"""
Contains functions to estimate aspects of blocks (like area and delay)
by either using internal models or by making calls out to external tool chains.
"""
import re
import os
import math
import tempfile
import subprocess
import sys
import collections
from .core import working_block
from .wire import Input, Output, Const, Register, WireVector
from .pyrtlexceptions import PyrtlError, PyrtlInternalError
from .importexport import output_to_verilog
from .memory import RomBlock
from .helperfuncs import _currently_in_jupyter_notebook, _print_netlist_latex
# --------------------------------------------------------------------
# __ ___ ___ __ ___ ___ __
# /\ |__) |__ /\ |__ /__` | | |\/| /\ | | / \ |\ |
# /~~\ | \ |___ /~~\ |___ .__/ | | | | /~~\ | | \__/ | \|
#
[docs]
def area_estimation(tech_in_nm=130, block=None):
"""Estimates the total area of the block.
:param float tech_in_nm: the size of the circuit technology to be estimated
(for example, 65 is 65nm and 250 is 0.25um)
:return: tuple of estimated areas (logic, mem) in terms of mm^2
The estimations are based off of 130nm standard cell designs for the logic,
and custom memory blocks from the literature. The results are not fully
validated and we do not recommend that this function be used in carrying
out science for publication.
"""
def mem_area_estimate(tech_in_nm, bits, ports, is_rom):
# http://www.cs.ucsb.edu/~sherwood/pubs/ICCD-srammodel.pdf
# ROM is assumed to be 1/10th of area of SRAM
tech_in_um = tech_in_nm / 1000.0
area_estimate = 0.001 * tech_in_um**2.07 * bits**0.9 * ports**0.7 + 0.0048
return area_estimate if not is_rom else area_estimate / 10.0
# Subset of the raw data gathered from yosys, mapping to vsclib 130nm library
# Width Adder_Area Mult_Area (area in "tracks" as discussed below)
# 8 211 2684
# 16 495 12742
# 32 1110 49319
# 64 2397 199175
# 128 4966 749828
def adder_stdcell_estimate(width):
return width * 34.4 - 25.8
def multiplier_stdcell_estimate(width):
if width == 1:
return 5
elif width == 2:
return 39
elif width == 3:
return 219
else:
return -958 + (150 * width) + (45 * width**2)
def stdcell_estimate(net):
if net.op in 'w~sc':
return 0
elif net.op in '&|n':
return 40 / 8.0 * len(net.args[0]) # 40 lambda
elif net.op in '^=<>x':
return 80 / 8.0 * len(net.args[0]) # 80 lambda
elif net.op == 'r':
return 144 / 8.0 * len(net.args[0]) # 144 lambda
elif net.op in '+-':
return adder_stdcell_estimate(len(net.args[0]))
elif net.op == '*':
return multiplier_stdcell_estimate(len(net.args[0]))
elif net.op in 'm@':
return 0 # memories handled elsewhere
else:
raise PyrtlInternalError('Unable to estimate the following net '
'due to unimplemented op :\n%s' % str(net))
block = working_block(block)
