1 # Copyright (c) 2015 Advanced Micro Devices, Inc.
4 # For use for simulation and test purposes only
6 # Redistribution and use in source and binary forms, with or without
7 # modification, are permitted provided that the following conditions are met:
9 # 1. Redistributions of source code must retain the above copyright notice,
10 # this list of conditions and the following disclaimer.
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.
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.
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.
32 from __future__
import print_function
33 from __future__
import absolute_import
35 import optparse
, os
, re
, getpass
41 from m5
.objects
import *
42 from m5
.util
import addToPath
48 from common
import Options
49 from common
import Simulation
50 from common
import GPUTLBOptions
, GPUTLBConfig
53 from common
import FileSystemConfig
55 ########################## Script Options ########################
56 def 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
)
61 opt
= parser
.get_option(opt_str
)
63 exec("parser.values.%s = %s" % (opt
.dest
, value
))
65 def 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
)
70 opt
= parser
.get_option(opt_str
)
72 exec("return_value = parser.values.%s" % opt
.dest
)
75 # Adding script options
76 parser
= optparse
.OptionParser()
77 Options
.addCommonOptions(parser
)
78 Options
.addSEOptions(parser
)
80 parser
.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")
83 parser
.add_option("-u", "--num-compute-units", type="int", default
=4,
84 help="number of GPU compute units"),
85 parser
.add_option("--num-cp", type="int", default
=0,
86 help="Number of GPU Command Processors (CP)")
87 parser
.add_option("--benchmark-root", help="Root of benchmark directory tree")
89 # not super important now, but to avoid putting the number 4 everywhere, make
91 parser
.add_option("--cu-per-sqc", type="int", default
=4, help="number of CUs" \
92 "sharing an SQC (icache, and thus icache TLB)")
93 parser
.add_option('--cu-per-scalar-cache', type='int', default
=4,
94 help='Number of CUs sharing a scalar cache')
95 parser
.add_option("--simds-per-cu", type="int", default
=4, help="SIMD units" \
97 parser
.add_option('--cu-per-sa', type='int', default
=4,
98 help='Number of CUs per shader array. This must be a '
99 'multiple of options.cu-per-sqc and options.cu-per-scalar')
100 parser
.add_option('--sa-per-complex', type='int', default
=1,
101 help='Number of shader arrays per complex')
102 parser
.add_option('--num-gpu-complexes', type='int', default
=1,
103 help='Number of GPU complexes')
104 parser
.add_option("--wf-size", type="int", default
=64,
105 help="Wavefront size(in workitems)")
106 parser
.add_option("--sp-bypass-path-length", type="int", default
=4, \
107 help="Number of stages of bypass path in vector ALU for "
108 "Single Precision ops")
109 parser
.add_option("--dp-bypass-path-length", type="int", default
=4, \
110 help="Number of stages of bypass path in vector ALU for "
111 "Double Precision ops")
112 # issue period per SIMD unit: number of cycles before issuing another vector
113 parser
.add_option("--issue-period", type="int", default
=4, \
114 help="Number of cycles per vector instruction issue period")
115 parser
.add_option("--glbmem-wr-bus-width", type="int", default
=32, \
116 help="VGPR to Coalescer (Global Memory) data bus width "
118 parser
.add_option("--glbmem-rd-bus-width", type="int", default
=32, \
119 help="Coalescer to VGPR (Global Memory) data bus width in "
121 # Currently we only support 1 local memory pipe
122 parser
.add_option("--shr-mem-pipes-per-cu", type="int", default
=1, \
123 help="Number of Shared Memory pipelines per CU")
124 # Currently we only support 1 global memory pipe
125 parser
.add_option("--glb-mem-pipes-per-cu", type="int", default
=1, \
126 help="Number of Global Memory pipelines per CU")
127 parser
.add_option("--wfs-per-simd", type="int", default
=10, help="Number of " \
130 parser
.add_option("--registerManagerPolicy", type="string", default
="static",
131 help="Register manager policy")
132 parser
.add_option("--vreg-file-size", type="int", default
=2048,
133 help="number of physical vector registers per SIMD")
134 parser
.add_option("--vreg-min-alloc", type="int", default
=4,
135 help="Minimum number of registers that can be allocated "
136 "from the VRF. The total number of registers will be "
137 "aligned to this value.")
