Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 1 | # Copyright (C) 2020 Arm Limited or its affiliates. All rights reserved. |
| 2 | # |
| 3 | # SPDX-License-Identifier: Apache-2.0 |
| 4 | # |
| 5 | # Licensed under the Apache License, Version 2.0 (the License); you may |
| 6 | # not use this file except in compliance with the License. |
| 7 | # You may obtain a copy of the License at |
| 8 | # |
| 9 | # www.apache.org/licenses/LICENSE-2.0 |
| 10 | # |
| 11 | # Unless required by applicable law or agreed to in writing, software |
| 12 | # distributed under the License is distributed on an AS IS BASIS, WITHOUT |
| 13 | # WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. |
| 14 | # See the License for the specific language governing permissions and |
| 15 | # limitations under the License. |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 16 | # Description: |
| 17 | # The scheduler costs various strategies for scheduling the network in order to select the block configuration. |
Diego Russo | ea6111a | 2020-04-14 18:41:58 +0100 | [diff] [blame] | 18 | import copy |
Diego Russo | e8a1045 | 2020-04-21 17:39:10 +0100 | [diff] [blame] | 19 | import enum |
| 20 | from functools import lru_cache |
Diego Russo | ea6111a | 2020-04-14 18:41:58 +0100 | [diff] [blame] | 21 | |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 22 | import numpy as np |
Diego Russo | ea6111a | 2020-04-14 18:41:58 +0100 | [diff] [blame] | 23 | |
| 24 | from . import live_range |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 25 | from . import npu_performance |
| 26 | from . import stats_writer |
Fredrik Svedberg | 880e735 | 2020-08-25 11:31:47 +0200 | [diff] [blame] | 27 | from .data_type import DataType |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 28 | from .high_level_command_stream_generator import calc_allowed_ofm_ifm_overlap_for_pass_list |
Diego Russo | e8a1045 | 2020-04-21 17:39:10 +0100 | [diff] [blame] | 29 | from .nn_graph import CascadedPass |
| 30 | from .nn_graph import PassPlacement |
| 31 | from .nn_graph import SchedulerRewrite |
| 32 | from .nn_graph import SchedulingStrategy |
| 33 | from .npu_performance import make_bandwidth_array |
| 34 | from .npu_performance import make_cycles_array |
| 35 | from .npu_performance import make_macs_array |
| 36 | from .npu_performance import make_metrics_arrays |
| 37 | from .npu_performance import PassCycles |
Jacob Bohlin | 1a66697 | 2020-09-11 10:04:15 +0200 | [diff] [blame] | 38 | from .numeric_util import full_shape |
Diego Russo | e8a1045 | 2020-04-21 17:39:10 +0100 | [diff] [blame] | 39 | from .operation import NpuBlockType |
Louis Verhaard | aee5d75 | 2020-09-30 09:01:52 +0200 | [diff] [blame^] | 40 | from .operation import Op |
Diego Russo | e8a1045 | 2020-04-21 17:39:10 +0100 | [diff] [blame] | 41 | from .shared_buffer_allocation import find_block_configs_suitable_for_pass_and_shared_buffer |
| 42 | from .shared_buffer_allocation import shared_buffer_allocation_for_pass_and_block_config |
| 43 | from .tensor import MemArea |
Patrik Gustavsson | eca2e95 | 2020-05-27 09:15:11 +0200 | [diff] [blame] | 44 | from .tensor import MemType |
Diego Russo | e8a1045 | 2020-04-21 17:39:10 +0100 | [diff] [blame] | 45 | from .tensor import TensorFormat |
| 46 | from .tensor import TensorPurpose |
| 47 | from .tensor import TensorSubPurpose |
Jacob Bohlin | 1a66697 | 2020-09-11 10:04:15 +0200 | [diff] [blame] | 48 | |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 49 | |
| 50 | class ParetoMetric(enum.Enum): |
| 51 | BwCycMem = 1 |
| 52 | BwCycMemBlkH = 2 |
| 53 | |
| 54 | def __str__(self): |
| 55 | return self.name |
| 56 | |
| 57 | |
| 58 | class SchedulerOptions: |
| 59 | def __init__( |
| 60 | self, |
| 61 | use_cascading=True, |
| 62 | use_ifm_ofm_overlap=True, |
| 63 | verbose_schedule=False, |
| 64 | verbose_pareto_frontier_schedules=False, |
| 65 | use_ifm_streaming=True, |
| 66 | pareto_metric=ParetoMetric.BwCycMem, |
Charles Xu | 7b8823f | 2020-05-29 13:53:10 +0200 | [diff] [blame] | 67 | use_nhcwb16_between_cascaded_passes=True, |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 68 | ): |
| 69 | self.use_cascading = use_cascading |
| 70 | self.use_ifm_ofm_overlap = use_ifm_ofm_overlap |
| 71 | self.verbose_schedule = verbose_schedule |
| 72 | self.verbose_pareto_frontier_schedules = verbose_pareto_frontier_schedules |
| 73 | self.use_ifm_streaming = use_ifm_streaming |
| 74 | self.pareto_metric = pareto_metric |
Charles Xu | 7b8823f | 2020-05-29 13:53:10 +0200 | [diff] [blame] | 75 | self.use_nhcwb16_between_cascaded_passes = use_nhcwb16_between_cascaded_passes |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 76 | |
| 77 | def __str__(self): |
| 78 | return type(self).__name__ + ": " + str(self.__dict__) |
| 79 | |
| 80 | __repr__ = __str__ |
| 81 | |
| 82 | |
| 83 | class Strategy: |
| 84 | __slots__ = "strat", "param", "passes", "block_configs", "rewrite_list", "bws", "macs", "cycles", "sram_used" |
| 85 | |
| 86 | def __init__(self, strat, param, passes, block_configs, rewrite_list, bws, macs, cycles, sram_used): |
| 87 | self.strat = strat |
| 88 | self.param = param |
| 89 | self.passes = passes |
| 90 | self.block_configs = block_configs |
| 91 | self.rewrite_list = ( |
| 92 | rewrite_list # list of (SchedulerRewrite, Tensor, new sub purpose, purpose param a, purpose param b, pass) |
| 93 | ) |
| 94 | self.bws = bws |
| 95 | self.macs = macs |
| 96 | self.cycles = cycles |
| 97 | self.sram_used = sram_used |
| 98 | |
| 99 | def __eq__(self, other): |
| 100 | if self.strat != other.strat: |
| 101 | return False |
| 102 | if self.param != other.param: |
| 103 | return False |
| 104 | if self.block_configs != other.block_configs: |
| 105 | return False |
| 106 | if self.passes != other.passes: |
| 107 | return False |
| 108 | if (self.bws != other.bws).any(): |
| 109 | return False |
| 110 | if (self.macs != other.macs).any(): |
| 111 | return False |
| 112 | if (self.cycles != other.cycles).any(): |
| 113 | return False |
| 114 | if self.sram_used != other.sram_used: |
| 115 | return False |
| 116 | return True |
| 117 | |
| 118 | def empty(self): |
| 119 | return not self.passes |
| 120 | |
| 121 | def key(self): |
| 122 | return self.passes[-1] |
| 123 | |
| 124 | def clone(self): |
| 125 | return Strategy( |