# The functions above were gathered and calibrated by mapping
# reference designs to an openly available 130nm stdcell library.
# http://www.vlsitechnology.org/html/vsc_description.html
# http://www.vlsitechnology.org/html/cells/vsclib013/lib_gif_index.html
# In a standard cell design, each gate takes up a length of standard "track"
# in the chip. The functions above return that length for each of the different
# types of functions in the units of "tracks". In the 130nm process used,
# 1 lambda is 55nm, and 1 track is 8 lambda.
# first, sum up the area of all of the logic elements (including registers)
total_tracks = sum(stdcell_estimate(a_net) for a_net in block.logic)
total_length_in_nm = total_tracks * 8 * 55
# each track is then 72 lambda tall, and converted from nm2 to mm2
area_in_mm2_for_130nm = (total_length_in_nm * (72 * 55)) / 1e12
# scaling from 130nm to the target tech
logic_area = area_in_mm2_for_130nm / (130.0 / tech_in_nm) ** 2
# now sum up the area of the memories
mem_area = 0
for mem in set(net.op_param[1] for net in block.logic_subset('@m')):
bits, ports, is_rom = _bits_ports_and_isrom_from_memory(mem)
mem_area += mem_area_estimate(tech_in_nm, bits, ports, is_rom)
return logic_area, mem_area
def _bits_ports_and_isrom_from_memory(mem):
""" Helper to extract mem bits and ports for estimation. """
is_rom = False
bits = 2**mem.addrwidth * mem.bitwidth
read_ports = len(mem.readport_nets)
write_ports = len(mem.writeport_nets)
if isinstance(mem, RomBlock):
is_rom = True
ports = max(read_ports, write_ports)
return bits, ports, is_rom
# --------------------------------------------------------------------
# ___ __ /\ __ __
# | | |\/| | |\ | / ` /~~\ |\ | /\ | \_/ /__` | /__`
# | | | | | | \| \__> / \| \| /~~\ |_ | .__/ | .__/
#
[docs]
class TimingAnalysis(object):
"""Timing analysis estimates the timing delays in the block
TimingAnalysis has an :py:attr:`~.TimingAnalysis.timing_map` object that
maps wires to the 'time' after a clock edge at which the signal in the wire
settles
"""
[docs]
def __init__(self, block=None, gate_delay_funcs=None):
""" Calculates timing delays in the block.
:param Block block: PyRTL block to analyze
:param gate_delay_funcs: a map with keys corresponding to the gate op and
a function returning the delay as the value.
It takes the gate as an argument.
If the delay is negative (-1), the gate will be treated as the end
of the block
Calculates the timing analysis while allowing for
different timing delays of different gates of each type.
Supports all valid presynthesis blocks.
Currently doesn't support memory post synthesis.
"""
self.block = working_block(block)
self.timing_map = None
self.block.sanity_check()
self._generate_timing_map(gate_delay_funcs)
def _generate_timing_map(self, gate_delay_funcs):
# The functions above were gathered and calibrated by mapping
# reference designs to an openly available 130nm stdcell library.
# Note that this is will compute the critical logic delay, but does
# not include setup/hold time.
if gate_delay_funcs is None:
gate_delay_funcs = {
'~': lambda width: 48.5,
'&': lambda width: 98.5,
'|': lambda width: 105.3,
'^': lambda width: 135.07,
'n': lambda width: 66.0,
'w': lambda width: 0,
'+': self._logconst_func(184.0, 18.9),
'-': self._logconst_func(184.0, 18.9),
'*': self._multiplier_stdcell_estimate,
'<': self._logconst_func(101.9, 105.4),
'>': self._logconst_func(101.9, 105.4),
'=': self._logconst_func(60.1, 147),
'x': lambda width: 138.0,
'c': lambda width: 0,
's': lambda width: 0,
'r': lambda width: -1,
'm': self._memory_read_estimate,
'@': lambda width: -1,
}
cleared = self.block.wirevector_subset((Input, Const, Register))
self.timing_map = {wirevector: 0 for wirevector in cleared}
for _gate in self.block: # ordered iteration
if _gate.op == 'm':
gate_delay = gate_delay_funcs['m'](_gate.op_param[1]) # reads require a memid
else:
gate_delay = gate_delay_funcs[_gate.op](len(_gate.args[0]))
if gate_delay < 0:
continue
time = max(self.timing_map[a_wire] for a_wire in _gate.args) + gate_delay
for dest_wire in _gate.dests:
self.timing_map[dest_wire] = time
@staticmethod
def _logconst_func(a, b):
return lambda x: a * math.log(float(x), 2) + b
@staticmethod
def _multiplier_stdcell_estimate(width):
if width == 1:
return 98.57
elif width == 2:
return 200.17
else:
return 549.1 * math.log(width, 2) - 391.7
@staticmethod
def _memory_read_estimate(mem):
# http://www.cs.ucsb.edu/~sherwood/pubs/ICCD-srammodel.pdf
# ROM is assumed to be same delay as SRAM (perhaps optimistic?)
bits, ports, is_rom = _bits_ports_and_isrom_from_memory(mem)
tech_in_um = 0.130
return 270 * tech_in_um**1.38 * bits**0.25 * ports**1.30 + 1.05
[docs]
def max_freq(self, tech_in_nm=130, ffoverhead=None):
"""Estimates the max frequency of a block in MHz.