139 parser
.add_option("--sreg-file-size", type="int", default
=2048,
140 help="number of physical vector registers per SIMD")
141 parser
.add_option("--sreg-min-alloc", type="int", default
=4,
142 help="Minimum number of registers that can be allocated "
143 "from the SRF. The total number of registers will be "
144 "aligned to this value.")
146 parser
.add_option("--bw-scalor", type="int", default
=0,
147 help="bandwidth scalor for scalability analysis")
148 parser
.add_option("--CPUClock", type="string", default
="2GHz",
150 parser
.add_option("--gpu-clock", type="string", default
="1GHz",
152 parser
.add_option("--cpu-voltage", action
="store", type="string",
154 help = """CPU voltage domain""")
155 parser
.add_option("--gpu-voltage", action
="store", type="string",
157 help = """CPU voltage domain""")
158 parser
.add_option("--CUExecPolicy", type="string", default
="OLDEST-FIRST",
159 help="WF exec policy (OLDEST-FIRST, ROUND-ROBIN)")
160 parser
.add_option("--SegFaultDebug",action
="store_true",
161 help="checks for GPU seg fault before TLB access")
162 parser
.add_option("--FunctionalTLB",action
="store_true",
163 help="Assumes TLB has no latency")
164 parser
.add_option("--LocalMemBarrier",action
="store_true",
165 help="Barrier does not wait for writethroughs to complete")
166 parser
.add_option("--countPages", action
="store_true",
167 help="Count Page Accesses and output in per-CU output files")
168 parser
.add_option("--TLB-prefetch", type="int", help = "prefetch depth for"\
170 parser
.add_option("--pf-type", type="string", help="type of prefetch: "\
171 "PF_CU, PF_WF, PF_PHASE, PF_STRIDE")
172 parser
.add_option("--pf-stride", type="int", help="set prefetch stride")
173 parser
.add_option("--numLdsBanks", type="int", default
=32,
174 help="number of physical banks per LDS module")
175 parser
.add_option("--ldsBankConflictPenalty", type="int", default
=1,
176 help="number of cycles per LDS bank conflict")
177 parser
.add_option('--fast-forward-pseudo-op', action
='store_true',
178 help = 'fast forward using kvm until the m5_switchcpu'
179 ' pseudo-op is encountered, then switch cpus. subsequent'
180 ' m5_switchcpu pseudo-ops will toggle back and forth')
181 parser
.add_option("--num-hw-queues", type="int", default
=10,
182 help="number of hw queues in packet processor")
184 Ruby
.define_options(parser
)
186 #add TLB options to the parser
187 GPUTLBOptions
.tlb_options(parser
)
189 (options
, args
) = parser
.parse_args()
191 # The GPU cache coherence protocols only work with the backing store
192 setOption(parser
, "--access-backing-store")
194 # if benchmark root is specified explicitly, that overrides the search path
195 if options
.benchmark_root
:
196 benchmark_path
= [options
.benchmark_root
]
198 # Set default benchmark search path to current dir
199 benchmark_path
= ['.']