| 126 | self.strat, |
| 127 | self.param, |
| 128 | self.passes, |
| 129 | self.block_configs, |
| 130 | self.rewrite_list, |
| 131 | self.bws, |
| 132 | self.macs, |
| 133 | self.cycles, |
| 134 | self.sram_used, |
| 135 | ) |
| 136 | |
| 137 | def __str__(self): |
| 138 | return "<scheduler.Strategy: %s %s %s %s %s %s %s>" % ( |
| 139 | self.strat, |
| 140 | self.passes, |
| 141 | self.rewrite_list, |
| 142 | self.bws, |
| 143 | self.macs, |
| 144 | self.cycles, |
| 145 | self.sram_used, |
| 146 | ) |
| 147 | |
| 148 | __repr__ = __str__ |
| 149 | |
| 150 | |
| 151 | class StrategySet: |
| 152 | __slots__ = "strats", "bws", "macs", "cycles", "max_sram_used", "total_sram_used" |
| 153 | |
| 154 | def __init__(self, strats=None): |
| 155 | if strats is None: |
| 156 | strats = dict() |
| 157 | self.strats = strats # final pass in packed pass -> Strategy |
| 158 | self.bws, self.macs, self.cycles = make_metrics_arrays() |
| 159 | self.max_sram_used = 0 |
| 160 | self.total_sram_used = 0 |
| 161 | |
| 162 | def update_statistics(self): |
| 163 | self.bws = make_bandwidth_array() |
| 164 | self.max_sram_used = 0 |
| 165 | for ps, strat in self.strats.items(): |
| 166 | self.bws += strat.bws |
| 167 | self.macs += strat.macs |
| 168 | self.cycles += strat.cycles |
| 169 | self.max_sram_used = max(self.max_sram_used, strat.sram_used) |
| 170 | self.total_sram_used += strat.sram_used |
| 171 | |
| 172 | def clone_add_strategy(self, new_strat): |
| 173 | key = new_strat.key() |
| 174 | if key in self.strats: |
| 175 | assert new_strat == self.strats[key] |
| 176 | return self |
| 177 | else: |
| 178 | new_strats = dict(self.strats) |
| 179 | new_strats[key] = new_strat |
| 180 | new_set = StrategySet(new_strats) |
| 181 | new_set.bws = self.bws + new_strat.bws |
| 182 | new_set.macs = self.macs + new_strat.macs |
| 183 | new_set.cycles = self.cycles + new_strat.cycles |
| 184 | new_set.max_sram_used = max(self.max_sram_used, new_strat.sram_used) |
| 185 | new_set.total_sram_used = self.total_sram_used + new_strat.sram_used |
| 186 | return new_set |
| 187 | |
| 188 | def __eq__(self, other): |
| 189 | if (self.bws != other.bws).any(): |
| 190 | return False |
| 191 | if (self.macs != other.macs).any(): |
| 192 | return False |
| 193 | if (self.cycles != other.cycles).any(): |
| 194 | return False |
| 195 | if self.max_sram_used != other.max_sram_used: |
| 196 | return False |
| 197 | if self.total_sram_used != other.total_sram_used: |
| 198 | return False |
| 199 | if self.strats != other.strats: |
| 200 | return False |
| 201 | return True |
| 202 | |
| 203 | def __str__(self): |
| 204 | return "<scheduler.StrategySet: max_sram_used=%s passes_covered=%s>" % ( |
| 205 | self.max_sram_used, |
| 206 | list(ps.name for ps in self.strats), |
| 207 | ) |
| 208 | |
| 209 | __repr__ = __str__ |
| 210 | |
| 211 | |
| 212 | empty_strategy = Strategy( |
| 213 | SchedulingStrategy.Unknown, None, [], [], [], make_bandwidth_array(), make_macs_array(), make_cycles_array(), 0 |
| 214 | ) |
| 215 | INFINITY = 1e30 |
| 216 | |
| 217 | ABORT_SEARCH = [] |
| 218 | |
| 219 | |
| 220 | def flatten_list_of_lists(lstlst): |
| 221 | lst = [] |
| 222 | for v in lstlst: |
| 223 | lst.extend(v) |
| 224 | return lst |
| 225 | |
| 226 | |
| 227 | class DynamicProgrammingScheduler: |
| 228 | def __init__(self, nng, sg, arch, sram_limit, options: SchedulerOptions): |
| 229 | self.nng = nng |
| 230 | self.sg = sg |
| 231 | self.arch = arch |
| 232 | self.sram_limit = sram_limit |
| 233 | self.options = copy.copy(options) |
| 234 | self.use_cascading = options.use_cascading |
| 235 | |
| 236 | if self.arch.feature_map_storage_mem_area != MemArea.Sram: |
| 237 | self.use_ifm_ofm_overlap = False # force off IFM/OFM overlap if IFMs and OFMs are not in the SRAM |
Patrik Gustavsson | 3ab9452 | 2020-06-29 17:36:55 +0200 | [diff] [blame] | 238 | else: |
| 239 | self.use_ifm_ofm_overlap = options.use_ifm_ofm_overlap |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 240 | |
| 241 | self.verbose_schedule = options.verbose_schedule |
| 242 | self.verbose_pareto_frontier_schedules = options.verbose_pareto_frontier_schedules |
| 243 | self.mem_area = MemArea.Sram |
| 244 | |
| 245 | self.bandwidth_weights = arch.bandwidth_weights |
| 246 | self.cycles_weight = arch.cycles_weight |
| 247 | self.max_sram_used_weight = arch.max_sram_used_weight |
| 248 | |
| 249 | self.n_combinations_searched = 0 |
| 250 | |
| 251 | self.feature_maps_not_in_fast_storage = ( |
| 252 | arch.tensor_storage_mem_area[TensorPurpose.FeatureMap] != arch.fast_storage_mem_area |
| 253 | ) |
| 254 | |
| 255 | self.pareto_max_candidates = 16 |
| 256 | |
| 257 | self.ifm_stream_npu_blocks = set( |
Louis Verhaard | aee5d75 | 2020-09-30 09:01:52 +0200 | [diff] [blame^] | 258 | (NpuBlockType.ConvolutionMxN, NpuBlockType.ConvolutionDepthWise, NpuBlockType.Pooling,) |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 259 | ) |
| 260 | |
| 261 | num_pareto_metrics = 4 |
| 262 | view_values = ",".join(["d"] * num_pareto_metrics) |
| 263 | order_values = ["f%d" % (idx,) for idx in range(num_pareto_metrics)] |
| 264 | |
| 265 | def pareto_metric(self, candidate): |
| 266 | strat, strat_set = candidate |
| 267 | total_cycles = strat.cycles[PassCycles.Total] + strat_set.cycles[PassCycles.Total] |
| 268 | bws = strat.bws + strat_set.bws |
| 269 | last_block_height = 0 |
| 270 | if self.options.pareto_metric == ParetoMetric.BwCycMemBlkH and len(strat.block_configs) > 0: |
| 271 | last_block_height = strat.block_configs[-1][0] |
| 272 | |
| 273 | return ( |
| 274 | np.tensordot(bws, self.bandwidth_weights, axes=3) + total_cycles * self.cycles_weight, |
| 275 | strat_set.max_sram_used, |
| 276 | strat.sram_used, |
| 277 | last_block_height, |
| 278 | ) |
| 279 | |
| 280 | def filter_pareto_frontier(self, candidates, remove_equally_good_candidates): |
| 281 | |
| 282 | candidates = [cand for cand in candidates if max(cand[0].sram_used, cand[1].max_sram_used) <= self.sram_limit] |
| 283 | |
| 284 | if len(candidates) <= 1: |
| 285 | return candidates |
| 286 | assert remove_equally_good_candidates |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 287 | pareto_vals = np.zeros((len(candidates), DynamicProgrammingScheduler.num_pareto_metrics)) |
| 288 | ids = np.arange(len(candidates), dtype=np.int32) |
