help="CPU clock") 125parser.add_option("--GPUClock", type="string", default="1GHz", 126 help="GPU clock") 127parser.add_option("--cpu-voltage", action="store", type="string", 128 default='1.0V', 129 help = """CPU voltage domain""") 130parser.add_option("--gpu-voltage", action="store", type="string", 131 default='1.0V', 132 help = """CPU voltage domain""") 133parser.add_option("--CUExecPolicy", type="string", default="OLDEST-FIRST", 134 help="WF exec policy (OLDEST-FIRST, ROUND-ROBIN)") 135parser.add_option("--xact-cas-mode", action="store_true", 136 help="enable load_compare mode (transactional CAS)") 137parser.add_option("--SegFaultDebug",action="store_true", 138 help="checks for GPU seg fault before TLB access") 139parser.add_option("--FunctionalTLB",action="store_true", 140 help="Assumes TLB has no latency") 141parser.add_option("--LocalMemBarrier",action="store_true", 142 help="Barrier does not wait for writethroughs to complete") 143parser.add_option("--countPages", action="store_true", 144 help="Count Page Accesses and output in per-CU output files") 145parser.add_option("--TLB-prefetch", type="int", help = "prefetch depth for"\ 146 "TLBs") 147parser.add_option("--pf-type", type="string", help="type of prefetch: "\ 148 "PF_CU, PF_WF, PF_PHASE, PF_STRIDE") 149parser.add_option("--pf-stride", type="int", help="set prefetch stride") 150parser.add_option("--numLdsBanks", type="int", default=32, 151 help="number of physical banks per LDS module") 152parser.add_option("--ldsBankConflictPenalty", type="int", default=1, 153 help="number of cycles per LDS bank conflict") 154parser.add_option('--fast-forward-pseudo-op', action='store_true', 155 help = 'fast forward using kvm until the m5_switchcpu' 156 ' pseudo-op is encountered, then switch cpus. subsequent' 157 ' m5_switchcpu pseudo-ops will toggle back and forth') 158parser.add_option('--outOfOrderDataDelivery', action='store_true', 159 default=False, help='enable OoO data delivery in the GM' 160 ' pipeline') 161 162Ruby.define_options(parser) 163 164#add TLB options to the parser 165GPUTLBOptions.tlb_options(parser) 166 167(options, args) = parser.parse_args() 168 169# The GPU cache coherence protocols only work with the backing store 170setOption(parser, "--access-backing-store") 171 172# if benchmark root is specified explicitly, that overrides the search path 173if options.benchmark_root: 174 benchmark_path = [options.benchmark_root] 175else: 176 # Set default benchmark search path to current dir 177 benchmark_path = ['.'] 178 179########################## Sanity Check ######################## 180 181# Currently the gpu model requires ruby 182if buildEnv['PROTOCOL'] == 'None': 183 fatal("GPU model requires ruby") 184 185# Currently the gpu model requires only timing or detailed CPU 186if not (options.cpu_type == "TimingSimpleCPU" or 187 options.cpu_type == "DerivO3CPU"): 188 fatal("GPU model requires TimingSimpleCPU or DerivO3CPU") 189 190# This file can support multiple compute units 191assert(options.num_compute_units >= 1) 192 193# Currently, the sqc (I-Cache of GPU) is shared by 194# multiple compute units(CUs). The protocol works just fine 195# even if sqc is not shared. Overriding this option here 196# so that the user need not explicitly set this (assuming 197# sharing sqc is the common usage) 198n_cu = options.num_compute_units 199num_sqc = int(math.ceil(float(n_cu) / options.cu_per_sqc)) 200options.num_sqc = num_sqc # pass this to Ruby 201 202########################## Creating the GPU system ######################## 203# shader is the GPU 204shader = Shader(n_wf = options.wfs_per_simd, 205 clk_domain = SrcClockDomain( 206 clock = options.GPUClock, 207 voltage_domain = VoltageDomain( 208 voltage = options.gpu_voltage))) 209 210# GPU_RfO(Read For Ownership) implements SC/TSO memory model. 211# Other GPU protocols implement release consistency at GPU side. 212# So, all GPU protocols other than GPU_RfO should make their writes 213# visible to the global memory and should read from global memory 214# during kernal boundary. The pipeline initiates(or do not initiate) 215# the acquire/release operation depending on this impl_kern_boundary_sync 216# flag. This flag=true means pipeline initiates a acquire/release operation 217# at kernel boundary. 218if buildEnv['PROTOCOL'] == 'GPU_RfO': 219 shader.impl_kern_boundary_sync = False 220else: 221 shader.impl_kern_boundary_sync = True 222 223# Switching off per-lane TLB by default 224per_lane = False 225if options.TLB_config == "perLane": 226 per_lane = True 227 228# List of compute units; one GPU can have multiple compute units 229compute_units = [] 230for i in xrange(n_cu): 231 compute_units.append(ComputeUnit(cu_id = i, perLaneTLB = per_lane, 232 num_SIMDs = options.simds_per_cu, 233 wfSize = options.wf_size, 234 spbypass_pipe_length = options.sp_bypass_path_length, 235 dpbypass_pipe_length = options.dp_bypass_path_length, 236 issue_period = options.issue_period, 237 coalescer_to_vrf_bus_width = \ 238 options.glbmem_rd_bus_width, 239 vrf_to_coalescer_bus_width = \ 240 options.glbmem_wr_bus_width, 241 num_global_mem_pipes = \ 242 options.glb_mem_pipes_per_cu, 243 num_shared_mem_pipes = \ 244 options.shr_mem_pipes_per_cu, 245 n_wf = options.wfs_per_simd, 246 execPolicy = options.CUExecPolicy, 247 xactCasMode = options.xact_cas_mode, 248 debugSegFault = options.SegFaultDebug, 249 functionalTLB = options.FunctionalTLB, 250 localMemBarrier = options.LocalMemBarrier, 251 countPages = options.countPages, 252 localDataStore = \ 253 LdsState(banks = options.numLdsBanks, 254 bankConflictPenalty = \ 255 options.ldsBankConflictPenalty), 256 out_of_order_data_delivery = 257 options.outOfOrderDataDelivery)) 258 wavefronts = [] 259 vrfs = [] 260 for j in xrange(options.simds_per_cu): 261 for k in xrange(shader.n_wf): 262 wavefronts.append(Wavefront(simdId = j, wf_slot_id = k, 263 wfSize = options.wf_size)) 264 vrfs.append(VectorRegisterFile(simd_id=j, 265 num_regs_per_simd=options.vreg_file_size)) 266 compute_units[-1].wavefronts = wavefronts 267 compute_units[-1].vector_register_file = vrfs 268 if options.TLB_prefetch: 269 compute_units[-1].prefetch_depth = options.TLB_prefetch 270 compute_units[-1].prefetch_prev_type = options.pf_type 271 272 # attach the LDS and the CU to the bus (actually a Bridge) 273 compute_units[-1].ldsPort = compute_units[-1].ldsBus.slave 274 compute_units[-1].ldsBus.master = compute_units[-1].localDataStore.cuPort 275 276# Attach compute units to GPU 277shader.CUs = compute_units 278 279########################## Creating the CPU system ######################## 280options.num_cpus = options.num_cpus 281 282# The shader core will be whatever is after the CPU cores are accounted for 283shader_idx = options.num_cpus 284 285# The command processor will be whatever is after the shader is accounted for 286cp_idx = shader_idx + 1 287cp_list = [] 288 289# List of CPUs 290cpu_list = [] 291 292CpuClass, mem_mode = Simulation.getCPUClass(options.cpu_type) 293if CpuClass == AtomicSimpleCPU: 294 fatal("AtomicSimpleCPU is not supported") 295if mem_mode != 'timing': 296 fatal("Only the timing memory mode is supported") 297shader.timing = True 298 299if options.fast_forward and options.fast_forward_pseudo_op: 300 fatal("Cannot fast-forward based both on the number of instructions and" 301 " on pseudo-ops") 302fast_forward = options.fast_forward or options.fast_forward_pseudo_op 303 304if fast_forward: 305 FutureCpuClass, future_mem_mode = CpuClass, mem_mode 306 307 CpuClass = X86KvmCPU 308 mem_mode = 'atomic_noncaching' 309 # Leave shader.timing untouched, because its value only matters at the 310 # start of the simulation and because we require switching cpus 311 # *before* the first kernel launch. 312 313 future_cpu_list = [] 314 315 # Initial CPUs to be used during fast-forwarding. 316 for i in xrange(options.num_cpus): 317 cpu = CpuClass(cpu_id = i, 318 clk_domain = SrcClockDomain( 319 clock = options.CPUClock, 320 voltage_domain = VoltageDomain( 321 voltage = options.cpu_voltage))) 322 cpu_list.append(cpu) 323 324 if options.fast_forward: 325 cpu.max_insts_any_thread = int(options.fast_forward) 326 327if fast_forward: 328 MainCpuClass = FutureCpuClass 329else: 330 MainCpuClass = CpuClass 331 332# CPs to be used throughout the simulation. 333for i in xrange(options.num_cp): 334 cp = MainCpuClass(cpu_id = options.num_cpus + i, 335 clk_domain = SrcClockDomain( 336 clock = options.CPUClock, 337 voltage_domain = VoltageDomain( 338 voltage = options.cpu_voltage))) 339 cp_list.append(cp) 340 341# Main CPUs (to be used after fast-forwarding if fast-forwarding is specified). 342for i in xrange(options.num_cpus): 343 cpu = MainCpuClass(cpu_id = i, 344 clk_domain = SrcClockDomain( 345 clock = options.CPUClock, 346 voltage_domain = VoltageDomain( 347 voltage = options.cpu_voltage))) 348 if fast_forward: 349 cpu.switched_out = True 350 future_cpu_list.append(cpu) 351 else: 352 cpu_list.append(cpu) 353 354########################## Creating the GPU dispatcher ######################## 355# Dispatcher dispatches work from host CPU to GPU 356host_cpu = cpu_list[0] 357dispatcher = GpuDispatcher() 358 359########################## Create and assign the workload ######################## 360# Check for rel_path in elements of base_list using test, returning 361# the first full path that satisfies test 362def find_path(base_list, rel_path, test): 363 for base in base_list: 364 if not base: 365 # base could be None if environment var not set 366 continue 367 full_path = os.path.join(base, rel_path) 368 if test(full_path): 369 return full_path 370 fatal("%s not found in %s" % (rel_path, base_list)) 371 372def find_file(base_list, rel_path): 373 return find_path(base_list, rel_path, os.path.isfile) 374 375executable = find_path(benchmark_path, options.cmd, os.path.exists) 376# it's common for a benchmark to be in a directory with the same 377# name as the executable, so we handle that automatically 378if os.path.isdir(executable): 379 benchmark_path = [executable] 380 executable = find_file(benchmark_path, options.cmd) 381if options.kernel_files: 382 kernel_files = [find_file(benchmark_path, f) 383 for f in options.kernel_files.split(':')] 384else: 385 # if kernel_files is not set, see if there's a unique .asm file 386 # in the same directory as the executable 387 kernel_path = os.path.dirname(executable) 388 kernel_files = glob.glob(os.path.join(kernel_path, '*.asm')) 389 if kernel_files:
|