201 ########################## Sanity Check ########################
203 # Currently the gpu model requires ruby
204 if buildEnv
['PROTOCOL'] == 'None':
205 fatal("GPU model requires ruby")
207 # Currently the gpu model requires only timing or detailed CPU
208 if not (options
.cpu_type
== "TimingSimpleCPU" or
209 options
.cpu_type
== "DerivO3CPU"):
210 fatal("GPU model requires TimingSimpleCPU or DerivO3CPU")
212 # This file can support multiple compute units
213 assert(options
.num_compute_units
>= 1)
215 # Currently, the sqc (I-Cache of GPU) is shared by
216 # multiple compute units(CUs). The protocol works just fine
217 # even if sqc is not shared. Overriding this option here
218 # so that the user need not explicitly set this (assuming
219 # sharing sqc is the common usage)
220 n_cu
= options
.num_compute_units
221 num_sqc
= int(math
.ceil(float(n_cu
) / options
.cu_per_sqc
))
222 options
.num_sqc
= num_sqc
# pass this to Ruby
223 num_scalar_cache
= int(math
.ceil(float(n_cu
) / options
.cu_per_scalar_cache
))
224 options
.num_scalar_cache
= num_scalar_cache
226 print('Num SQC = ', num_sqc
, 'Num scalar caches = ', num_scalar_cache
,
229 ########################## Creating the GPU system ########################
231 shader
= Shader(n_wf
= options
.wfs_per_simd
,
232 clk_domain
= SrcClockDomain(
233 clock
= options
.gpu_clock
,
234 voltage_domain
= VoltageDomain(
235 voltage
= options
.gpu_voltage
)))
237 # GPU_RfO(Read For Ownership) implements SC/TSO memory model.
238 # Other GPU protocols implement release consistency at GPU side.
239 # So, all GPU protocols other than GPU_RfO should make their writes
240 # visible to the global memory and should read from global memory
241 # during kernal boundary. The pipeline initiates(or do not initiate)
242 # the acquire/release operation depending on these impl_kern_launch_rel
243 # and impl_kern_end_rel flags. The flag=true means pipeline initiates
244 # a acquire/release operation at kernel launch/end.
245 # VIPER protocols (GPU_VIPER, GPU_VIPER_Region and GPU_VIPER_Baseline)
246 # are write-through based, and thus only imple_kern_launch_acq needs to
248 if buildEnv
['PROTOCOL'] == 'GPU_RfO':
249 shader
.impl_kern_launch_acq
= False
250 shader
.impl_kern_end_rel
= False
251 elif (buildEnv
['PROTOCOL'] != 'GPU_VIPER' or
252 buildEnv
['PROTOCOL'] != 'GPU_VIPER_Region' or
253 buildEnv
['PROTOCOL'] != 'GPU_VIPER_Baseline'):
254 shader
.impl_kern_launch_acq
= True
255 shader
.impl_kern_end_rel
= False
257 shader
.impl_kern_launch_acq
= True
258 shader
.impl_kern_end_rel
= True
260 # Switching off per-lane TLB by default
262 if options
.TLB_config
== "perLane":
265 # List of compute units; one GPU can have multiple compute units
267 for i
in range(n_cu
):
268 compute_units
.append(ComputeUnit(cu_id
= i
, perLaneTLB
= per_lane
,
269 num_SIMDs
= options
.simds_per_cu
,
270 wf_size
= options
.wf_size
,
271 spbypass_pipe_length
= \
272 options
.sp_bypass_path_length
,
273 dpbypass_pipe_length
= \
274 options
.dp_bypass_path_length
,
275 issue_period
= options
.issue_period
,
276 coalescer_to_vrf_bus_width
= \
277 options
.glbmem_rd_bus_width
,
278 vrf_to_coalescer_bus_width
= \
279 options
.glbmem_wr_bus_width
,
280 num_global_mem_pipes
= \
281 options
.glb_mem_pipes_per_cu
,
282 num_shared_mem_pipes
= \
283 options
.shr_mem_pipes_per_cu
,
284 n_wf
= options
.wfs_per_simd
,
285 execPolicy
= options
.CUExecPolicy
,
286 debugSegFault
= options
.SegFaultDebug
,
287 functionalTLB
= options
