Tim Hall | d8339a7 | 2021-05-27 18:49:40 +0100 | [diff] [blame] | 1 | # Copyright (C) 2021 Arm Limited or its affiliates. All rights reserved. |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 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 | d8339a7 | 2021-05-27 18:49:40 +0100 | [diff] [blame] | 16 | # |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 17 | # Description: |
Tim Hall | d8339a7 | 2021-05-27 18:49:40 +0100 | [diff] [blame] | 18 | # The scheduler creates and searches for an optimal plan for the network, selecting block configurations and |
| 19 | # subdivisions for the Operators |
Jonas Ohlsson | 845e232 | 2022-03-01 12:39:55 +0100 | [diff] [blame] | 20 | # For Class name forward references for the type annotations. (see PEP 563). |
| 21 | from __future__ import annotations |
| 22 | |
Diego Russo | ea6111a | 2020-04-14 18:41:58 +0100 | [diff] [blame] | 23 | import copy |
Johan Alfvén | 5e0ae55 | 2022-02-09 21:20:10 +0100 | [diff] [blame] | 24 | from collections import namedtuple |
Tim Hall | d8339a7 | 2021-05-27 18:49:40 +0100 | [diff] [blame] | 25 | from enum import auto |
| 26 | from enum import IntEnum |
| 27 | from typing import Dict |
| 28 | from typing import List |
| 29 | from typing import Optional |
| 30 | from typing import Tuple |
Jonas Ohlsson | 845e232 | 2022-03-01 12:39:55 +0100 | [diff] [blame] | 31 | from typing import TYPE_CHECKING |
| 32 | |
| 33 | # Import needed for Type annotations. Only import for Type checking to avoid run-time errors due to cyclic import. |
| 34 | if TYPE_CHECKING: |
| 35 | from .npu_performance import CycleCost |
Diego Russo | ea6111a | 2020-04-14 18:41:58 +0100 | [diff] [blame] | 36 | |
erik.andersson@arm.com | de6cb64 | 2022-02-02 14:03:15 +0100 | [diff] [blame] | 37 | import numpy as np |
| 38 | |
Diego Russo | ea6111a | 2020-04-14 18:41:58 +0100 | [diff] [blame] | 39 | from . import live_range |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 40 | from . import npu_performance |
Tim Hall | d8339a7 | 2021-05-27 18:49:40 +0100 | [diff] [blame] | 41 | from . import tensor_allocation |
| 42 | from . import weight_compressor |
| 43 | from .architecture_allocator import ArchitectureBlockConfig |
| 44 | from .architecture_allocator import find_block_config |
| 45 | from .architecture_allocator import get_ifm_area_required |
Tim Hall | d8339a7 | 2021-05-27 18:49:40 +0100 | [diff] [blame] | 46 | from .architecture_features import ArchitectureFeatures |
| 47 | from .architecture_features import Block |
| 48 | from .cascade_builder import CascadeBuilder |
| 49 | from .cascade_builder import CascadeInfo |
Fredrik Svedberg | 880e735 | 2020-08-25 11:31:47 +0200 | [diff] [blame] | 50 | from .data_type import DataType |
Diego Russo | e8a1045 | 2020-04-21 17:39:10 +0100 | [diff] [blame] | 51 | from .nn_graph import CascadedPass |
Tim Hall | d8339a7 | 2021-05-27 18:49:40 +0100 | [diff] [blame] | 52 | from .nn_graph import Graph |
| 53 | from .nn_graph import Pass |
Diego Russo | e8a1045 | 2020-04-21 17:39:10 +0100 | [diff] [blame] | 54 | from .nn_graph import PassPlacement |
Diego Russo | e8a1045 | 2020-04-21 17:39:10 +0100 | [diff] [blame] | 55 | from .nn_graph import SchedulingStrategy |
Tim Hall | d8339a7 | 2021-05-27 18:49:40 +0100 | [diff] [blame] | 56 | from .nn_graph import Subgraph |
| 57 | from .numeric_util import round_down |
| 58 | from .numeric_util import round_up |
Diego Russo | e8a1045 | 2020-04-21 17:39:10 +0100 | [diff] [blame] | 59 | from .operation import NpuBlockType |
Louis Verhaard | aee5d75 | 2020-09-30 09:01:52 +0200 | [diff] [blame] | 60 | from .operation import Op |
Tim Hall | d8339a7 | 2021-05-27 18:49:40 +0100 | [diff] [blame] | 61 | from .shape4d import Shape4D |
