-
Notifications
You must be signed in to change notification settings - Fork 44
/
environment_os.py
504 lines (414 loc) · 17.9 KB
/
environment_os.py
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
# coding=utf-8
# Copyright 2021 The Circuit Training Team Authors.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
"""Circuit training Environmnet with gin config."""
import datetime
import math
import os
from typing import Any, Callable, Dict, Text, Tuple, Optional
from absl import logging
from Plc_client import coordinate_descent_placer as cd_placer
from Plc_client import observation_config
from Plc_client import observation_extractor_os as observation_extractor
from Plc_client import placement_util_os as placement_util
from Plc_client import plc_client as plc_client
from Plc_client import plc_client_os as plc_client_os
import gin
import gym
import numpy as np
import tensorflow as tf
from inspect import currentframe, getframeinfo
# from tf_agents.environments import suite_gym
# from tf_agents.environments import wrappers
ObsType = Dict[Text, np.ndarray]
InfoType = Dict[Text, float]
class InfeasibleActionError(ValueError):
"""An infeasible action were passed to the env."""
def __init__(self, action, mask):
"""Initialize an infeasible action error.
Args:
action: Infeasible action that was performed.
mask: The mask associated with the current observation. mask[action] is
`0` for infeasible actions.
"""
ValueError.__init__(self, action, mask)
self.action = action
self.mask = mask
def __str__(self):
return 'Infeasible action (%s) when the mask is (%s)' % (self.action,
self.mask)
@gin.configurable
def cost_info_function(
plc: plc_client.PlacementCost,
done: bool,
wirelength_weight: float = 1.0,
density_weight: float = 0.5,
congestion_weight: float = 0.5) -> Tuple[float, Dict[Text, float]]:
"""Returns the RL cost and info.
Args:
plc: Placement cost object.
done: Set if it is the terminal step.
wirelength_weight: Weight of wirelength in the reward function.
density_weight: Weight of density in the reward function.
congestion_weight: Weight of congestion in the reward function used only for
legalizing the placement in greedy std cell placer.
Returns:
The RL cost.
Raises:
ValueError: When the cost mode is not supported.
Notes: we found the default congestion and density weights more stable.
"""
proxy_cost = 0.0
if not done:
return proxy_cost, {
'wirelength': -1.0,
'congestion': -1.0,
'density': -1.0,
}
wirelength = -1.0
congestion = -1.0
density = -1.0
if wirelength_weight > 0.0:
wirelength = plc.get_cost()
proxy_cost += wirelength_weight * wirelength
if congestion_weight > 0.0:
congestion = plc.get_congestion_cost()
proxy_cost += congestion_weight * congestion
if density_weight > 0.0:
density = plc.get_density_cost()
proxy_cost += density_weight * density
info = {
'wirelength': wirelength,
'congestion': congestion,
'density': density,
}
return proxy_cost, info
@gin.configurable
class CircuitEnv(object):
"""Defines the CircuitEnv class."""
INFEASIBLE_REWARD = -1.0
def __init__(
self,
_plc_client = None,
netlist_file: Text = '',
init_placement: Text = '',
_plc = None,
create_placement_cost_fn: Callable[
..., plc_client.PlacementCost] = placement_util.create_placement_cost,
std_cell_placer_mode: Text = 'fd',
cost_info_fn: Callable[[plc_client.PlacementCost, bool],
Tuple[float, Dict[Text,
float]]] = cost_info_function,
global_seed: int = 0,
is_eval: bool = False,
save_best_cost: bool = False,
output_plc_file: Text = '',
make_soft_macros_square: bool = True,
cd_finetune: bool = False,
cd_plc_file: Text = 'ppo_cd_placement.plc',
train_step: Optional[tf.Variable] = None,
unplace_all_nodes_in_init: bool = True):
"""Creates a CircuitEnv.