:param float tech_in_nm: the size of the circuit technology to be estimated
(for example, 65 is 65nm and 250 is 0.25um)
:param float ffoverhead: setup and ff propagation delay in picoseconds
:return: a number representing an estimate of the max frequency in Mhz
All params are optional and have reasonable default values. Estimation
is based on Dennard Scaling assumption and does not include wiring
effect -- as a result the estimates may be optimistic (especially below
65nm).
"""
cp_length = self.max_length()
scale_factor = 130.0 / tech_in_nm
if ffoverhead is None:
clock_period_in_ps = scale_factor * (cp_length + 189 + 194)
else:
clock_period_in_ps = (scale_factor * cp_length) + ffoverhead
return 1e6 * 1.0 / clock_period_in_ps
[docs]
def max_length(self):
"""Returns the max timing delay of the circuit in ps.
The result assumes that the circuit is implemented in a 130nm process,
and that there is no setup or hold time associated with the circuit.
The resulting value is in picoseconds. If an proper estimation of
timing is required it is recommended to us
:py:meth:`.TimingAnalysis.max_freq` to determine the clock period as it
more accurately considers scaling and setup/hold.
"""
return max(self.timing_map.values())
[docs]
def print_max_length(self):
"""Prints the max timing delay of the circuit """
print("The total block timing delay is ", self.max_length())
class _TooManyCPsError(Exception):
pass
[docs]
def critical_path(self, print_cp=True, cp_limit=100):
""" Takes a timing map and returns the critical paths of the system.
:param bool print_cp: Whether to print the critical path to the terminal
after calculation
:return: a list containing tuples with the 'first' wire as the
first value and the critical paths (which themselves are lists
of nets) as the second
"""
critical_paths = [] # storage of all completed critical paths
wire_src_map, dst_map = self.block.net_connections()
def critical_path_pass(old_critical_path, first_wire):
if isinstance(first_wire, (Input, Const, Register)):
critical_paths.append((first_wire, old_critical_path))
return
if len(critical_paths) >= cp_limit:
raise self._TooManyCPsError()
source = wire_src_map[first_wire]
critical_path = [source]
critical_path.extend(old_critical_path)
arg_max_time = max(self.timing_map[arg_wire] for arg_wire in source.args)
for arg_wire in source.args:
# if the time for both items are the max, both will be on a critical path
if self.timing_map[arg_wire] == arg_max_time:
critical_path_pass(critical_path, arg_wire)
max_time = self.max_length()
try:
for wire_pair in self.timing_map.items():
if wire_pair[1] == max_time:
critical_path_pass([], wire_pair[0])
except self._TooManyCPsError:
print("Critical path count limit reached")
if print_cp:
self.print_critical_paths(critical_paths)
return critical_paths
[docs]
@staticmethod
def print_critical_paths(critical_paths):
""" Prints the results of the critical path length analysis.
Done by default by the :py:meth:`.TimingAnalysis.critical_path` function.
"""
line_indent = " " * 2
# print the critical path
for cp_with_num in enumerate(critical_paths):
print("Critical path", cp_with_num[0], ":")
print(line_indent, "The first wire is:", cp_with_num[1][0])
if _currently_in_jupyter_notebook():
_print_netlist_latex(list(net for net in cp_with_num[1][1]))
else:
for net in cp_with_num[1][1]:
print(line_indent, (net))
print()
# --------------------------------------------------------------------
# __ __ __
# \ / / \ /__` \ / /__`
# | \__/ .__/ | .__/
#
[docs]
def yosys_area_delay(library, abc_cmd=None, leave_in_dir=None, block=None):
"""Synthesize with `Yosys <https://yosyshq.net/yosys/>`_ and return
estimate of area and delay.