Diego Russo | e8a1045 | 2020-04-21 17:39:10 +0100 | [diff] [blame] | 62 | from .tensor import MemArea |
Patrik Gustavsson | eca2e95 | 2020-05-27 09:15:11 +0200 | [diff] [blame] | 63 | from .tensor import MemType |
Tim Hall | d8339a7 | 2021-05-27 18:49:40 +0100 | [diff] [blame] | 64 | from .tensor import Tensor |
Diego Russo | e8a1045 | 2020-04-21 17:39:10 +0100 | [diff] [blame] | 65 | from .tensor import TensorFormat |
| 66 | from .tensor import TensorPurpose |
| 67 | from .tensor import TensorSubPurpose |
Jonas Ohlsson | 845e232 | 2022-03-01 12:39:55 +0100 | [diff] [blame] | 68 | from .weight_compressor import NpuWeightTensor |
Jacob Bohlin | 1a66697 | 2020-09-11 10:04:15 +0200 | [diff] [blame] | 69 | |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 70 | |
Tim Hall | d8339a7 | 2021-05-27 18:49:40 +0100 | [diff] [blame] | 71 | def shape_for_format(shape: Shape4D, tensor_format: TensorFormat) -> Shape4D: |
| 72 | if tensor_format == TensorFormat.NHCWB16: |
| 73 | return shape.with_depth(round_up(shape.depth, 16)) |
| 74 | |
| 75 | return shape |
| 76 | |
| 77 | |
| 78 | class OptimizationStrategy(IntEnum): |
| 79 | """Enum defining the different optimization strategies for the Scheduler""" |
| 80 | |
| 81 | Size = auto() |
| 82 | Performance = auto() |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 83 | |
| 84 | def __str__(self): |
| 85 | return self.name |
| 86 | |
| 87 | |
Tim Hall | d8339a7 | 2021-05-27 18:49:40 +0100 | [diff] [blame] | 88 | class SchedulerOpInfo: |
| 89 | """Contains metadata about a SchedulerOperation that is unique to one Schedule""" |
| 90 | |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 91 | def __init__( |
| 92 | self, |
Tim Hall | d8339a7 | 2021-05-27 18:49:40 +0100 | [diff] [blame] | 93 | block_config: ArchitectureBlockConfig, |
| 94 | weights_size: int, |
| 95 | stripe_input: Shape4D, |
| 96 | stripe_input2: Optional[Shape4D], |
| 97 | stripe: Shape4D, |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 98 | ): |
Tim Hall | d8339a7 | 2021-05-27 18:49:40 +0100 | [diff] [blame] | 99 | self.block_config = block_config |
| 100 | self.weights_size = weights_size |
| 101 | self.stripe_input = stripe_input |
| 102 | self.stripe_input2 = stripe_input2 |
| 103 | self.stripe = stripe |
| 104 | self.cascade = 0 # Assigned by CascadeBuilder. 0 means not part of a cascade |
| 105 | self.time_index = None # Set by update_op_memory_snapshot |
| 106 | self.ofm_depth_slices: List[int] = [0, stripe.depth] |
Jonas Ohlsson | 845e232 | 2022-03-01 12:39:55 +0100 | [diff] [blame] | 107 | self.npu_weights_tensor: Optional[NpuWeightTensor] = None |
| 108 | self.npu_scales_tensor: Optional[NpuWeightTensor] = None |
Tim Hall | b5df773 | 2022-05-04 16:20:43 +0100 | [diff] [blame^] | 109 | self.buffered_weight_tensor: Optional[Tensor] = None |
Jonas Ohlsson | 845e232 | 2022-03-01 12:39:55 +0100 | [diff] [blame] | 110 | self.cycles: Optional[CycleCost] = None |
Tim Hall | d8339a7 | 2021-05-27 18:49:40 +0100 | [diff] [blame] | 111 | self.slack_buffering_cycles = 0 |
| 112 | self.slack_buffering_memory = 0 |
| 113 | self.full_weight_transfer_cycles = 0 |
| 114 | |
| 115 | def copy(self): |
Jonas Ohlsson | d857507 | 2022-03-30 10:30:25 +0200 | [diff] [blame] | 116 | res = SchedulerOpInfo( |
| 117 | self.block_config, |
| 118 | self.weights_size, |
| 119 | self.stripe_input, |
| 120 | self.stripe_input2, |
| 121 | self.stripe, |
| 122 | ) |
Tim Hall | d8339a7 | 2021-05-27 18:49:40 +0100 | [diff] [blame] | 123 | res.cascade = self.cascade |
| 124 | return res |
| 125 | |
| 126 | def __str__(self): |
| 127 | res = f"\t\tBlock Config = {self.block_config}\n" |
| 128 | res += f"\t\tOFM Block = {self.block_config.ofm_block}\n" |
| 129 | res += f"\t\tIFM Stripe = {self.stripe_input}\n" |