| 289 | for idx, cand in enumerate(candidates): |
| 290 | pareto_vals[idx] = self.pareto_metric(cand) |
| 291 | |
| 292 | sort_order = np.argsort( |
| 293 | pareto_vals.view(DynamicProgrammingScheduler.view_values), |
| 294 | order=DynamicProgrammingScheduler.order_values, |
| 295 | axis=0, |
| 296 | kind="stable", |
| 297 | ).flatten() |
| 298 | pareto_vals = pareto_vals[sort_order] |
| 299 | ids = ids[sort_order] |
| 300 | |
| 301 | pareto_frontier = [] |
| 302 | while len(ids) > 0: |
| 303 | pareto_frontier.append(candidates[ids[0]]) |
| 304 | not_dominated_by_first = (pareto_vals < pareto_vals[0]).any(axis=1) |
| 305 | ids = ids[not_dominated_by_first] |
| 306 | pareto_vals = pareto_vals[not_dominated_by_first] |
| 307 | |
| 308 | if len(pareto_frontier) > self.pareto_max_candidates: |
| 309 | pareto_frontier = self.sort_by_candidate_metric(pareto_frontier) |
| 310 | pareto_frontier = pareto_frontier[: self.pareto_max_candidates] |
| 311 | |
| 312 | return pareto_frontier |
| 313 | |
| 314 | def candidate_metric(self, candidate): |
| 315 | strat, strat_set = candidate |
| 316 | max_sram_used = max(strat_set.max_sram_used, strat.sram_used) |
| 317 | bws = strat.bws + strat_set.bws |
| 318 | total_cycles = strat.cycles[PassCycles.Total] + strat_set.cycles[PassCycles.Total] |
| 319 | |
| 320 | return ( |
| 321 | max_sram_used * self.max_sram_used_weight |
| 322 | + np.tensordot(bws, self.bandwidth_weights, axes=3) |
| 323 | + total_cycles * self.cycles_weight |
| 324 | ) |
| 325 | |
| 326 | def sort_by_candidate_metric(self, candidate_list): |
| 327 | sorted_list = list(sorted(candidate_list, key=self.candidate_metric)) |
| 328 | return sorted_list |
| 329 | |
| 330 | def best_candidate(self, candidate_list): |
| 331 | if len(candidate_list) == 0: |
| 332 | return ABORT_SEARCH |
| 333 | if len(candidate_list) == 1: |
| 334 | return candidate_list[0] |
| 335 | sorted_list = self.sort_by_candidate_metric(candidate_list) |
| 336 | return sorted_list[0] |
| 337 | |
| 338 | def graduate_strat(self, strat_type, sram_used, old_strat_data): |
| 339 | res = [] |
| 340 | for old_strat, old_strat_set in old_strat_data: |
| 341 | if old_strat.sram_used + sram_used > self.sram_limit: |
| 342 | continue # This strategy is bad, drop it |
| 343 | if old_strat_set.max_sram_used > self.sram_limit: |
| 344 | continue # This strategy is bad, drop it |
| 345 | assert old_strat.strat == SchedulingStrategy.Unknown |
| 346 | |
| 347 | new_strat = old_strat.clone() |
| 348 | new_strat.strat = strat_type |
| 349 | new_strat.sram_used = old_strat.sram_used + sram_used |
| 350 | |
| 351 | if self.use_ifm_ofm_overlap: |
| 352 | overlap = calc_allowed_ofm_ifm_overlap_for_pass_list( |
| 353 | new_strat.strat, new_strat.passes, new_strat.block_configs |
| 354 | ) |
| 355 | new_strat.sram_used -= overlap |
| 356 | |
| 357 | new_strat_set = old_strat_set.clone_add_strategy(new_strat) |
| 358 | res.append((empty_strategy, new_strat_set)) |
| 359 | return self.filter_pareto_frontier(res, remove_equally_good_candidates=True) |
| 360 | |
| 361 | def append_sram(self, sram_used, old_strat_data): |
| 362 | res = [] |
| 363 | for old_strat, strat_set in old_strat_data: |
| 364 | assert old_strat.strat == SchedulingStrategy.Unknown |
| 365 | assert old_strat.sram_used == 0 |
| 366 | new_strat = old_strat.clone() |
| 367 | new_strat.sram_used = old_strat.sram_used + sram_used |
| 368 | |
| 369 | res.append((new_strat, strat_set)) |
| 370 | return res |
| 371 | |
| 372 | def append_sram_block_config_performance_metrics(self, sram_used, block_config, metrics, old_strat_data): |
| 373 | res = [] |
| 374 | for old_strat, strat_set in old_strat_data: |
| 375 | assert old_strat.strat == SchedulingStrategy.Unknown |
| 376 | new_strat = old_strat.clone() |
| 377 | bws, macs, cycles = metrics[:3] |
| 378 | |
| 379 | new_strat.sram_used = old_strat.sram_used + sram_used |
| 380 | new_strat.block_configs = old_strat.block_configs + [block_config] |
| 381 | new_strat.bws = old_strat.bws + bws |
| 382 | new_strat.macs = old_strat.macs + macs |
| 383 | new_strat.cycles = old_strat.cycles + cycles |
| 384 | new_strat.bws, new_strat.macs, new_strat.cycles = npu_performance.collate_stats_for_cascaded_pass( |
| 385 | self.arch, new_strat.bws, new_strat.macs, new_strat.cycles |
| 386 | ) |
| 387 | |
| 388 | res.append((new_strat, strat_set)) |
| 389 | return res |
| 390 | |
| 391 | def append_sram_pass_block_config_performance_metrics_rewrite_list( |
| 392 | self, sram_used, new_pass, block_config, metrics, rewrite_list, old_strat_data |
| 393 | ): |
| 394 | res = [] |
| 395 | for old_strat, strat_set in old_strat_data: |
| 396 | assert old_strat.strat == SchedulingStrategy.Unknown |
| 397 | new_strat = old_strat.clone() |
| 398 | bws, macs, cycles = metrics[:3] |
| 399 | new_strat.sram_used = old_strat.sram_used + sram_used |
| 400 | new_strat.block_configs = old_strat.block_configs + [block_config] |
| 401 | new_strat.bws = old_strat.bws + bws |
| 402 | new_strat.macs = old_strat.macs + macs |
| 403 | new_strat.cycles = old_strat.cycles + cycles |
| 404 | new_strat.passes = old_strat.passes + [new_pass] |
| 405 | new_strat.bws, new_strat.macs, new_strat.cycles = npu_performance.collate_stats_for_cascaded_pass( |
| 406 | self.arch, new_strat.bws, new_strat.macs, new_strat.cycles |
| 407 | ) |
| 408 | new_strat.rewrite_list = old_strat.rewrite_list + rewrite_list |
| 409 | res.append((new_strat, strat_set)) |
| 410 | return res |
| 411 | |
| 412 | def append_sram_rewrite_list(self, sram_used, rewrite_list, old_strat_data): |
| 413 | res = [] |
| 414 | for old_strat, strat_set in old_strat_data: |
| 415 | assert old_strat.strat == SchedulingStrategy.Unknown |
| 416 | new_strat = old_strat.clone() |
| 417 | new_strat.sram_used = old_strat.sram_used + sram_used |
| 418 | new_strat.rewrite_list = old_strat.rewrite_list + rewrite_list |
| 419 | res.append((new_strat, strat_set)) |
| 420 | return res |
| 421 | |
| 422 | def pass_to_strat(self, strat_data): |
| 423 | res = {} |
| 424 | for strat in strat_data[1].strats.values(): |
| 425 | for ps in strat.passes: |
| 426 | res[ps] = strat |
| 427 | return res |
| 428 | |
| 429 | def compatible_strats(self, a, b): |
| 430 | intersection = a.keys() & b.keys() |
| 431 | for k in intersection: |
| 432 | if a[k] != b[k]: |