Args:
netlist_file: Path to the input netlist file.
init_placement: Path to the input inital placement file, used to read grid
and canas size.
create_placement_cost_fn: A function that given the netlist and initial
placement file create the placement_cost object.
std_cell_placer_mode: Options for fast std cells placement: `fd` (uses the
force-directed algorithm).
cost_info_fn: The cost function that given the plc object returns the RL
cost.
global_seed: Global seed for initializing env features. This seed should
be the same across actors. Not used currently.
is_eval: If set, save the final placement in output_dir.
save_best_cost: Boolean, if set, saves the palcement if its cost is better
than the previously saved palcement.
output_plc_file: The path to save the final placement.
make_soft_macros_square: If True, make the shape of soft macros square
before using analytical std cell placers like FD.
cd_finetune: If True, runs coordinate descent to finetune macro
orientations. Supposed to run in eval only, not training.
cd_plc_file: Name of the CD fine-tuned plc file, the file will be save in
the same dir as output_plc_file
train_step: A tf.Variable indicating the training step, only used for
saving plc files in the evaluation.
unplace_all_nodes_in_init: Unplace all nodes after initialization.
"""
del global_seed
if not netlist_file:
raise ValueError('netlist_file must be provided.')
self.netlist_file = netlist_file
self._std_cell_placer_mode = std_cell_placer_mode
self._cost_info_fn = cost_info_fn
self._is_eval = is_eval
self._save_best_cost = save_best_cost
self._output_plc_file = output_plc_file
self._output_plc_dir = os.path.dirname(output_plc_file)
self._make_soft_macros_square = make_soft_macros_square
self._cd_finetune = cd_finetune
self._cd_plc_file = cd_plc_file
self._train_step = train_step
self._plc = create_placement_cost_fn(plc_client=_plc,
netlist_file=netlist_file, init_placement=init_placement)
# We call ObservationExtractor before unplace_all_nodes, so we use the
# inital placement in the static features (location_x and location_y).
# This results in better placements.
self._observation_config = observation_config.ObservationConfig()
self._observation_extractor = observation_extractor.ObservationExtractor(
plc=self._plc)
# print(getframeinfo(currentframe()).lineno, '\n', np.array(self._plc.get_node_mask(13333)).reshape(35,33))
if self._make_soft_macros_square:
# It is better to make the shape of soft macros square before using
# analytical std cell placers like FD.
self._plc.make_soft_macros_square()
self._grid_cols, self._grid_rows = self._plc.get_grid_num_columns_rows()
self._canvas_width, self._canvas_height = self._plc.get_canvas_width_height(
)
self._hard_macro_indices = [
m for m in self._plc.get_macro_indices()
if not self._plc.is_node_soft_macro(m)
]
self._num_hard_macros = len(self._hard_macro_indices)
# sorted by size
self._sorted_node_indices = placement_util.get_ordered_node_indices(
mode='descending_size_macro_first', plc=self._plc)
# print(len(self._sorted_node_indices))
self._sorted_soft_macros = self._sorted_node_indices[self._num_hard_macros:]
# print(len(self._sorted_soft_macros))
# Generate a map from actual macro_index to its position in
# self.macro_indices. Needed because node adjacency matrix is in the same
# node order of plc.get_macro_indices.
self._macro_index_to_pos = {}
for i, macro_index in enumerate(self._plc.get_macro_indices()):
self._macro_index_to_pos[macro_index] = i
# Padding for mapping the placement canvas on the agent canvas.
rows_pad = self._observation_config.max_grid_size - self._grid_rows
cols_pad = self._observation_config.max_grid_size - self._grid_cols
self._up_pad = rows_pad // 2
self._right_pad = cols_pad // 2
self._low_pad = rows_pad - self._up_pad
self._left_pad = cols_pad - self._right_pad
self._saved_cost = np.inf
self._current_actions = []
self._current_node = 0
self._done = False
self._current_mask = self._get_mask()
# print(getframeinfo(currentframe()).lineno, '\n', np.array(self._plc.get_node_mask(0)).reshape(35,33))
self._infeasible_state = False
if unplace_all_nodes_in_init:
# TODO(b/223026568) Remove unplace_all_nodes from init
self._plc.unplace_all_nodes()
logging.warning('* Unplaced all Nodes in init *')
logging.info('***Num node to place***:%s', self._num_hard_macros)
@property
def observation_space(self) -> gym.spaces.Space:
"""Env Observation space."""
return self._observation_config.observation_space
@property
def action_space(self) -> gym.spaces.Space:
return gym.spaces.Discrete(self._observation_config.max_grid_size**2)
@property
def environment_name(self) -> Text:
return self.netlist_file
def get_static_obs(self):
"""Get the static observation for the environment.
