work in progress

This commit is contained in:
Julian Hammer
2019-10-14 17:08:40 +02:00
parent 85dd53dc4e
commit 1c673382b4
3 changed files with 118 additions and 164 deletions

View File

@@ -4,115 +4,29 @@ import re
import sys
import xml.etree.ElementTree as ET
from distutils.version import StrictVersion
from itertools import groupby, product
from ruamel import yaml
from osaca.db_interface import add_entries_to_db
from osaca.parser import ParserAArch64v81, ParserX86ATT
from osaca.parser import get_parser
from osaca.semantics import MachineModel
ARCH_DICT = {
'vulcan': 'aarch64',
'snb': 'x86',
'ivb': 'x86',
'hsw': 'x86',
'bdw': 'x86',
'skl': 'x86',
'skx': 'x86',
'csx': 'x86',
}
def port_pressure_from_tag_attributes(attrib):
# '1*p015+1*p1+1*p23+1*p4+3*p5' ->
# [(1, '015'), (1, '1'), (1, '23'), (1, '4'), (3, '5')]
port_occupation = []
for p in attrib['ports'].split('+'):
cycles, ports = p.split('*p')
port_occupation.append((int(cycles), ports))
def port_pressure_from_tag_attributes(attrib, arch, ports):
# apply cycles for D ports
data_port = re.compile(r'[0-9]D$')
data_ports = [x[:-1] for x in filter(data_port.match, ports)]
# format attributes
cycles = attrib['ports'].split('+')
cycles = [c.split('*') for c in cycles]
for i, c in enumerate(cycles):
cycles[i][0] = int(c[0])
if str(c[1]).startswith('p'):
cycles[i][1] = [p for p in c[1][1:]]
if data_ports and data_ports == cycles[i][1]:
# uops for data ports
cycles.append([c[0], [x + 'D' for x in data_ports]])
cycles[i][0] = [
cycles[i][0] / num for num in range(1, len(cycles[i][1]) + 1) for _ in range(num)
]
cycles = [list(product(c[0], c[1])) for c in cycles]
all_options = []
# iterate over all combinations of all uop options
for cycles_combs in cycles:
options = []
tmp_opt = []
total = cycles_combs[0][0]
# iterate over all combinations of each uop option
for comb in cycles_combs:
# add options until they reach the total num of uops
tmp_opt.append(comb)
if sum([c[0] for c in tmp_opt]) == total:
# copy this option as one of several to the cycle option list
options.append(tmp_opt.copy())
tmp_opt = []
if len(tmp_opt) != 0:
raise ValueError('Cannot compute port pressure')
options = [x for x, _ in groupby(options)]
all_options.append(options)
all_options = list(product(*all_options))
# find best scheduling
port_pressure = {}
for p in ports:
port_pressure[p] = 0.0
first = calculate_port_pressure(all_options[0])
for key in first:
port_pressure[key] = first[key]
for option in all_options[1:]:
tmp = calculate_port_pressure(option)
if (max(list(tmp.values())) <= max(list(port_pressure.values()))) and (
len(tmp) > len([x for x in port_pressure.values() if x != 0.0])
):
for k in port_pressure:
port_pressure[k] = tmp[k] if k in tmp else 0.0
# check if calculation equals given throughput
if abs(max(list(port_pressure.values())) - float(attrib['TP_ports'])) > 0.01:
print('Contradicting TP value compared to port_pressure. Ignore port pressure.')
for p in port_pressure:
port_pressure[p] = 0.0
return port_pressure
# Also consider DIV pipeline
# Also
if 'div_cycles' in attrib:
div_port = re.compile(r'[0-9]DV$')
div_ports = [x for x in filter(div_port.match, ports)]
for dp in div_ports:
port_pressure[dp] += int(attrib['div_cycles']) / len(div_ports)
return port_pressure
port_occupation.append((int(attrib['div_cycles']), ('DV',)))
return port_occupation
def calculate_port_pressure(pp_option):
ports = {}
for option in pp_option:
for port in option:
if port[1] in ports:
ports[port[1]] += port[0]
else:
ports[port[1]] = port[0]
return ports
def extract_paramters(instruction_tag, arch):
isa = ARCH_DICT[arch.lower()]
parser = ParserX86ATT()
if isa == 'aarch64':
parser = ParserAArch64v81()
elif isa == 'x86':
parser = ParserX86ATT()
def extract_paramters(instruction_tag, isa):
parser = get_parser(isa)
# Extract parameter components
parameters = [] # used to store string representations
parameter_tags = sorted(instruction_tag.findall("operand"), key=lambda p: int(p.attrib['idx']))