:param library: stdcell library file to target in liberty format
:param abc_cmd: string of commands for :program:`yosys` to pass to
:program:`abc` for synthesis
:param dir: the directory where temporary files should be left
:param block: PyRTL block to analyze
:return: a tuple of numbers: area, delay
If `dir` is specified, that directory will be used to create any temporary
files, and the resulting files will be left behind there (which can be
useful for manual exploration or debugging)
The area and delay are returned in units as defined by the stdcell
library. In the standard vsc 130nm library, the area is in a number of
"tracks", each of which is about 1.74 square um (see area estimation
for more details) and the delay is in ps.
http://www.vlsitechnology.org/html/vsc_description.html
May raise `PyrtlError` if :program:`yosys` is not configured correctly, and
`PyrtlInternalError` if the call to :program:`yosys` was not successful
"""
if abc_cmd is None:
abc_cmd = 'strash;scorr;ifraig;retime;dch,-f;map;print_stats;'
else:
# first, replace whitespace with commas as per yosys requirements
re.sub(r"\s+", ',', abc_cmd)
# then append with "print_stats" to generate the area and delay info
abc_cmd = '%s;print_stats;' % abc_cmd
def extract_area_delay_from_yosys_output(yosys_output):
report_lines = [line
for line in yosys_output.decode().split('\n')
if 'ABC: netlist' in line]
area = re.match(r'.*area\s*=\s*([0-9\.]*)', report_lines[0]).group(1)
delay = re.match(r'.*delay\s*=\s*([0-9\.]*)', report_lines[0]).group(1)
return float(area), float(delay)
yosys_arg_template = """-p
read_verilog %s;
synth -top toplevel;
dfflibmap -liberty %s;
abc -liberty %s -script +%s
""".replace('\n', ' ')
temp_d, temp_path = tempfile.mkstemp(prefix='pyrtl_verilog', suffix='.v',
dir=leave_in_dir, text=True)
try:
# write the verilog to a temp
yosys_arg = yosys_arg_template % (temp_path, library, library, abc_cmd)
with open(temp_path, 'w') as f:
print('// generated via pyrtl yosys_area_delay', file=f)
print('// yosys %s' % yosys_arg, file=f)
output_to_verilog(f, block=block)
os.close(temp_d)
# call yosys on the temp, and grab the output
yosys_output = subprocess.check_output(['yosys', yosys_arg])
area, delay = extract_area_delay_from_yosys_output(yosys_output)
except (subprocess.CalledProcessError, ValueError) as e:
print('Error with call to yosys...', file=sys.stderr)
print('---------------------------------------------', file=sys.stderr)
print(str(e.output).replace('\\n', '\n'), file=sys.stderr)
print('---------------------------------------------', file=sys.stderr)
raise PyrtlError('Yosys callfailed')
except OSError as e:
print('Error with call to yosys...', file=sys.stderr)
raise PyrtlError('Call to yosys failed (not installed or on path?)')
finally:
if leave_in_dir is None:
os.remove(temp_path)
return area, delay
[docs]
class PathsResult(dict):
[docs]
def print(self, file=sys.stdout):
""" Pretty print the result of calling :py:func:`.paths`
:param f: the open file to print to (defaults to stdout)
:return: None
"""
# All this work, to make sure it's determinstic
def path_sort_key(path):
dst_names = [net.dests[0].name if net.dests else '' for net in path]
return (len(path), dst_names)
for start in sorted(self.keys(), key=lambda w: w.name):
print("From %s" % start.name, file=file)
for end in sorted(self[start].keys(), key=lambda w: w.name):
print(" To %s" % end.name, file=file)
paths = self[start][end]
if len(paths) > 0:
for i, paths in enumerate(sorted(paths, key=path_sort_key)):
print(" Path %d" % i, file=file)
for path in paths:
print(" %s" % str(path), file=file)
else:
print(" (No paths)", file=file)
[docs]
def paths(src=None, dst=None, dst_nets=None, block=None):
"""Get the list of all paths from `src` to `dst`.
:param Union[WireVector, Iterable[WireVector]] src: source wire(s) from which to
trace your paths; if None, will get paths from all Inputs
:param Union[WireVector, Iterable[WireVector]] dst: destination wire(s) to which to
trace your paths; if None, will get paths to all Outputs
:param dict[WireVector, LogicNet] dst_nets: map from wire to set of nets where the
wire is an argument; will compute it internally if not given via a
call to pyrtl.net_connections()
:param Block block: block to use (defaults to working block)
:return: a map of the form `{src_wire: {dst_wire: [path]}}` for each `src_wire` in `src`
(or all inputs if `src` is None), `dst_wire` in `dst` (or all outputs if `dst` is None),
where `path` is a list of nets. This map is also an instance of :py:class:`.PathsResult`,
so you can call :py:meth:`.PathsResult.print` on it to pretty print it.