.FunctionalTLB
,
288 localMemBarrier
= options
.LocalMemBarrier
,
289 countPages
= options
.countPages
,
291 LdsState(banks
= options
.numLdsBanks
,
292 bankConflictPenalty
= \
293 options
.ldsBankConflictPenalty
)))
299 for j
in xrange(options
.simds_per_cu
):
300 for k
in xrange(shader
.n_wf
):
301 wavefronts
.append(Wavefront(simdId
= j
, wf_slot_id
= k
,
302 wf_size
= options
.wf_size
))
303 vrf_pool_mgrs
.append(SimplePoolManager(pool_size
= \
304 options
.vreg_file_size
,
306 options
.vreg_min_alloc
))
308 vrfs
.append(VectorRegisterFile(simd_id
=j
, wf_size
=options
.wf_size
,
309 num_regs
=options
.vreg_file_size
))
311 srf_pool_mgrs
.append(SimplePoolManager(pool_size
= \
312 options
.sreg_file_size
,
314 options
.vreg_min_alloc
))
315 srfs
.append(ScalarRegisterFile(simd_id
=j
, wf_size
=options
.wf_size
,
316 num_regs
=options
.sreg_file_size
))
318 compute_units
[-1].wavefronts
= wavefronts
319 compute_units
[-1].vector_register_file
= vrfs
320 compute_units
[-1].scalar_register_file
= srfs
321 compute_units
[-1].register_manager
= \
322 RegisterManager(policy
=options
.registerManagerPolicy
,
323 vrf_pool_managers
=vrf_pool_mgrs
,
324 srf_pool_managers
=srf_pool_mgrs
)
325 if options
.TLB_prefetch
:
326 compute_units
[-1].prefetch_depth
= options
.TLB_prefetch
327 compute_units
[-1].prefetch_prev_type
= options
.pf_type
329 # attach the LDS and the CU to the bus (actually a Bridge)
330 compute_units
[-1].ldsPort
= compute_units
[-1].ldsBus
.slave
331 compute_units
[-1].ldsBus
.master
= compute_units
[-1].localDataStore
.cuPort
333 # Attach compute units to GPU
334 shader
.CUs
= compute_units
336 ########################## Creating the CPU system ########################
337 options
.num_cpus
= options
.num_cpus
339 # The shader core will be whatever is after the CPU cores are accounted for
340 shader_idx
= options
.num_cpus
342 # The command processor will be whatever is after the shader is accounted for
343 cp_idx
= shader_idx
+ 1
349 CpuClass
, mem_mode
= Simulation
.getCPUClass(options
.cpu_type
)
350 if CpuClass
== AtomicSimpleCPU
:
351 fatal("AtomicSimpleCPU is not supported")
352 if mem_mode
!= 'timing':
353 fatal("Only the timing memory mode is supported")
356 if options
.fast_forward
and options
.fast_forward_pseudo_op
:
357 fatal("Cannot fast-forward based both on the number of instructions and"
359 fast_forward
= options
.fast_forward
or options
.fast_forward_pseudo_op
362 FutureCpuClass
, future_mem_mode
= CpuClass
, mem_mode
365 mem_mode
= 'atomic_noncaching'
366 # Leave shader.timing untouched, because its value only matters at the
367 # start of the simulation and because we require switching cpus
368 # *before* the first kernel launch.
372 # Initial CPUs to be used during fast-forwarding.
373 for i
in range(options
.num_cpus
):
374 cpu
= CpuClass(cpu_id
= i
,
375 clk_domain
= SrcClockDomain(
376 clock
= options
.CPUClock
,
377 voltage_domain
= VoltageDomain(
378 voltage
= options
.cpu_voltage
)))
381 if options
.fast_forward
:
382 cpu
.max_insts_any_thread
= int(options
.fast_forward
)
385 MainCpuClass
= FutureCpuClass
387 MainCpuClass
= CpuClass
389 # CPs to be used throughout the simulation.
390 for i
in range(options
.num_cp
):
391 cp
= MainCpuClass(cpu_id
= options
.num_cpus
+ i
,
392 clk_domain
= SrcClockDomain(
393 clock
= options
.CPUClock
,
394 voltage_domain
= VoltageDomain(
395 voltage
= options
.cpu_voltage
)))
398 # Main CPUs (to be used after fast-forwarding if fast-forwarding is specified).
