def create(args):
# System
CpuClass, mem_mode, _ = Simulation.setCPUClass(args)
sys_cfg = Benchmarks.SysConfig(args.script, args.mem_size)
system = FSConfig.makeArmSystem(mem_mode,
"VExpress_GEM5_V2",
args.num_cpus,
sys_cfg,
bare_metal=True,
security=True)
system.voltage_domain = VoltageDomain(voltage=args.sys_voltage)
system.clk_domain = SrcClockDomain(clock=args.sys_clock,
voltage_domain=system.voltage_domain)
system.highest_el_is_64 = True
system.have_virtualization = True
system.workload.object_file = args.kernel
# CPU cluster
system.cpu_voltage_domain = VoltageDomain()
system.cpu_clk_domain = SrcClockDomain(
clock=args.cpu_clock, voltage_domain=system.cpu_voltage_domain)
system.cpu = [
CpuClass(clk_domain=system.cpu_clk_domain, cpu_id=i)
for i in range(args.num_cpus)
]
for cpu in system.cpu:
cpu.createThreads()
# (gem5 v20.1) Disable FEAT_VHE, prevents booting
features = cpu.isa[0].id_aa64mmfr1_el1.getValue()
cpu.isa[0].id_aa64mmfr1_el1 = features & ~0xf00
CacheConfig.config_cache(args, system)
# Devices
system.realview.atp_adapter = ProfileGen(config_files=args.atp_file,
exit_when_done=False,
init_only=True,
disable_watchdog=True,
disable_mem_check=True)
system.realview.atp_device = ATPDevice(pio_addr=0x2b500000,
interrupt=ArmSPI(num=104),
atp_id="STREAM")
system.realview.attachSmmu([system.realview.atp_device], system.membus)
# (gem5 v20.1) Ensure 128 CMDQ entries for compatibility from Linux v5.4
system.realview.smmu.smmu_idr1 = 0x00E00000
# (gem5 v20.2+) Enable SMMUv3 interrupt interface to boot Linux
if hasattr(system.realview.smmu, "irq_interface_enable"):
system.realview.smmu.irq_interface_enable = True
connect_adapter(system.realview.atp_adapter, system.realview.smmu)
if args.disk_image:
system.disk = [
PciVirtIO(vio=VirtIOBlock(image=create_cow_image(disk)))
for disk in args.disk_image
]
for disk in system.disk:
system.realview.attachPciDevice(disk, system.iobus)
# Memory
MemConfig.config_mem(args, system)
return system
else:
fatal("This test is only for FPGA in Ruby. Please set --ruby.\n")
(options, args) = parser.parse_args()
if args:
print "Error: script doesn't take any positional arguments"
sys.exit(1)
numThreads = 1
process1 = LiveProcess()
process1.pid = 1100
process1.cmd = ['tests/test-progs/polybench-c-4.2/fdtd_ref']
(CPUClass, test_mem_mode, FutureClass) = Simulation.setCPUClass(options)
CPUClass.numThreads = numThreads
np = options.num_cpus
system = System(
cpu=[DerivO3CPU() for i in xrange(np)],
#system = System(cpu = [TimingSimpleCPU() for i in xrange(np)],
mem_mode='timing',
mem_ranges=[AddrRange('512MB')],
cache_line_size=64)
system.fpga = [FpgaCPU() for i in xrange(options.num_fpgas)]
system.voltage_domain = VoltageDomain(voltage=options.sys_voltage)
system.clk_domain = SrcClockDomain(clock=options.sys_clock,
voltage_domain=system.voltage_domain)
system.cpu_voltage_domain = VoltageDomain()
if args:
print("Error: script doesn't take any positional arguments")
sys.exit(1)
numThreads = 1
if options.cpu_type != "TraceCPU":
fatal("This is a script for elastic trace replay simulation, use "\
"--cpu-type=TraceCPU\n");
if options.num_cpus > 1:
fatal("This script does not support multi-processor trace replay.\n")
# In this case FutureClass will be None as there is not fast forwarding or
# switching
(CPUClass, test_mem_mode, FutureClass) = Simulation.setCPUClass(options)
CPUClass.numThreads = numThreads
system = System(cpu = CPUClass(cpu_id=0),
mem_mode = test_mem_mode,
mem_ranges = [AddrRange(options.mem_size)],
cache_line_size = options.cacheline_size)
# Create a top-level voltage domain
system.voltage_domain = VoltageDomain(voltage = options.sys_voltage)
# Create a source clock for the system. This is used as the clock period for
# xbar and memory
system.clk_domain = SrcClockDomain(clock = options.sys_clock,
voltage_domain = system.voltage_domain)
return drive_sys
# Add args
parser = argparse.ArgumentParser()
Options.addCommonOptions(parser)
Options.addFSOptions(parser)
# Add the ruby specific and protocol specific args
if '--ruby' in sys.argv:
Ruby.define_options(parser)
args = parser.parse_args()