| 433 | return False |
| 434 | return True |
| 435 | |
| 436 | def collate_strats_for_passes(self, all_passes): |
| 437 | if len(all_passes) == 0: |
| 438 | return [(empty_strategy, StrategySet(dict()))] |
| 439 | if len(all_passes) == 1: |
| 440 | return all_passes[0] # save some space in the common case |
| 441 | all_strands = [[self.pass_to_strat(strat_data) for strat_data in strand] for strand in all_passes] |
| 442 | prev_combos = [dict()] |
| 443 | for j, strand in enumerate(all_strands): |
| 444 | new_combos = [] |
| 445 | for i, alt in enumerate(strand): |
| 446 | for prev in prev_combos: |
| 447 | if self.compatible_strats(prev, alt): |
| 448 | cmb = dict(prev) |
| 449 | cmb.update(all_passes[j][i][1].strats) |
| 450 | new_combos.append(cmb) |
| 451 | prev_combos = new_combos |
| 452 | |
| 453 | res = [] |
| 454 | for d in prev_combos: |
| 455 | s = StrategySet(d) |
| 456 | s.update_statistics() |
| 457 | res.append((empty_strategy, s)) |
| 458 | return res |
| 459 | |
| 460 | def search_all_but_one_predecessor(self, ps, pred_pass, pred_pass_data): |
| 461 | # get the rest of the predecessors |
| 462 | other_predecessors = [pred for pred in ps.dag_predecessors if pred != pred_pass] |
| 463 | other_predecessor_data = self.search_pass_list(other_predecessors) |
| 464 | |
| 465 | # pred strat data has an incomplete strategy, which we need |
| 466 | # to continue on, whereas the other ones have completed strategies. |
| 467 | # we need to merge these, but keep the incomplete strategy too. |
| 468 | |
| 469 | res = [] |
| 470 | for pred_pass_strat, pred_pass_strat_set in pred_pass_data: |
| 471 | all_strats = [ |
| 472 | [(empty_strategy, pred_pass_strat_set)], # pred strat data but with a dummy empty strategy |
| 473 | other_predecessor_data, # this one is fine to use as-is |
| 474 | ] |
| 475 | collated_strat_data = self.collate_strats_for_passes(all_strats) |
| 476 | strat_data = [(pred_pass_strat, strat_set) for _, strat_set in collated_strat_data] |
| 477 | res.extend(strat_data) |
| 478 | return res |
| 479 | |
| 480 | def calc_non_local_mem_usage(self): |
| 481 | ignore_subgraph_input_output_tensors = self.sg.placement == PassPlacement.Cpu |
| 482 | range_set = live_range.extract_live_ranges_from_passes( |
| 483 | self.sg, |
| 484 | self.mem_area, |
| 485 | mark_output_tensors_overlapping_with_input_tensors=True, |
| 486 | ignore_subgraph_input_output_tensors=ignore_subgraph_input_output_tensors, |
| 487 | ) |
| 488 | range_dict = range_set.ranges |
| 489 | |
| 490 | # find which ranges overlap passes but aren't input/outputs of the passes. |
| 491 | # these won't be counted by the dynamic programming search and must be counted in manually. |
| 492 | end_pos = max(ps.time for ps in self.sg.passes) + 2 |
| 493 | mem_usage = np.zeros(end_pos) + self.sg.base_sram_used |
| 494 | non_local_mem_usage = np.zeros(end_pos, dtype=np.int64) |
| 495 | |
| 496 | for tens, rng in range_dict.items(): |
| 497 | storage_size = tens.storage_size() |
| 498 | assert tens.mem_area == self.mem_area |
| 499 | mem_usage[rng.start_time : rng.end_time] += storage_size |
| 500 | |
| 501 | for ps in self.sg.passes: |
| 502 | local_mem_usage = 0 |
| 503 | for tens in ps.inputs + ps.outputs + ps.intermediates: |
| 504 | if tens.mem_area != self.mem_area: |
| 505 | continue |
| 506 | |
| 507 | local_mem_usage += tens.storage_size() |
| 508 | |
| 509 | non_local_mem_usage[ps.time] = mem_usage[ps.time] - local_mem_usage |
| 510 | |
| 511 | self.non_local_mem_usage = non_local_mem_usage |
| 512 | |
| 513 | def search(self): |
| 514 | self.calc_non_local_mem_usage() |
| 515 | starting_passes = [ps for ps in self.sg.passes if not ps.successors] |
| 516 | strat_data = self.search_pass_list(starting_passes) |
| 517 | |
| 518 | _, best_set = self.best_candidate(strat_data) |
| 519 | |
| 520 | if self.verbose_pareto_frontier_schedules: |
| 521 | print( |
| 522 | "Scheduler searched %d combinations and found %d candidate schedules along the pareto frontier" |
Diqing Zhong | 504d6b6 | 2020-09-17 12:21:10 +0200 | [diff] [blame] | 523 | % (self.n_combinations_searched, len(strat_data)) |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 524 | ) |
| 525 | for idx, (_, strat_set) in enumerate(strat_data): |
| 526 | extra = "" |
| 527 | if strat_set == best_set: |
| 528 | extra = "(Best candidate)" |
| 529 | print("Candidate", idx, extra) |
| 530 | memory_used = {MemArea.Sram: strat_set.max_sram_used} |
| 531 | stats_writer.print_performance_metrics_for_strat( |
| 532 | self.arch, |
| 533 | "", |
| 534 | strat_set.cycles, |
| 535 | strat_set.macs, |
| 536 | strat_set.bws, |
| 537 | self.nng.batch_size, |
| 538 | memory_used, |
| 539 | len(self.sg.passes), |
| 540 | len(strat_set.strats), |
| 541 | ) |
| 542 | |
| 543 | return best_set |
| 544 | |
| 545 | def search_pass_list(self, pass_list): |
| 546 | all_strats = [] |
| 547 | for ps in pass_list: |
| 548 | strat = self.search_output(ps) |
| 549 | all_strats.append(strat) |
| 550 | strat_data = self.collate_strats_for_passes(all_strats) |
| 551 | for strd in strat_data: |
| 552 | for ps in pass_list: |
| 553 | assert ps in strd[1].strats # should have strategies for everything we asked to search |
| 554 | return strat_data |
| 555 | |
| 556 | def search_predecessors(self, ps): |
| 557 | |
| 558 | # protect against graphs with loops. collate_strats_for_passes will sort this out later so that |
| 559 | # we have strats for all passes |
| 560 | |
| 561 | pass_list = ps.dag_predecessors |
| 562 | strat_data = self.search_pass_list(pass_list) |
| 563 | |
| 564 | return strat_data |
| 565 | |
| 566 | @lru_cache(maxsize=None) |
| 567 | def search_output(self, ps): |
| 568 | |
| 569 | assert ps in self.sg.passes |
| 570 | candidate_list = [] |
| 571 | |
| 572 | candidate_list.extend(self.search_weight_streaming_output(ps)) |
| 573 | |
| 574 | if self.options.use_ifm_streaming: |
| 575 | candidate_list.extend(self.search_ifm_streaming_output(ps)) |
| 576 | |
| 577 | best = self.filter_pareto_frontier(candidate_list, remove_equally_good_candidates=True) |
| 578 | |
| 579 | if not best: |
| 580 | print( |
| 581 | "Warning: Dynamic search programming algorithm failed for pass %s, invoking fallback strategy" |
| 582 | % (ps.name,) |
| 583 | ) |
| 584 | return self.search_predecessors(ps) |
| 585 | |