@@ -138,16 +52,12 @@ def extract_paramters(instruction_tag, arch):
parameter['class'] = 'register'
possible_regs = [parser.parse_register('%' + r) for r in parameter_tag.text.split(',')]
if possible_regs[0] is None:
raise ValueError(
'Unknown register type for {} with {}.'.format(
parameter_tag.attrib, parameter_tag.text
)
)
raise ValueError('Unknown register type for {} with {}.'.format(
parameter_tag.attrib, parameter_tag.text))
if isa == 'x86':
if parser.is_vector_register(possible_regs[0]['register']):
possible_regs[0]['register']['name'] = possible_regs[0]['register'][
'name'
].lower()[:3]
possible_regs[0]['register']['name'] = \
possible_regs[0]['register']['name'].lower()[:3]
if 'mask' in possible_regs[0]['register']:
possible_regs[0]['register']['mask'] = True
else:
@@ -175,9 +85,8 @@ def extract_paramters(instruction_tag, arch):
def extract_model(tree, arch):
mm = MachineModel(arch.lower())
ports = mm._data['ports']
model_data = []
mm = MachineModel()
for instruction_tag in tree.findall('.//instruction'):
ignore = False
@@ -185,8 +94,8 @@ def extract_model(tree, arch):
# Extract parameter components
try:
parameters = extract_paramters(instruction_tag, arch)
if ARCH_DICT[arch.lower()] == 'x86':
parameters = extract_paramters(instruction_tag, mm.get_isa_for_arch(arch))
if mm.get_isa_for_arch(arch).lower() == 'x86':
parameters.reverse()
except ValueError as e:
print(e, file=sys.stderr)
@@ -201,16 +110,12 @@ def extract_model(tree, arch):
if 'TP_ports' in measurement_tag.attrib:
throughput = measurement_tag.attrib['TP_ports']
else:
throughput = (
measurement_tag.attrib['TP'] if 'TP' in measurement_tag.attrib else None
)
uops = (
int(measurement_tag.attrib['uops']) if 'uops' in measurement_tag.attrib else None
)
throughput = (measurement_tag.attrib['TP']
if 'TP' in measurement_tag.attrib else None)
uops = (int(measurement_tag.attrib['uops'])
if 'uops' in measurement_tag.attrib else None)
if 'ports' in measurement_tag.attrib:
port_pressure.append(
port_pressure_from_tag_attributes(measurement_tag.attrib, arch, ports)
)
port_pressure.append(port_pressure_from_tag_attributes(measurement_tag.attrib))
latencies = [
int(l_tag.attrib['cycles'])
for l_tag in measurement_tag.iter('latency')
@@ -230,39 +135,30 @@ def extract_model(tree, arch):
# Ordered by IACA version (newest last)
for iaca_tag in sorted(
arch_tag.iter('IACA'), key=lambda i: StrictVersion(i.attrib['version'])
):
arch_tag.iter('IACA'), key=lambda i: StrictVersion(i.attrib['version'])):
if 'ports' in iaca_tag.attrib:
port_pressure.append(
port_pressure_from_tag_attributes(iaca_tag.attrib, arch, ports)
)
port_pressure.append(port_pressure_from_tag_attributes(iaca_tag.attrib))
if ignore:
continue
# Add missing ports:
[p[1] for p in for pp in port_pressure]
mm.add_port()
# Check if all are equal
if port_pressure:
if port_pressure[1:] != port_pressure[:-1]:
print(
"Contradicting port occupancies, using latest IACA:", mnemonic, file=sys.stderr
)
print("Contradicting port occupancies, using latest IACA:", mnemonic,
file=sys.stderr)
port_pressure = port_pressure[-1]
throughput = max(list(port_pressure.values()) + [0.0])
throughput = max(mm.average_port_pressure(port_pressure))
else:
# print("No data available for this architecture:", mnemonic, file=sys.stderr)
continue
# ---------------------------------------------
model_data.append(
{
'name': mnemonic,
'operands': parameters,
'uops': uops,
'throughput': throughput,
'latency': latency,
'port_pressure': port_pressure,
}
)
mm.set_instruction(mnemonic, parameters, latency, port_pressure, throughput, uops)
return model_data
return mm
def architectures(tree):
@@ -282,12 +178,10 @@ def main():
tree = ET.parse(args.xml)
if args.arch:
model_data = extract_model(tree, args.arch)
print(yaml.dump(model_data, allow_unicode=True))
model = extract_model(tree, args.arch)
print(model.dump())
else:
for arch in architectures(tree):
model_data = extract_model(tree, arch)
add_entries_to_db(arch, model_data)
raise NotImplementedError()
if __name__ == '__main__':