# system under test can be any CPU
(TestCPUClass, test_mem_mode, FutureClass) = Simulation.setCPUClass(args)
# Match the memories with the CPUs, based on the options for the test system
TestMemClass = Simulation.setMemClass(args)
if args.benchmark:
try:
bm = Benchmarks[args.benchmark]
except KeyError:
print("Error benchmark %s has not been defined." % args.benchmark)
print("Valid benchmarks are: %s" % DefinedBenchmarks)
sys.exit(1)
else:
if args.dual:
bm = [
SysConfig(disks=args.disk_image,
parser.add_argument(
"--bare-metal",
action="store_true",
help="Provide the raw system without the linux specific bits")
parser.add_argument("--dtb-filename", action="store", type=str,
help="Specifies device tree blob file to use with device-tree-"\
"enabled kernels")
parser.add_argument("--virtio-rng",
action="store_true",
help="Enable VirtIORng device")
# ---------------------------- Parse Options --------------------------- #
args = parser.parse_args()
# CPU and Memory
(CPUClass, mem_mode, FutureClass) = Simulation.setCPUClass(args)
MemClass = Simulation.setMemClass(args)
np = args.num_cpus
# ---------------------------- Setup System ---------------------------- #
# Default Setup
system = System()
mdesc = SysConfig(disks=args.disk_image,
rootdev=args.root_device,
mem=args.mem_size,
os_type=args.os_type)
system.mem_mode = mem_mode
system.mem_ranges = [AddrRange(start=0x80000000, size=mdesc.mem())]
if args.bare_metal:
else:
exec("workload = %s(buildEnv['TARGET_ISA', 'linux', '%s')" %
(app, options.spec_input))
multiprocesses.append(workload.makeLiveProcess())
except:
print >> sys.stderr, "Unable to find workload for %s: %s" % (
buildEnv['TARGET_ISA'], app)
sys.exit(1)
elif options.cmd:
multiprocesses, numThreads = get_processes(options)
else:
print >> sys.stderr, "No workload specified. Exiting!\n"
sys.exit(1)
(CPUClass, test_mem_mode,
FutureClass) = Simulation.setCPUClass(options) #------
CPUClass.numThreads = numThreads
# Check -- do not allow SMT with multiple CPUs
if options.smt and options.num_cpus > 1:
fatal("You cannot use SMT with multiple CPUs!")
np = options.num_cpus
system = System(cpu=[CPUClass(cpu_id=i) for i in xrange(np)],
mem_mode=test_mem_mode,
mem_ranges=[AddrRange(options.mem_size)],
cache_line_size=options.cacheline_size)
if numThreads > 1:
system.multi_thread = True
Options.addFSOptions(parser)
Options.addSPMOptions(parser)
# Add the ruby specific and protocol specific options
if '--ruby' in sys.argv:
Ruby.define_options(parser)
(options, args) = parser.parse_args()
if args:
print("Error: script doesn't take any positional arguments")
sys.exit(1)
# system under test can be any CPU
(TestCPUClass, test_mem_mode, FutureClass,
FutureClass2) = Simulation.setCPUClass(options)
# Match the memories with the CPUs, based on the options for the test system
TestMemClass = Simulation.setMemClass(options)
if options.benchmark:
try:
bm = Benchmarks[options.benchmark]
except KeyError:
print("Error benchmark %s has not been defined." % options.benchmark)
print("Valid benchmarks are: %s" % DefinedBenchmarks)
sys.exit(1)
else:
if options.dual:
bm = [
SysConfig(disk=options.disk_image,
def makeGpuFSSystem(args):
# Boot options are standard gem5 options plus:
# - Framebuffer device emulation 0 to reduce driver code paths.
# - Blacklist amdgpu as it cannot (currently) load in KVM CPU.
# - Blacklist psmouse as amdgpu driver adds proprietary commands that
# cause gem5 to panic.
boot_options = ['earlyprintk=ttyS0', 'console=ttyS0,9600',
'lpj=7999923', 'root=/dev/sda1',
'drm_kms_helper.fbdev_emulation=0',
'modprobe.blacklist=amdgpu',
'modprobe.blacklist=psmouse']
cmdline = ' '.join(boot_options)
if MemorySize(args.mem_size) < MemorySize('2GB'):
panic("Need at least 2GB of system memory to load amdgpu module")
# Use the common FSConfig to setup a Linux X86 System
(TestCPUClass, test_mem_mode, FutureClass) = Simulation.setCPUClass(args)
bm = SysConfig(disks=[args.disk_image], mem=args.mem_size)
system = makeLinuxX86System(test_mem_mode, args.num_cpus, bm, True,
cmdline=cmdline)
system.workload.object_file = binary(args.kernel)