399 for i
in range(options
.num_cpus
):
400 cpu
= MainCpuClass(cpu_id
= i
,
401 clk_domain
= SrcClockDomain(
402 clock
= options
.CPUClock
,
403 voltage_domain
= VoltageDomain(
404 voltage
= options
.cpu_voltage
)))
406 cpu
.switched_out
= True
407 future_cpu_list
.append(cpu
)
411 host_cpu
= cpu_list
[0]
413 hsapp_gpu_map_vaddr
= 0x200000000
414 hsapp_gpu_map_size
= 0x1000
415 hsapp_gpu_map_paddr
= int(Addr(options
.mem_size
))
417 # HSA kernel mode driver
418 gpu_driver
= GPUComputeDriver(filename
="kfd")
420 # Creating the GPU kernel launching components: that is the HSA
421 # packet processor (HSAPP), GPU command processor (CP), and the
423 gpu_hsapp
= HSAPacketProcessor(pioAddr
=hsapp_gpu_map_paddr
,
424 numHWQueues
=options
.num_hw_queues
)
425 dispatcher
= GPUDispatcher()
426 gpu_cmd_proc
= GPUCommandProcessor(hsapp
=gpu_hsapp
,
427 dispatcher
=dispatcher
)
428 gpu_driver
.device
= gpu_cmd_proc
429 shader
.dispatcher
= dispatcher
430 shader
.gpu_cmd_proc
= gpu_cmd_proc
432 # Create and assign the workload Check for rel_path in elements of
433 # base_list using test, returning the first full path that satisfies test
434 def find_path(base_list
, rel_path
, test
):
435 for base
in base_list
:
437 # base could be None if environment var not set
439 full_path
= os
.path
.join(base
, rel_path
)
442 fatal("%s not found in %s" % (rel_path
, base_list
))
444 def find_file(base_list
, rel_path
):
445 return find_path(base_list
, rel_path
, os
.path
.isfile
)
447 executable
= find_path(benchmark_path
, options
.cmd
, os
.path
.exists
)
448 # It's common for a benchmark to be in a directory with the same
449 # name as the executable, so we handle that automatically
450 if os
.path
.isdir(executable
):
451 benchmark_path
= [executable
]
452 executable
= find_file(benchmark_path
, options
.cmd
)
455 with
open(options
.env
, 'r') as f
:
456 env
= [line
.rstrip() for line
in f
]
458 env
= ['LD_LIBRARY_PATH=%s' % ':'.join([
459 "/proj/radl_tools/rocm-1.6/lib",
460 "/proj/radl_tools/rocm-1.6/hcc/lib64",
461 "/tool/pandora64/.package/libunwind-1.1/lib",
462 "/tool/pandora64/.package/gcc-6.4.0/lib64"
464 "HSA_ENABLE_INTERRUPT=0"]
466 process
= Process(executable
= executable
, cmd
= [options
.cmd
]
467 + options
.options
.split(), drivers
= [gpu_driver
], env
= env
)
471 cpu
.workload
= process
474 cp
.workload
= host_cpu
.workload
477 for i
in range(len(future_cpu_list
)):
478 future_cpu_list
[i
].workload
= cpu_list
[i
].workload
479 future_cpu_list
[i
].createThreads()
481 ########################## Create the overall system ########################
482 # List of CPUs that must be switched when moving between KVM and simulation
485 [(cpu_list
[i
], future_cpu_list
[i
]) for i
in range(options
.num_cpus
)]
487 # Full list of processing cores in the system.