| 586 | return best |
| 587 | |
| 588 | def search_ifm_streaming_output(self, ps): |
| 589 | if ps.placement != PassPlacement.Npu: |
| 590 | return ABORT_SEARCH |
| 591 | if ps.npu_block_type not in self.ifm_stream_npu_blocks: |
| 592 | return ABORT_SEARCH |
| 593 | strat_data = self.search_ifm_streaming_body(ps, False) |
| 594 | |
| 595 | sram_used = self.non_local_mem_usage[ps.time] |
| 596 | for tens in ps.outputs: |
| 597 | if tens.mem_area == self.mem_area: |
| 598 | sram_used += tens.storage_size() |
| 599 | |
| 600 | return self.graduate_strat(SchedulingStrategy.IfmStream, sram_used, strat_data) |
| 601 | |
| 602 | @lru_cache(maxsize=None) |
| 603 | def search_ifm_streaming_body(self, ps, force_outputs_to_fast_storage): |
| 604 | if ps.placement != PassPlacement.Npu: |
| 605 | return ABORT_SEARCH |
| 606 | if ps.npu_block_type not in self.ifm_stream_npu_blocks: |
| 607 | return ABORT_SEARCH |
| 608 | ifm_input_search_resuls = self.search_ifm_streaming_input(ps) |
| 609 | res = [] |
| 610 | |
| 611 | base_sram_used = 0 |
| 612 | for tens in ps.intermediates: |
| 613 | if tens.mem_area == self.mem_area: |
Patrik Gustavsson | 90831bc | 2020-08-24 16:26:11 +0200 | [diff] [blame] | 614 | if tens.purpose == TensorPurpose.Weights: |
| 615 | base_sram_used = tens.storage_size(self.arch.weight_estimation_scaling) |
| 616 | else: |
| 617 | base_sram_used += tens.storage_size() |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 618 | |
| 619 | all_block_configs = self.get_block_configs(ps) |
| 620 | for block_config in all_block_configs: |
| 621 | all_strats = [] |
| 622 | |
| 623 | if self.use_cascading: |
| 624 | all_strats.extend(self.search_ifm_streaming_partial(ps, block_config)) |
| 625 | |
| 626 | all_strats.extend(ifm_input_search_resuls) |
| 627 | |
| 628 | rewrite_list = [] |
| 629 | sram_used = base_sram_used |
| 630 | |
| 631 | metrics = npu_performance.performance_metrics_for_pass( |
| 632 | self.arch, |
| 633 | ps, |
| 634 | block_config, |
| 635 | rewrite_list=rewrite_list, |
| 636 | force_outputs_to_fast_storage=force_outputs_to_fast_storage, |
| 637 | ) |
| 638 | |
| 639 | res.extend( |
| 640 | self.append_sram_pass_block_config_performance_metrics_rewrite_list( |
| 641 | sram_used, ps, block_config, metrics, rewrite_list, all_strats |
| 642 | ) |
| 643 | ) |
| 644 | |
| 645 | self.n_combinations_searched += len(res) |
| 646 | res = self.filter_pareto_frontier(res, remove_equally_good_candidates=True) |
| 647 | return res |
| 648 | |
Diqing Zhong | 504d6b6 | 2020-09-17 12:21:10 +0200 | [diff] [blame] | 649 | def avoid_for_cascading(self, pred_candidate): |
Patrik Gustavsson | c0bb899 | 2020-08-11 16:45:35 +0200 | [diff] [blame] | 650 | for op in pred_candidate.ops: |
Diqing Zhong | 504d6b6 | 2020-09-17 12:21:10 +0200 | [diff] [blame] | 651 | if ( |
Louis Verhaard | aee5d75 | 2020-09-30 09:01:52 +0200 | [diff] [blame^] | 652 | op.type == Op.ConcatSliceWrite |
Diqing Zhong | 504d6b6 | 2020-09-17 12:21:10 +0200 | [diff] [blame] | 653 | and self.arch.feature_map_storage_mem_area != self.arch.fast_storage_mem_area |
| 654 | ): |
| 655 | # For SRAM spilling, concat op is avoided as predecessor |
Patrik Gustavsson | c0bb899 | 2020-08-11 16:45:35 +0200 | [diff] [blame] | 656 | return True |
Jacob Bohlin | 1a66697 | 2020-09-11 10:04:15 +0200 | [diff] [blame] | 657 | if len(op.outputs) > 1 or len(op.outputs[0].consumer_list) > 1: |
| 658 | # The op has consumers in other subgraphs |
| 659 | return True |
Patrik Gustavsson | c0bb899 | 2020-08-11 16:45:35 +0200 | [diff] [blame] | 660 | return False |
| 661 | |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 662 | def search_ifm_streaming_partial(self, ps, block_config): |
| 663 | if ps.placement != PassPlacement.Npu: |
| 664 | return ABORT_SEARCH |
| 665 | |
| 666 | if len(ps.inputs) < 1: |
| 667 | return ABORT_SEARCH |
| 668 | |
| 669 | ifm_tensor = ps.ifm_tensor |
| 670 | |
| 671 | if ifm_tensor is None: |
| 672 | return ABORT_SEARCH |
| 673 | if ifm_tensor.purpose != TensorPurpose.FeatureMap: |
| 674 | return ABORT_SEARCH |
| 675 | if not ifm_tensor.storage_shape or len(ifm_tensor.storage_shape) != 4: |
| 676 | return ABORT_SEARCH |
| 677 | |
| 678 | pred_pass_list = [] |
| 679 | for pred_candidate in ps.dag_predecessors: |
| 680 | if len(pred_candidate.outputs) == 1 and pred_candidate.outputs[0] == ifm_tensor: |
| 681 | # we found a predecessor that produces this IFM tensor |
Patrik Gustavsson | 458a208 | 2020-08-13 13:41:05 +0200 | [diff] [blame] | 682 | if not ifm_tensor.avoid_NHCWB16: |
| 683 | # and NHCWB16 format is not to be avoided |
| 684 | if len(pred_candidate.successors) == 1 and pred_candidate.successors[0] == ps: |
| 685 | # and it only has one successor, namely us |
| 686 | if pred_candidate.placement == PassPlacement.Npu: |
| 687 | if pred_candidate.npu_block_type in self.ifm_stream_npu_blocks: |
| 688 | # and it is on the Npu |
Diqing Zhong | 504d6b6 | 2020-09-17 12:21:10 +0200 | [diff] [blame] | 689 | if not self.avoid_for_cascading(pred_candidate): |
Patrik Gustavsson | 458a208 | 2020-08-13 13:41:05 +0200 | [diff] [blame] | 690 | # and fusable - it's a candidate |
| 691 | pred_pass_list.append(pred_candidate) |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 692 | |
| 693 | if not pred_pass_list: |
| 694 | return ABORT_SEARCH |
| 695 | |
| 696 | all_candidates = [] |
| 697 | for pred_pass in pred_pass_list: |
| 698 | # recurse into the next pass |
| 699 | ifm_strat_data = self.search_ifm_streaming_body(pred_pass, self.feature_maps_not_in_fast_storage) |
| 700 | |
| 701 | strat_data = self.search_all_but_one_predecessor(ps, pred_pass, ifm_strat_data) |
| 702 | for strat_opt in strat_data: |
| 703 | |
| 704 | pred_pass_block_config = strat_opt[0].block_configs[-1] |
| 705 | rolling_buffer_dims = npu_performance.rolling_buffer_dims_from_passes( |
| 706 | self.arch, pred_pass, pred_pass_block_config, ps, block_config |
| 707 | ) |
| 708 | if rolling_buffer_dims is None: |
| 709 | continue # this does not pack properly, skip it. |
| 710 | |
| 711 | sram_used = 0 |
| 712 | for tens in ps.inputs: |
| 713 | if tens != ifm_tensor: |
| 714 | if tens.mem_area == self.mem_area: |
| 715 | sram_used += tens.storage_size() |
| 716 | |
| 717 | rolling_buffer_y, rolling_buffer_x = rolling_buffer_dims |
| 718 | |
| 719 | rewrite_list = [ |