View File

@@ -8,4 +8,12 @@ from .base_parser import BaseParser
from .parser_x86att import ParserX86ATT
from .parser_AArch64v81 import ParserAArch64v81
__all__ = ['AttrDict', 'BaseParser', 'ParserX86ATT', 'ParserAArch64v81']
__all__ = ['AttrDict', 'BaseParser', 'ParserX86ATT', 'ParserAArch64v81', 'get_parser']
def get_parser(isa):
if isa.lower() == 'x86':
return ParserX86ATT()
elif isa.lower() == 'aarch64':
return ParserAArch64v81()
else:
raise ValueError("Unknown ISA {!r}.".format(isa))

View File

@@ -2,28 +2,46 @@
import os
import re
from itertools import product
from ruamel import yaml
from osaca import utils
from osaca import utils, __version__
from osaca.parser import ParserX86ATT
class MachineModel(object):
def __init__(self, arch=None, path_to_yaml=None):
if not arch and not path_to_yaml:
raise ValueError('Either arch or path_to_yaml required.')
if arch and path_to_yaml:
raise ValueError('Only one of arch and path_to_yaml is allowed.')
self._path = path_to_yaml
self._arch = arch
if arch:
self._arch = arch.lower()
with open(utils.find_file(self._arch+'.yml'), 'r') as f:
self._data = yaml.load(f, Loader=yaml.Loader)
elif path_to_yaml:
with open(self._path, 'r') as f:
self._data = yaml.load(f, Loader=yaml.Loader)
self._data = {
'osaca_version': str(__version__),
'micro_architecture': None,
'arch_code': None,
'isa': None,
'ROB_size': None,
'retired_uOps_per_cycle': None,
'scheduler_size': None,
'hidden_loads': None,
'load_latency': {},
'load_throughput': [
{'base': b, 'index': i, 'offset': o, 'scale': s, 'port_pressure': []}
for b, i, o, s in product(['gpr'], ['gpr', None], ['imd', None], [1, 8])],
'ports': [],
'port_model_scheme': None,
'instruction_forms': [],
}
else:
if arch and path_to_yaml:
raise ValueError('Only one of arch and path_to_yaml is allowed.')
self._path = path_to_yaml
self._arch = arch
if arch:
self._arch = arch.lower()
with open(utils.find_file(self._arch+'.yml'), 'r') as f:
self._data = yaml.load(f, Loader=yaml.Loader)
elif path_to_yaml:
with open(self._path, 'r') as f:
self._data = yaml.load(f, Loader=yaml.Loader)
def __getitem__(self, key):
"""Return configuration entry."""
@@ -36,20 +54,51 @@ class MachineModel(object):
######################################################
def get_instruction(self, name, operands):
"""Find and return instruction data from name and operands."""
if name is None:
return None
try:
return next(
instruction_form
for instruction_form in self._data['instruction_forms']
if instruction_form['name'].lower() == name.lower()
if instruction_form['name'].upper() == name.upper()
and self._match_operands(instruction_form['operands'], operands)
)
except StopIteration:
return None
except TypeError:
except TypeError as e:
print('\nname: {}\noperands: {}'.format(name, operands))
raise TypeError
raise TypeError from e
def average_port_pressure(self, port_pressure):
"""Construct average port pressure list from instruction data."""
port_list = self._data['ports']
average_pressure = [0.0]*len(port_list)
for cycles, ports in port_pressure:
for p in ports:
average_pressure[port_list.index(p)] += cycles/len(ports)
return average_pressure
def set_instruction(self, name, operands,
latency=None, port_pressure=None, throughput=None, uops=None):
"""Import instruction form information."""
# If it already exists. Overwrite information.
instr_data = self.get_instruction(name, operands)
if instr_data is None:
instr_data = {}
self._data['instruction_forms'].append(instr_data)
instr_data['name'] = name
instr_data['operands'] = operands
instr_data['latency'] = latency
instr_data['port_pressure'] = port_pressure
instr_data['throughput'] = throughput
instr_data['uops'] = uops
def add_port(self, port):
if port not in self._data['ports']:
self._data['ports'].append(port)
def get_ISA(self):
return self._data['isa']
@@ -103,6 +152,9 @@ class MachineModel(object):
return arch_dict[arch].lower()
return None
def dump(self):
return yaml.dump(self._data, Dumper=yaml.Dumper, allow_unicode=True)
######################################################
def _check_for_duplicate(self, name, operands):