488 cpu_list
= cpu_list
+ [shader
] + cp_list
490 # creating the overall system
491 # notice the cpu list is explicitly added as a parameter to System
492 system
= System(cpu
= cpu_list
,
493 mem_ranges
= [AddrRange(options
.mem_size
)],
494 cache_line_size
= options
.cacheline_size
,
497 system
.future_cpu
= future_cpu_list
498 system
.voltage_domain
= VoltageDomain(voltage
= options
.sys_voltage
)
499 system
.clk_domain
= SrcClockDomain(clock
= options
.sys_clock
,
500 voltage_domain
= system
.voltage_domain
)
503 have_kvm_support
= 'BaseKvmCPU' in globals()
504 if have_kvm_support
and buildEnv
['TARGET_ISA'] == "x86":
506 for i
in range(len(host_cpu
.workload
)):
507 host_cpu
.workload
[i
].useArchPT
= True
508 host_cpu
.workload
[i
].kvmInSE
= True
510 fatal("KvmCPU can only be used in SE mode with x86")
512 # configure the TLB hierarchy
513 GPUTLBConfig
.config_tlb_hierarchy(options
, system
, shader_idx
)
516 system
.piobus
= IOXBar(width
=32, response_latency
=0,
517 frontend_latency
=0, forward_latency
=0)
518 dma_list
= [gpu_hsapp
, gpu_cmd_proc
]
519 Ruby
.create_system(options
, None, system
, None, dma_list
, None)
520 system
.ruby
.clk_domain
= SrcClockDomain(clock
= options
.ruby_clock
,
521 voltage_domain
= system
.voltage_domain
)
522 gpu_cmd_proc
.pio
= system
.piobus
.master
523 gpu_hsapp
.pio
= system
.piobus
.master
525 for i
, dma_device
in enumerate(dma_list
):
526 exec('system.dma_cntrl%d.clk_domain = system.ruby.clk_domain' % i
)
528 # attach the CPU ports to Ruby
529 for i
in range(options
.num_cpus
):
530 ruby_port
= system
.ruby
._cpu
_ports
[i
]
532 # Create interrupt controller
533 system
.cpu
[i
].createInterruptController()
535 # Connect cache port's to ruby
536 system
.cpu
[i
].icache_port
= ruby_port
.slave
537 system
.cpu
[i
].dcache_port
= ruby_port
.slave
539 ruby_port
.mem_master_port
= system
.piobus
.slave
540 if buildEnv
['TARGET_ISA'] == "x86":
541 system
.cpu
[i
].interrupts
[0].pio
= system
.piobus
.master
542 system
.cpu
[i
].interrupts
[0].int_master
= system
.piobus
.slave
543 system
.cpu
[i
].interrupts
[0].int_slave
= system
.piobus
.master
545 system
.cpu
[i
].itb
.walker
.port
= ruby_port
.slave
546 system
.cpu
[i
].dtb
.walker
.port
= ruby_port
.slave
548 # attach CU ports to Ruby
549 # Because of the peculiarities of the CP core, you may have 1 CPU but 2
550 # sequencers and thus 2 _cpu_ports created. Your GPUs shouldn't be
551 # hooked up until after the CP. To make this script generic, figure out
552 # the index as below, but note that this assumes there is one sequencer
553 # per compute unit and one sequencer per SQC for the math to work out
555 gpu_port_idx
= len(system
.ruby
._cpu
_ports
) \
556 - options
.num_compute_units
- options
.num_sqc \
557 - options
.num_scalar_cache
558 gpu_port_idx
= gpu_port_idx
- options
.num_cp
* 2
560 wavefront_size
= options
.wf_size
561 for i
in range(n_cu
):
562 # The pipeline issues wavefront_size number of uncoalesced requests
563 # in one GPU issue cycle. Hence wavefront_size mem ports.
564 for j
in range(wavefront_size
):
565 system
.cpu
[shader_idx
].CUs
[i
].memory_port
[j
] = \
566 system
.ruby
._cpu
_ports
[gpu_port_idx
].slave
[j
]
569 for i
in range(n_cu
):
570 if i
> 0 and not i
% options
.cu_per_sqc
:
571 print("incrementing idx on ", i
)
573 system
.cpu
[shader_idx
].CUs
[i
].sqc_port
= \
574 system
.ruby
._cpu
_ports
[gpu_port_idx
].slave
575 gpu_port_idx
= gpu_port_idx
+ 1
577 for i
in xrange(n_cu
):
578 if i
> 0 and not i
% options
.cu_per_scalar_cache
:
579 print("incrementing idx on ", i
)
581 system
.cpu
[shader_idx
].CUs
[i
].scalar_port
= \
582 system
.ruby
._cpu
_ports
[gpu_port_idx
].slave
583 gpu_port_idx
= gpu_port_idx
+ 1
585 # attach CP ports to Ruby
586 for i
in range(options
.num_cp
):
587 system
.cpu
[cp_idx
].createInterruptController()
588 system
.cpu
[cp_idx
].dcache_port
= \
589 system
.ruby
._cpu
_ports
[gpu_port_idx
+ i
* 2].slave
590 system
.cpu
[cp_idx
].icache_port
= \
591 system
.ruby
._cpu
_ports
[gpu_port_idx
+ i
* 2 + 1].slave
592 system
.cpu
[cp_idx
].interrupts
[0].pio
= system
.piobus
.master
593 system
.cpu
[cp_idx
].interrupts
[0].int_master
= system
.piobus
.slave
594 system
.cpu
[cp_idx
].interrupts
[0].int_slave
= system
.piobus
.master
597 ################# Connect the CPU and GPU via GPU Dispatcher ##################
598 # CPU rings the GPU doorbell to notify a pending task
599 # using this interface.