| 720 | ( |
| 721 | SchedulerRewrite.ChangeTensorSubPurpose, |
| 722 | ifm_tensor, |
| 723 | TensorSubPurpose.RollingBufferY, |
| 724 | rolling_buffer_y, |
| 725 | None, |
| 726 | ps, |
| 727 | ) |
| 728 | ] |
| 729 | sram_used += ifm_tensor.storage_size_for_sub_purpose( |
Patrik Gustavsson | 90831bc | 2020-08-24 16:26:11 +0200 | [diff] [blame] | 730 | self.arch, TensorSubPurpose.RollingBufferY, rolling_buffer_y, None |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 731 | ) |
| 732 | |
| 733 | all_candidates.extend(self.append_sram_rewrite_list(sram_used, rewrite_list, [strat_opt])) |
| 734 | |
| 735 | self.n_combinations_searched += len(all_candidates) |
| 736 | return all_candidates |
| 737 | |
| 738 | def get_block_configs(self, ps): |
| 739 | if ps.placement != PassPlacement.Npu: |
Diego Russo | ea6111a | 2020-04-14 18:41:58 +0100 | [diff] [blame] | 740 | return [(1, 1, 1, 1)] # default |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 741 | |
| 742 | block_configs = find_block_configs_suitable_for_pass_and_shared_buffer(self.arch, ps) |
| 743 | |
| 744 | # Take a limited number of the largest blocks |
| 745 | if self.arch.block_config_limit > 0: |
| 746 | # Sort by block area, followed by depth |
| 747 | block_configs.sort(key=lambda cfg: (cfg[0] * cfg[1]) << 8 | cfg[3], reverse=True) |
| 748 | bound = min(len(block_configs), self.arch.block_config_limit) |
| 749 | # We take 'n' from the fat end of the list, and 'n' from the thin end of the list. |
| 750 | tmp = block_configs[:bound] |
| 751 | tmp.extend(block_configs[max(bound, len(block_configs) - bound) :]) |
| 752 | block_configs = tmp |
| 753 | |
| 754 | return block_configs |
| 755 | |
| 756 | def search_ifm_streaming_input(self, ps): |
| 757 | sram_used = 0 |
| 758 | for tens in ps.inputs: |
| 759 | if tens.mem_area == self.mem_area: |
| 760 | sram_used += tens.storage_size() |
| 761 | |
| 762 | return self.append_sram(sram_used, self.search_predecessors(ps)) |
| 763 | |
| 764 | def search_weight_streaming_output(self, ps): |
| 765 | strat_data = self.search_weight_streaming_body(ps) |
| 766 | |
| 767 | sram_used = self.non_local_mem_usage[ps.time] |
| 768 | for tens in ps.outputs: |
| 769 | if tens.mem_area == self.mem_area: |
| 770 | sram_used += tens.storage_size() |
| 771 | |
| 772 | return self.graduate_strat(SchedulingStrategy.WeightStream, sram_used, strat_data) |
| 773 | |
| 774 | @lru_cache(maxsize=None) |
| 775 | def search_weight_streaming_body(self, ps): |
| 776 | |
| 777 | strat_data = self.search_weight_streaming_input(ps) |
| 778 | |
| 779 | res = [] |
| 780 | |
| 781 | all_block_configs = self.get_block_configs(ps) |
| 782 | |
| 783 | for block_config in all_block_configs: |
| 784 | |
| 785 | sram_used = 0 |
| 786 | rewrite_list = [] |
| 787 | |
| 788 | for tens in ps.intermediates: |
| 789 | if tens.mem_area == self.mem_area: |
| 790 | if tens.purpose == TensorPurpose.Weights: |
Patrik Gustavsson | 90831bc | 2020-08-24 16:26:11 +0200 | [diff] [blame] | 791 | sram_used += tens.storage_size_for_sub_purpose( |
| 792 | self.arch, TensorSubPurpose.DoubleBuffer, block_config[3] |
| 793 | ) |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 794 | rewrite_list.append( |
| 795 | ( |
| 796 | SchedulerRewrite.ChangeTensorSubPurpose, |
| 797 | tens, |
| 798 | TensorSubPurpose.DoubleBuffer, |
| 799 | block_config[3], |
| 800 | None, |
| 801 | ps, |
| 802 | ) |
| 803 | ) |
| 804 | else: |
| 805 | sram_used += tens.storage_size() |
| 806 | |
| 807 | metrics = npu_performance.performance_metrics_for_pass( |
| 808 | self.arch, ps, block_config, rewrite_list=rewrite_list |
| 809 | ) |
| 810 | |
| 811 | res.extend( |
| 812 | self.append_sram_pass_block_config_performance_metrics_rewrite_list( |
| 813 | sram_used, ps, block_config, metrics, rewrite_list, strat_data |
| 814 | ) |
| 815 | ) |
| 816 | |
| 817 | self.n_combinations_searched += len(res) |
| 818 | res = self.filter_pareto_frontier(res, remove_equally_good_candidates=True) |
| 819 | return res |
| 820 | |
| 821 | def search_weight_streaming_input(self, ps): |
| 822 | sram_used = 0 |
| 823 | for tens in ps.inputs: |
| 824 | if tens.mem_area == self.mem_area: |
| 825 | sram_used += tens.storage_size() |
| 826 | |
| 827 | return self.append_sram(sram_used, self.search_predecessors(ps)) |
| 828 | |
| 829 | def apply_result(self, strat_set, arch): |
| 830 | pass_to_cascaded_pass = dict() |
| 831 | for _, strat in strat_set.strats.items(): |
| 832 | # rewrite the tensors that need this first. e.g. make rolling buffers |
| 833 | inputs = [] |
| 834 | intermediates = [] |
| 835 | outputs = [] |
| 836 | |
| 837 | for ps in strat.passes: |
| 838 | inputs += ps.inputs |
| 839 | intermediates += ps.intermediates |
| 840 | outputs += ps.outputs |
| 841 | |
| 842 | for tens in set(inputs) & set(outputs): |
| 843 | # tensors that are in both sets are intermediates |
| 844 | |
| 845 | # find pass with input/output tensor, and check if they are both placed on NPU |
| 846 | input_placement = None |
| 847 | output_placement = None |
| 848 | for ps in strat.passes: |
| 849 | if tens in ps.inputs: |
| 850 | input_placement = ps.placement |
| 851 | if tens in ps.outputs: |
| 852 | output_placement = ps.placement |
| 853 | if input_placement == output_placement == PassPlacement.Npu: |
| 854 | tens.set_format(TensorFormat.NHCWB16, arch) |
| 855 | |
| 856 | intermediates.append(tens) |
| 857 | inputs.remove(tens) |
| 858 | outputs.remove(tens) |
| 859 | |
| 860 | for rewrite_op, tens, sub_purpose, param_a, param_b, ps in strat.rewrite_list: |
| 861 | if rewrite_op == SchedulerRewrite.ChangeTensorSubPurpose: |
| 862 | tens.mem_area = self.arch.fast_storage_mem_area |
Patrik Gustavsson | eca2e95 | 2020-05-27 09:15:11 +0200 | [diff] [blame] | 863 | tens.mem_type = MemType.Scratch_fast |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 864 | tens.set_new_sub_purpose(sub_purpose, param_a, param_b) |
| 865 | else: |
| 866 | assert 0, "unknown rewrite_op " + str(rewrite_op) |
| 867 | |
| 868 | is_element_wise = True |
| 869 | for ps in strat.passes: |
| 870 | assert ps.placement == strat.passes[0].placement |
| 871 | if not ps.is_element_wise: |
| 872 | is_element_wise = False |
| 873 | break |
| 874 | |
| 875 | cascaded_pass = CascadedPass( |
| 876 | strat.passes[0].name, |
| 877 | strat.strat, |