# Set the cache line size for the entire system.
system.cache_line_size = args.cacheline_size
# Create a top-level voltage and clock domain.
system.voltage_domain = VoltageDomain(voltage = args.sys_voltage)
system.clk_domain = SrcClockDomain(clock = args.sys_clock,
voltage_domain = system.voltage_domain)
# Create a CPU voltage and clock domain.
system.cpu_voltage_domain = VoltageDomain()
system.cpu_clk_domain = SrcClockDomain(clock = args.cpu_clock,
voltage_domain =
system.cpu_voltage_domain)
# Setup VGA ROM region
system.shadow_rom_ranges = [AddrRange(0xc0000, size = Addr('128kB'))]
# Create specified number of CPUs. GPUFS really only needs one.
system.cpu = [TestCPUClass(clk_domain=system.cpu_clk_domain, cpu_id=i)
for i in range(args.num_cpus)]
if ObjectList.is_kvm_cpu(TestCPUClass) or \
ObjectList.is_kvm_cpu(FutureClass):
system.kvm_vm = KvmVM()
# Create AMDGPU and attach to southbridge
shader = createGPU(system, args)
connectGPU(system, args)
# This arbitrary address is something in the X86 I/O hole
hsapp_gpu_map_paddr = 0xe00000000
gpu_hsapp = HSAPacketProcessor(pioAddr=hsapp_gpu_map_paddr,
numHWQueues=args.num_hw_queues)
dispatcher = GPUDispatcher()
gpu_cmd_proc = GPUCommandProcessor(hsapp=gpu_hsapp,
dispatcher=dispatcher)
shader.dispatcher = dispatcher
shader.gpu_cmd_proc = gpu_cmd_proc
# GPU, HSAPP, and GPUCommandProc are DMA devices
system._dma_ports.append(gpu_hsapp)
system._dma_ports.append(gpu_cmd_proc)
system._dma_ports.append(system.pc.south_bridge.gpu)
gpu_hsapp.pio = system.iobus.mem_side_ports
gpu_cmd_proc.pio = system.iobus.mem_side_ports
system.pc.south_bridge.gpu.pio = system.iobus.mem_side_ports
# Create Ruby system using Ruby.py for now
Ruby.create_system(args, True, system, system.iobus,
system._dma_ports)
# Create a seperate clock domain for Ruby
system.ruby.clk_domain = SrcClockDomain(clock = args.ruby_clock,
voltage_domain = system.voltage_domain)
for (i, cpu) in enumerate(system.cpu):
#
# Tie the cpu ports to the correct ruby system ports
#
cpu.clk_domain = system.cpu_clk_domain
cpu.createThreads()
cpu.createInterruptController()
system.ruby._cpu_ports[i].connectCpuPorts(cpu)
# The shader core will be whatever is after the CPU cores are accounted for
shader_idx = args.num_cpus
system.cpu.append(shader)
gpu_port_idx = len(system.ruby._cpu_ports) \
- args.num_compute_units - args.num_sqc \
- args.num_scalar_cache
gpu_port_idx = gpu_port_idx - args.num_cp * 2
# Connect token ports. For this we need to search through the list of all
# sequencers, since the TCP coalescers will not necessarily be first. Only
# TCP coalescers use a token port for back pressure.
token_port_idx = 0
for i in range(len(system.ruby._cpu_ports)):
if isinstance(system.ruby._cpu_ports[i], VIPERCoalescer):
system.cpu[shader_idx].CUs[token_port_idx].gmTokenPort = \
system.ruby._cpu_ports[i].gmTokenPort
token_port_idx += 1
wavefront_size = args.wf_size
for i in range(args.num_compute_units):
# The pipeline issues wavefront_size number of uncoalesced requests
# in one GPU issue cycle. Hence wavefront_size mem ports.
for j in range(wavefront_size):
system.cpu[shader_idx].CUs[i].memory_port[j] = \
system.ruby._cpu_ports[gpu_port_idx].in_ports[j]
gpu_port_idx += 1
for i in range(args.num_compute_units):
if i > 0 and not i % args.cu_per_sqc:
gpu_port_idx += 1
system.cpu[shader_idx].CUs[i].sqc_port = \
system.ruby._cpu_ports[gpu_port_idx].in_ports
gpu_port_idx = gpu_port_idx + 1
for i in range(args.num_compute_units):
if i > 0 and not i % args.cu_per_scalar_cache:
gpu_port_idx += 1
system.cpu[shader_idx].CUs[i].scalar_port = \
system.ruby._cpu_ports[gpu_port_idx].in_ports
gpu_port_idx = gpu_port_idx + 1
return system