600 # And GPU uses this interface to notify the CPU of task completion
601 # The communcation happens through emulated driver.
603 # Note this implicit setting of the cpu_pointer, shader_pointer and tlb array
604 # parameters must be after the explicit setting of the System cpu list
606 shader
.cpu_pointer
= future_cpu_list
[0]
608 shader
.cpu_pointer
= host_cpu
610 ########################## Start simulation ########################
612 chroot
= os
.path
.expanduser(options
.chroot
)
613 redirect_paths
= [RedirectPath(src
= "/proc",
614 dests
= ["%s/fs/proc" % m5
.options
.outdir
]),
615 RedirectPath(src
= "/sys",
616 dests
= ["%s/fs/sys" % m5
.options
.outdir
]),
617 RedirectPath(src
= "/tmp",
618 dests
= ["%s/fs/tmp" % m5
.options
.outdir
]),
619 RedirectPath(src
= "/dev/shm",
620 dests
= ["/dev/shm/%s/gem5_%s" %
621 (getpass
.getuser(), os
.getpid())])]
623 system
.redirect_paths
= redirect_paths
625 root
= Root(system
=system
, full_system
=False)
627 hsaTopology
.createHsaTopology(options
)
629 m5
.ticks
.setGlobalFrequency('1THz')
630 if options
.abs_max_tick
:
631 maxtick
= options
.abs_max_tick
635 # Benchmarks support work item annotations
636 Simulation
.setWorkCountOptions(system
, options
)
638 # Checkpointing is not supported by APU model
639 if (options
.checkpoint_dir
!= None or
640 options
.checkpoint_restore
!= None):
641 fatal("Checkpointing not supported by apu model")
643 checkpoint_dir
= None
644 m5
.instantiate(checkpoint_dir
)
646 # Map workload to this address space
647 host_cpu
.workload
[0].map(0x10000000, 0x200000000, 4096)
649 if options
.fast_forward
:
650 print("Switch at instruction count: %d" % cpu_list
[0].max_insts_any_thread
)
652 exit_event
= m5
.simulate(maxtick
)
654 if options
.fast_forward
:
655 if exit_event
.getCause() == "a thread reached the max instruction count":
656 m5
.switchCpus(system
, switch_cpu_list
)
657 print("Switched CPUS @ tick %s" % (m5
.curTick()))
659 exit_event
= m5
.simulate(maxtick
- m5
.curTick())
660 elif options
.fast_forward_pseudo_op
:
661 while exit_event
.getCause() == "switchcpu":
662 # If we are switching *to* kvm, then the current stats are meaningful
663 # Note that we don't do any warmup by default
664 if type(switch_cpu_list
[0][0]) == FutureCpuClass
:
665 print("Dumping stats...")
667 m5
.switchCpus(system
, switch_cpu_list
)
668 print("Switched CPUS @ tick %s" % (m5
.curTick()))
670 # This lets us switch back and forth without keeping a counter
671 switch_cpu_list
= [(x
[1], x
[0]) for x
in switch_cpu_list
]
672 exit_event
= m5
.simulate(maxtick
- m5
.curTick())
674 print("Ticks:", m5
.curTick())
675 print('Exiting because ', exit_event
.getCause())
677 FileSystemConfig
.cleanup_filesystem(options
)
679 sys
.exit(exit_event
.getCode())