| 130 | res += f"\t\tIFM2 Stripe = {self.stripe_input2}\n" |
| 131 | res += f"\t\tOFM Stripe = {self.stripe}\n" |
| 132 | res += f"\t\tEncoded Weights = {self.npu_weights_tensor and len(self.npu_weights_tensor.buffer)} bytes\n" |
Tim Hall | b5df773 | 2022-05-04 16:20:43 +0100 | [diff] [blame^] | 133 | res += ( |
| 134 | f"\t\tWeight buffer = {self.buffered_weight_tensor and self.buffered_weight_tensor.storage_size()} bytes\n" |
| 135 | ) |
Tim Hall | d8339a7 | 2021-05-27 18:49:40 +0100 | [diff] [blame] | 136 | res += f"\t\tDepth slices = {self.ofm_depth_slices}\n" |
| 137 | res += f"\t\tAssigned Cascade = {self.cascade}" |
| 138 | return res |
| 139 | |
| 140 | |
| 141 | class SchedulerOptions: |
| 142 | """Contains options for the Scheduler""" |
| 143 | |
| 144 | def __init__( |
Jonas Ohlsson | d857507 | 2022-03-30 10:30:25 +0200 | [diff] [blame] | 145 | self, |
| 146 | optimization_strategy, |
| 147 | sram_target, |
| 148 | verbose_schedule, |
Tim Hall | d8339a7 | 2021-05-27 18:49:40 +0100 | [diff] [blame] | 149 | ): |
| 150 | self.optimization_strategy = optimization_strategy |
| 151 | self.optimization_sram_limit = sram_target |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 152 | self.verbose_schedule = verbose_schedule |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 153 | |
Tim Hall | d8339a7 | 2021-05-27 18:49:40 +0100 | [diff] [blame] | 154 | def __str__(self) -> str: |
| 155 | return f"{type(self).__name__}: {str(self.__dict__)}" |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 156 | |
| 157 | __repr__ = __str__ |
| 158 | |
| 159 | |
Tim Hall | d8339a7 | 2021-05-27 18:49:40 +0100 | [diff] [blame] | 160 | class SchedulerTensor: |
| 161 | def __init__(self, shape, dt, mem_area, _format): |
| 162 | self.dtype = dt |
| 163 | self.mem_area = mem_area |
| 164 | self.shape = shape |
| 165 | self.format = _format |
| 166 | self.connection = None |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 167 | |
Tim Hall | d8339a7 | 2021-05-27 18:49:40 +0100 | [diff] [blame] | 168 | |
| 169 | class SchedulerOperation: |
| 170 | """Scheduler internal representation of 'Operation' |
| 171 | This class can be seen as a node within the Scheduler Graph representation |
| 172 | """ |
| 173 | |
| 174 | def __init__(self, ps: Pass, arch: ArchitectureFeatures, nng: Graph): |
| 175 | self.arch = arch |
| 176 | self.parent_ps = ps |
| 177 | self.parent_op = ps.primary_op |
| 178 | self.name = ps.primary_op.name |
| 179 | self.op_type = ps.primary_op.type |
| 180 | self.activation = ps.primary_op.activation |
| 181 | self.kernel = ps.primary_op.kernel |
Tim Hall | 3c5cfe9 | 2022-03-16 16:31:57 +0000 | [diff] [blame] | 182 | self.resampling_mode = ps.primary_op.ifm_resampling_mode |
Tim Hall | d8339a7 | 2021-05-27 18:49:40 +0100 | [diff] [blame] | 183 | self.uses_scalar = ps.primary_op.ifm2 is not None and ( |
| 184 | ps.primary_op.ifm.shape == [] or ps.primary_op.ifm2.shape == [] |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 185 | ) |
Tim Hall | d8339a7 | 2021-05-27 18:49:40 +0100 | [diff] [blame] | 186 | self.ifm_ublock = arch.ifm_ublock |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 187 | |
Jonas Ohlsson | d857507 | 2022-03-30 10:30:25 +0200 | [diff] [blame] | 188 | self.ifm = SchedulerTensor( |
| 189 | ps.ifm_shapes[0], |
| 190 | ps.ifm_tensor.dtype, |
| 191 | ps.ifm_tensor.mem_area, |
| 192 | ps.ifm_tensor.format, |
| 193 | ) |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 194 | |
Tim Hall | d8339a7 | 2021-05-27 18:49:40 +0100 | [diff] [blame] | 195 | self.ifm2 = None |
| 196 | if ps.ifm2_tensor: |
| 197 | self.ifm2 = SchedulerTensor( |
Jonas Ohlsson | d857507 | 2022-03-30 10:30:25 +0200 | [diff] [blame] | 198 | ps.ifm_shapes[1], |
| 199 | ps.ifm2_tensor.dtype, |