| 878 | inputs, |
| 879 | intermediates, |
| 880 | outputs, |
| 881 | strat.passes, |
| 882 | strat.passes[0].placement, |
| 883 | is_element_wise, |
| 884 | ) |
| 885 | assert strat.sram_used >= 0 |
| 886 | cascaded_pass.sram_used = strat.sram_used |
| 887 | |
| 888 | for idx, ps in enumerate(strat.passes): |
| 889 | assert ps not in pass_to_cascaded_pass |
| 890 | pass_to_cascaded_pass[ps] = cascaded_pass |
| 891 | ps.cascade = cascaded_pass |
| 892 | ps.block_config = strat.block_configs[idx] |
| 893 | |
| 894 | if ps.placement == PassPlacement.Npu: |
| 895 | ps.shared_buffer = shared_buffer_allocation_for_pass_and_block_config( |
| 896 | self.arch, ps, ps.block_config |
| 897 | ) |
| 898 | assert ps.shared_buffer is not None |
| 899 | |
Diqing Zhong | 504d6b6 | 2020-09-17 12:21:10 +0200 | [diff] [blame] | 900 | sram_used = max(self.non_local_mem_usage[ps.time], 0) |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 901 | for op in ps.ops: |
| 902 | subgraph = op.attrs.get("subgraph") |
| 903 | if subgraph: |
Diqing Zhong | 504d6b6 | 2020-09-17 12:21:10 +0200 | [diff] [blame] | 904 | subgraph.base_sram_used = sram_used |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 905 | |
| 906 | # all passes should have a cascaded pass now |
| 907 | if len(pass_to_cascaded_pass) != len(self.sg.passes): |
| 908 | print( |
| 909 | "mismatch: we have %d passes, but only %d have cascaded passes associated" |
| 910 | % (len(self.sg.passes), len(pass_to_cascaded_pass)) |
| 911 | ) |
| 912 | for ps in self.sg.passes: |
Diego Russo | ea6111a | 2020-04-14 18:41:58 +0100 | [diff] [blame] | 913 | if ps not in pass_to_cascaded_pass: |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 914 | print("%3d pass missing cascaded pass %s" % (ps.time, ps)) |
| 915 | |
| 916 | assert len(pass_to_cascaded_pass) == len(self.sg.passes) |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 917 | |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 918 | cascaded_passes = [] |
Charles Xu | 19515e8 | 2020-06-10 10:48:33 +0200 | [diff] [blame] | 919 | if self.sg.placement == PassPlacement.Cpu: |
| 920 | # Retain the pass order for CPU subgraph |
| 921 | cascaded_passes = [ps.cascade for ps in self.sg.passes] |
| 922 | else: |
| 923 | # we have all the passes, but we need to put them in order and build predecessor/successor links. |
| 924 | visit_pass_set = set() |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 925 | |
Charles Xu | 19515e8 | 2020-06-10 10:48:33 +0200 | [diff] [blame] | 926 | def visit_pass(ps): |
| 927 | if ps in visit_pass_set: |
| 928 | return |
| 929 | visit_pass_set.add(ps) |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 930 | |
Charles Xu | 19515e8 | 2020-06-10 10:48:33 +0200 | [diff] [blame] | 931 | cps = ps.cascade |
| 932 | dont_traverse = set(cps.passes) |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 933 | |
Charles Xu | 19515e8 | 2020-06-10 10:48:33 +0200 | [diff] [blame] | 934 | for ps in cps.passes: |
| 935 | for pred in ps.predecessors: |
| 936 | if pred in dont_traverse: |
| 937 | continue |
| 938 | visit_pass(pred) |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 939 | |
Charles Xu | 19515e8 | 2020-06-10 10:48:33 +0200 | [diff] [blame] | 940 | cascaded_passes.append(cps) |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 941 | |
Charles Xu | 19515e8 | 2020-06-10 10:48:33 +0200 | [diff] [blame] | 942 | starting_passes = [ps for ps in self.sg.passes if not ps.successors] |
| 943 | for ps in starting_passes: |
| 944 | visit_pass(ps) |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 945 | |
| 946 | # reorder so startup init cascaded passes come first |
| 947 | def is_startup_cascaded_pass(cps): |
| 948 | if not cps.passes: |
| 949 | return False |
| 950 | return cps.placement == PassPlacement.StartupInit |
| 951 | |
| 952 | cascaded_passes = [cps for cps in cascaded_passes if is_startup_cascaded_pass(cps)] + [ |
| 953 | cps for cps in cascaded_passes if not is_startup_cascaded_pass(cps) |
| 954 | ] |
| 955 | |
| 956 | self.sg.cascaded_passes = cascaded_passes |
| 957 | self.sg.build_cascaded_pass_links() |
| 958 | |
Louis Verhaard | 0b9c9a3 | 2020-09-15 14:05:38 +0200 | [diff] [blame] | 959 | # Check if NHCWB16 and/or fast storage can be used in between cascaded passes |
| 960 | # (NHCWB16 within cascaded passes has been handled earlier in this function) |
| 961 | if self.sg.placement == PassPlacement.Npu: |
| 962 | # Dictionary tensor -> list of ops, containing feature maps that can be attempted |
| 963 | # to be moved to fast storage |
| 964 | fast_storage_tensor_rewrites = {} |
| 965 | last_op_in_subgraph = self.sg.cascaded_passes[-1].passes[-1].primary_op |
| 966 | for ps in self.sg.cascaded_passes: |
| 967 | if ps.placement != PassPlacement.Npu: |
| 968 | continue |
| 969 | for output in ps.outputs: |
| 970 | if output.purpose != TensorPurpose.FeatureMap or output.avoid_NHCWB16: |
Patrik Gustavsson | feeb06d | 2020-04-22 12:53:47 +0200 | [diff] [blame] | 971 | continue |
Louis Verhaard | 0b9c9a3 | 2020-09-15 14:05:38 +0200 | [diff] [blame] | 972 | |
| 973 | use_NHCWB16 = True |
| 974 | use_fast_storage = True |
| 975 | rewrites = [] |
| 976 | for op in output.consumer_list: |
| 977 | if op is None: |
| 978 | use_NHCWB16 = False |
| 979 | use_fast_storage = False |
Charles Xu | 7b8823f | 2020-05-29 13:53:10 +0200 | [diff] [blame] | 980 | continue |
Louis Verhaard | aee5d75 | 2020-09-30 09:01:52 +0200 | [diff] [blame^] | 981 | if op.type == Op.ReduceSum and output.dtype == DataType.int32: |
Louis Verhaard | 0b9c9a3 | 2020-09-15 14:05:38 +0200 | [diff] [blame] | 982 | use_NHCWB16 = False |
Louis Verhaard | aee5d75 | 2020-09-30 09:01:52 +0200 | [diff] [blame^] | 983 | elif op.type == Op.Reshape: |
Louis Verhaard | 0b9c9a3 | 2020-09-15 14:05:38 +0200 | [diff] [blame] | 984 | # Detect no-op reshapes by comparing their full input and output tensor shapes. |
| 985 | inshape = full_shape(4, op.inputs[0].shape, 1) |
| 986 | outshape = full_shape(4, op.outputs[0].shape, 1) |
| 987 | # Using NHCWB16 format for a no-op reshape is only an option if subsequent |
| 988 | # consumers do not also need to perform a reshape or if the OFM is going to |
| 989 | # be processed by CPU operations. No-op reshape consumers with empty lists |