Static observations are invariant across steps on the same netlist, such as
netlist metadata and the adj graphs. This should only be used for
generalized RL.
Returns:
Numpy array representing the observation
"""
return self._observation_extractor.get_static_features()
def get_cost_info(self,
done: bool = False) -> Tuple[float, Dict[Text, float]]:
return self._cost_info_fn(plc=self._plc, done=done) # pytype: disable=wrong-keyword-args # trace-all-classes
def _get_mask(self) -> np.ndarray:
"""Gets the node mask for the current node.
Returns:
List of 0s and 1s indicating if action is feasible or not.
"""
if self._done:
# reset the board
mask = np.zeros(self._observation_config.max_grid_size**2, dtype=np.int32)
else:
node_index = self._sorted_node_indices[self._current_node]
mask = np.asarray(self._plc.get_node_mask(node_index), dtype=np.int32)
# print("current node mask: \n", mask.reshape(35, 33))
mask = np.reshape(mask, [self._grid_rows, self._grid_cols])
pad = ((self._up_pad, self._low_pad), (self._right_pad, self._left_pad))
mask = np.pad(mask, pad, mode='constant', constant_values=0)
return np.reshape(
mask, (self._observation_config.max_grid_size**2,)).astype(np.int32)
def _get_obs(self) -> ObsType:
"""Returns the observation."""
if self._current_node > 0:
previous_node_sorted = self._sorted_node_indices[self._current_node - 1]
previous_node_index = self._macro_index_to_pos[previous_node_sorted]
else:
previous_node_index = -1
if self._current_node < self._num_hard_macros:
current_node_sorted = self._sorted_node_indices[self._current_node]
current_node_index = self._macro_index_to_pos[current_node_sorted]
else:
current_node_index = 0
# YW EDIT: get observation
res = self._observation_extractor.get_all_features(
previous_node_index=previous_node_index,
current_node_index=current_node_index,
mask=self._current_mask)
# self._plc.set_routes_per_micron(44, 44) Does not do anything
# logging.info("####### Observation Extractor #######")
# print(res)
# logging.info("####### End #######")
return res
def _run_cd(self):
"""Runs coordinate descent to finetune the current placement."""
# CD only modifies macro orientation.
# Plc modified by CD will be reset at the end of the episode.
def cost_fn(plc):
return self._cost_info_fn(plc=plc, done=True) # pytype: disable=wrong-keyword-args # trace-all-classes
cd = cd_placer.CoordinateDescentPlacer(
plc=self._plc,
cost_fn=cost_fn,
use_stdcell_placer=True,
optimize_only_orientation=True)
cd.place()
def _save_placement(self, cost: float) -> None:
"""Saves the current placement.
Args:
cost: the current placement cost.
Raises:
IOError: If we cannot write the placement to file.
"""
if not self._save_best_cost or (cost < self._saved_cost and
(math.fabs(cost - self._saved_cost) /
(cost) > 5e-3)):
user_comments = ''
if self._train_step:
user_comments = f'Train step : {self._train_step.numpy()}'
placement_util.save_placement(self._plc, self._output_plc_file,
user_comments)
ts = datetime.datetime.now().strftime('%Y%m%d_%H%M%S')
ppo_snapshot_file = os.path.join(
self._output_plc_dir,
f'snapshot_ppo_opt_placement_timestamp_{ts}_cost_{cost:.4f}.plc')
placement_util.save_placement(self._plc, ppo_snapshot_file, user_comments)
self._saved_cost = cost
# Only runs CD if this is the best RL placement seen so far.
if self._cd_finetune:
self._run_cd()
cost = self._cost_info_fn(plc=self._plc, done=True)[0] # pytype: disable=wrong-keyword-args # trace-all-classes
cd_plc_file = os.path.join(self._output_plc_dir, self._cd_plc_file)
placement_util.save_placement(self._plc, cd_plc_file, user_comments)
cd_snapshot_file = os.path.join(
self._output_plc_dir,
f'snapshot_ppo_cd_placement_timestamp_{ts}_cost_{cost:.4f}.plc')
placement_util.save_placement(self._plc, cd_snapshot_file,
user_comments)
def call_analytical_placer_and_get_cost(self):
"""Calls analytical placer.