You can provide `dst_nets` (the result of calling :py:func:`.net_connections`, if you plan
on calling this function repeatedly on a block that hasn't changed, to speed things up.
This function can accept one or more `src` wires, and one or more `dst` wires,
such that it returns a map that can be accessed like so::
paths[src][dst] = [<path>, <path>, ...]
where `path` is a list of nets. Thus there can be multiple paths from a given `src` wire
to a given `dst` wire.
If `src` and `dst` are both single wires, you still need to access the
result via `paths[src][dst]`.
This also finds and returns the loop paths in the case of registers or memories that feed into
themselves, i.e. `paths[src][src]` is not necessarily empty.
It does not distinguish between loops that include synchronous vs asynchronous memories.
"""
block = working_block(block)
if dst_nets is None:
# Note: if you set `include_virtual_nodes=True`, Output wires will actually
# be present as the destination "net" of Output wires in the dst_nets map.
# That would overly complicate this algorithm: we will assume all values()
# in the dst_nets map are logic nets only. We set this to False for explicitness...
_, dst_nets = block.net_connections(include_virtual_nodes=False)
else:
# ... or make sure it's not present otherwise.
for output in block.wirevector_subset(cls=Output):
dst_nets.pop(output, None)
if src is None:
src = block.wirevector_subset(cls=Input)
elif isinstance(src, WireVector):
src = {src}
else:
src = set(src)
if dst is None:
dst = block.wirevector_subset(cls=Output)
elif isinstance(dst, WireVector):
dst = {dst}
else:
dst = set(dst)
def paths_src_dst(src, dst):
paths = []
# Use DFS to get the paths [each a list of nets] from src wire to dst wire
def dfs(w, curr_path):
if w is dst:
# Found valid path
paths.append(curr_path)
for dst_net in dst_nets.get(w, []):
# Avoid loops and the mem net (has no output wire)
if dst_net not in curr_path:
if dst_net.op == '@': # dests will be the read ports
for read_net in dst_net.op_param[1].readport_nets:
dfs(read_net.dests[0], curr_path + [dst_net, read_net])
else:
dfs(dst_net.dests[0], curr_path + [dst_net])
dfs(src, [])
return paths
all_paths = collections.defaultdict(dict)
for src_wire in src:
for dst_wire in dst:
paths = paths_src_dst(src_wire, dst_wire)
# Remove empty paths...
paths = list(filter(lambda x: len(x) > 0, paths))
# ...and those that are supersets of others (resulting from an inner loop).
if src_wire is not dst_wire:
paths = sorted(paths, key=lambda p: len(p), reverse=True)
keep = []
for i in range(len(paths)):
# Check if there is a path in paths[i+1:] that is the suffix
# of paths[i] (paths[i] is at least as large as each path in
# paths[i+1:]). If so, paths[i] contains a loop since both start
# at src_wire, so don't keep it.
if not any(paths[i][-len(p):] == p for p in paths[i + 1:]):
keep.append(paths[i])
paths = keep
all_paths[src_wire][dst_wire] = paths
return PathsResult(all_paths)
[docs]
def distance(src, dst, f, block=None):
""" Calculate the 'distance' along each path from `src` to `dst` according to `f`
:param WireVector src: wire to start from
:param WireVector dst: wire to end on
:param Callable[[LogicNet], int] f: function from a net to number,
representing the 'value' of a net that you want to sum
across all nets in the path
:param Block block: block to use (defaults to working block)
:return: a map from each path (a tuple) to its calculated distance
This calls the given function `f` on each net in a path, summing the result.
"""
ps = paths(src, dst, block=block)
ps = ps[src][dst]
# Turning path into tuple so it can be the key
m = {tuple(path): sum(map(f, path)) for path in ps}
return m
[docs]
def fanout(w):
""" Get the number of places a wire is used as an argument.
:param WireVector w: WireVector to check fanout for
:return: integer fanout count
"""
_, dst_nets = w._block.net_connections()
if w not in dst_nets:
return 0
all_args = [arg for net in dst_nets[w] for arg in net.args]
return len(list(filter(lambda arg: arg is w, all_args)))