| 990 | # (those that have no consumers, or null-consumers used as list terminators) |
| 991 | # must use normal NHWC output. |
| 992 | incompatible_consumers = [ |
| 993 | ( |
| 994 | not consumer.run_on_npu |
Louis Verhaard | aee5d75 | 2020-09-30 09:01:52 +0200 | [diff] [blame^] | 995 | or consumer.type == Op.Reshape |
Louis Verhaard | 0b9c9a3 | 2020-09-15 14:05:38 +0200 | [diff] [blame] | 996 | or (consumer is last_op_in_subgraph) |
| 997 | ) |
| 998 | for consumer in op.outputs[0].consumer_list |
| 999 | if consumer is not None |
| 1000 | ] |
| 1001 | if (outshape == inshape) and incompatible_consumers and not any(incompatible_consumers): |
| 1002 | rewrites.append(op) |
Tim Hall | ba69518 | 2020-08-26 17:27:19 +0100 | [diff] [blame] | 1003 | else: |
Louis Verhaard | 0b9c9a3 | 2020-09-15 14:05:38 +0200 | [diff] [blame] | 1004 | use_NHCWB16 = False |
| 1005 | use_fast_storage = False |
| 1006 | use_NHCWB16 &= op.run_on_npu |
| 1007 | use_fast_storage &= op.run_on_npu |
Patrik Gustavsson | feeb06d | 2020-04-22 12:53:47 +0200 | [diff] [blame] | 1008 | |
Louis Verhaard | 0b9c9a3 | 2020-09-15 14:05:38 +0200 | [diff] [blame] | 1009 | if use_fast_storage: |
| 1010 | fast_storage_tensor_rewrites[output] = rewrites |
| 1011 | if use_NHCWB16 and self.options.use_nhcwb16_between_cascaded_passes: |
| 1012 | output.set_format(TensorFormat.NHCWB16, arch) |
| 1013 | for rewrite_op in rewrites: |
| 1014 | rewrite_op.outputs[0].set_format(TensorFormat.NHCWB16, arch) |
| 1015 | if self.feature_maps_not_in_fast_storage: |
| 1016 | # Remember feature maps that can be moved to fast storage for later use |
| 1017 | # in use_fast_storage_for_feature_maps |
| 1018 | self.sg.scheduling_info["feature_map_rewrites"] = fast_storage_tensor_rewrites |
Patrik Gustavsson | feeb06d | 2020-04-22 12:53:47 +0200 | [diff] [blame] | 1019 | |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 1020 | |
| 1021 | def schedule_passes(nng, arch, options: SchedulerOptions): |
| 1022 | |
| 1023 | for sg in nng.subgraphs: |
| 1024 | sg.base_sram_used = 0 |
| 1025 | |
| 1026 | for sg in nng.subgraphs: |
| 1027 | # re-entering the same nodes from different contexts requires us to |
| 1028 | # build a simplified directed acyclic (DAG) version of the graph to |
| 1029 | # use for traversal, rather than using a visit dictionary. this avoids |
| 1030 | # recursing infinitely due to loops. |
| 1031 | sg.build_pass_dag_predecessors() |
| 1032 | |
| 1033 | dps = DynamicProgrammingScheduler(nng, sg, arch, arch.sram_size, options) |
| 1034 | |
| 1035 | strat_set = dps.search() |
| 1036 | |
| 1037 | dps.apply_result(strat_set, arch) |
| 1038 | |
| 1039 | if options.verbose_schedule: |
| 1040 | sg.print_cascaded_passes() |
Louis Verhaard | 0b9c9a3 | 2020-09-15 14:05:38 +0200 | [diff] [blame] | 1041 | |
| 1042 | |
| 1043 | def _calc_tens_to_cps(sg, tensor_rewrites): |
| 1044 | # Determines for each tensor the list of affected cascaded passes, in terms of SRAM consumption. |
| 1045 | # Returns dictionary tensor -> list of cascaded passes |
| 1046 | # Note: if cascaded passes are A, B, C, D, and a tensor is output |
| 1047 | # of A and input to D, then it also consumes SRAM in passes B and C. |
| 1048 | if "tens_to_cps" in sg.scheduling_info: |
| 1049 | return sg.scheduling_info["tens_to_cps"] |
| 1050 | # Determine life-time of tensors |
| 1051 | min_index = {} |
| 1052 | max_index = {} |
| 1053 | index = 0 |
| 1054 | cps_list = [cps for cps in sg.cascaded_passes if cps.placement == PassPlacement.Npu] |
| 1055 | for cps in cps_list: |
| 1056 | for tens in cps.inputs + cps.outputs: |
| 1057 | if tens in tensor_rewrites: |
| 1058 | min_index[tens] = min(index, min_index.get(tens, len(cps_list))) |
| 1059 | max_index[tens] = index |
| 1060 | index += 1 |
| 1061 | # Convert to affected cps-es |
| 1062 | tens_to_cps = {} |
| 1063 | for tens in min_index: |
| 1064 | tens_to_cps[tens] = cps_list[min_index[tens] : max_index[tens] + 1] |
| 1065 | sg.scheduling_info["tens_to_cps"] = tens_to_cps |
| 1066 | return tens_to_cps |
| 1067 | |
| 1068 | |
| 1069 | def use_fast_storage_for_feature_maps(sg, sram_limit, arch): |
| 1070 | # Attempts to use as much fast storage as possible for feature maps shared between cascaded passes. |
| 1071 | tensor_rewrites = sg.scheduling_info.get("feature_map_rewrites", {}) |
| 1072 | tens_to_cps = _calc_tens_to_cps(sg, tensor_rewrites) |
| 1073 | # Sort tensors first on life-time (smallest first), then on size (biggest first) |
| 1074 | tens_list = sorted([(len(tens_to_cps[tens]), -tens.storage_size(), tens.name, tens) for tens in tens_to_cps]) |
| 1075 | for _, _, _, tens in tens_list: |
| 1076 | cps_list = tens_to_cps[tens] |
| 1077 | if len(cps_list) <= 1: |
| 1078 | continue |
| 1079 | sz = tens.storage_size() |
| 1080 | fits_in_fast_storage = all([cps.sram_used + sz <= sram_limit for cps in cps_list]) |
| 1081 | if fits_in_fast_storage: |
| 1082 | tens.mem_area = arch.fast_storage_mem_area |
| 1083 | tens.mem_type = MemType.Scratch_fast |
| 1084 | tens.set_new_sub_purpose(TensorSubPurpose.Standard, None, None) |
| 1085 | assert tens in tensor_rewrites |
| 1086 | # Also rewrite reshapes |
| 1087 | for rewrite_op in tensor_rewrites[tens]: |
| 1088 | tens2 = rewrite_op.outputs[0] |
| 1089 | tens2.mem_area = arch.fast_storage_mem_area |
| 1090 | tens2.mem_type = MemType.Scratch_fast |
| 1091 | tens2.set_new_sub_purpose(TensorSubPurpose.Standard, None, None) |
| 1092 | for cps in cps_list: |
| 1093 | cps.sram_used += sz |
| 1094 | |
| 1095 | |
| 1096 | def undo_use_fast_storage(sg, arch): |
| 1097 | # Undoes the effects of a previous call to use_fast_storage_for_feature_maps |
| 1098 | tensor_rewrites = sg.scheduling_info.get("feature_map_rewrites", {}) |
| 1099 | tens_to_cps = _calc_tens_to_cps(sg, tensor_rewrites) |
| 1100 | mem_area = arch.tensor_storage_mem_area[TensorPurpose.FeatureMap] |
| 1101 | for tens, cps_list in tens_to_cps.items(): |
| 1102 | if tens.mem_type == MemType.Scratch_fast: |
| 1103 | sz = tens.storage_size() |
| 1104 | tens.mem_area = mem_area |
| 1105 | tens.mem_type = MemType.Scratch |
| 1106 | # Also undo reshapes |
| 1107 | for rewrite_op in tensor_rewrites[tens]: |
| 1108 | tens2 = rewrite_op.outputs[0] |
| 1109 | tens2.mem_area = mem_area |
| 1110 | tens2.mem_type = MemType.Scratch |
| 1111 | for cps in cps_list: |
| 1112 | cps.sram_used -= sz |