Calls analystical placer and evaluates cost when all nodes are placed. Also,
saves the placement file for eval if all the macros are placed by RL.
Returns:
A tuple for placement cost and info.
"""
if self._done:
self.analytical_placer()
# Only evaluates placement cost when all nodes are placed.
# All samples in the episode receive the same reward equal to final cost.
# This is realized by setting intermediate steps cost as zero, and
# propagate the final cost with discount factor set to 1 in replay buffer.
cost, info = self._cost_info_fn(self._plc, self._done)
# We only save placement if all nodes by placed RL, because the dreamplace
# mix-sized placement may not be legal.
if self._current_node == self._num_hard_macros and self._is_eval:
self._save_placement(cost)
return -cost, info
def reset(self) -> ObsType:
"""Restes the environment.
Returns:
An initial observation.
"""
self._plc.unplace_all_nodes()
self._current_actions = []
self._current_node = 0
self._done = False
self._current_mask = self._get_mask()
return self._get_obs()
def translate_to_original_canvas(self, action: int) -> int:
"""Translates a raw location to real one in the original canvas."""
up_pad = (self._observation_config.max_grid_size - self._grid_rows) // 2
right_pad = (self._observation_config.max_grid_size - self._grid_cols) // 2
a_i = action // self._observation_config.max_grid_size - up_pad
a_j = action % self._observation_config.max_grid_size - right_pad
if 0 <= a_i < self._grid_rows or 0 <= a_j < self._grid_cols:
action = a_i * self._grid_cols + a_j
else:
raise InfeasibleActionError(action, self._current_mask)
return action
def place_node(self, node_index: int, action: int) -> None:
self._plc.place_node(node_index, self.translate_to_original_canvas(action))
def analytical_placer(self) -> None:
if self._std_cell_placer_mode == 'fd':
placement_util.fd_placement_schedule(self._plc)
else:
raise ValueError('%s is not a supported std_cell_placer_mode.' %
(self._std_cell_placer_mode))
def step(self, action: int) -> Tuple[ObsType, float, bool, Any]:
"""Steps the environment.
Args:
action: The action to take (should be a list of size 1).
Returns:
observation, reward, done, and info.
Raises:
RuntimeError: action taken after episode was done
InfeasibleActionError: bad action taken (action is not in feasible
actions)
"""
if self._done:
raise RuntimeError('Action taken after episode is done.')
action = int(action)
self._current_actions.append(action)
if self._current_mask[action] == 0:
raise InfeasibleActionError(action, self._current_mask)
node_index = self._sorted_node_indices[self._current_node]
self.place_node(node_index, action)
self._current_node += 1
self._done = (self._current_node == self._num_hard_macros)
self._current_mask = self._get_mask()
if not self._done and not np.any(self._current_mask):
# Please note that _infeasible_state is not reset in reset function so,
# the caller of step() is responsible for resetting it.
self._infeasible_state = True
logging.info('Actions took before becoming infeasible: %s',
self._current_actions)
info = {
'wirelength': -1.0,
'congestion': -1.0,
'density': -1.0,
}
return self.reset(), self.INFEASIBLE_REWARD, True, info
cost, info = self.call_analytical_placer_and_get_cost()
return self._get_obs(), cost, self._done, info
# def create_circuit_environment(*args, **kwarg) -> wrappers.ActionClipWrapper:
# """Create an `CircuitEnv` wrapped as a Gym environment.
# Args:
# *args: Arguments.
# **kwarg: keyworded Arguments.
# Returns:
# PyEnvironment used for training.
# """
# env = CircuitEnv(*args, **kwarg)
# return wrappers.ActionClipWrapper(suite_gym.wrap_env(env))