| 200 | ps.ifm2_tensor.mem_area, |
| 201 | ps.ifm2_tensor.format, |
Tim Hall | d8339a7 | 2021-05-27 18:49:40 +0100 | [diff] [blame] | 202 | ) |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 203 | |
Jonas Ohlsson | d857507 | 2022-03-30 10:30:25 +0200 | [diff] [blame] | 204 | self.ofm = SchedulerTensor( |
| 205 | ps.ofm_shapes[0], |
| 206 | ps.ofm_tensor.dtype, |
| 207 | ps.ofm_tensor.mem_area, |
| 208 | ps.ofm_tensor.format, |
| 209 | ) |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 210 | |
Tim Hall | d8339a7 | 2021-05-27 18:49:40 +0100 | [diff] [blame] | 211 | # Input volume width and height required to produce the smallest possible stripe |
| 212 | self.min_stripe_input_w, self.min_stripe_input_h = self._calculate_min_stripe_input() |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 213 | |
Tim Hall | d8339a7 | 2021-05-27 18:49:40 +0100 | [diff] [blame] | 214 | # Flags that marks whether this SchedulerOperation requires full IFM/OFM |
| 215 | self.requires_full_ifm = False |
| 216 | self.requires_full_ifm2 = False |
| 217 | self.requires_full_ofm = False |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 218 | |
Tim Hall | d8339a7 | 2021-05-27 18:49:40 +0100 | [diff] [blame] | 219 | self.index = 0 |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 220 | |
Tim Hall | d8339a7 | 2021-05-27 18:49:40 +0100 | [diff] [blame] | 221 | def add_ifm_connection(self, conn: "Connection"): |
| 222 | """Add input connection to another SchedulerOperation or Subgraph Input""" |
| 223 | conn.consumers.append(self) |
| 224 | self.ifm.connection = conn |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 225 | |
Tim Hall | d8339a7 | 2021-05-27 18:49:40 +0100 | [diff] [blame] | 226 | def add_ifm2_connection(self, conn: "Connection"): |
| 227 | """Add input connection to another SchedulerOperation or Subgraph Input""" |
| 228 | if self.ifm2: |
| 229 | conn.consumers.append(self) |
| 230 | self.ifm2.connection = conn |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 231 | else: |
Tim Hall | d8339a7 | 2021-05-27 18:49:40 +0100 | [diff] [blame] | 232 | assert False, f"Trying to set an IFM2 Connection to {self} which has no IFM2" |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 233 | |
Tim Hall | d8339a7 | 2021-05-27 18:49:40 +0100 | [diff] [blame] | 234 | def add_ofm_connection(self, conn: "Connection"): |
| 235 | """Add output connection to another SchedulerOperation or Subgraph Output""" |
| 236 | conn.producers.append(self) |
| 237 | self.ofm.connection = conn |
| 238 | |
| 239 | def get_dependants(self): |
| 240 | """Returns a list of the Ops that depend on this Operation's OFM""" |
| 241 | return self.ofm.connection.consumers |
| 242 | |
| 243 | def ifm_size_in_bytes(self) -> int: |
| 244 | """Returns size of the IFM in bytes""" |
| 245 | ifm_storage_shape = shape_for_format(self.ifm.shape, self.ifm.format) |
| 246 | return round_up(ifm_storage_shape.elements() * self.ifm.dtype.size_in_bytes(), Tensor.AllocationQuantum) |
| 247 | |
| 248 | def ifm2_size_in_bytes(self) -> int: |
| 249 | """Returns size of the IFM2 in bytes""" |
| 250 | if self.ifm2: |
| 251 | ifm2_storage_shape = shape_for_format(self.ifm2.shape, self.ifm2.format) |
| 252 | return round_up(ifm2_storage_shape.elements() * self.ifm2.dtype.size_in_bytes(), Tensor.AllocationQuantum) |
| 253 | |
| 254 | return 0 |
| 255 | |
| 256 | def ofm_size_in_bytes(self) -> int: |
| 257 | """Returns size of the OFM in bytes""" |
| 258 | ofm_storage_shape = shape_for_format(self.ofm.shape, self.ofm.format) |
| 259 | return round_up(ofm_storage_shape.elements() * self.ofm.dtype.size_in_bytes(), Tensor.AllocationQuantum) |
| 260 | |
| 261 | def create_scheduler_info(self, nng: Graph, stripe: Shape4D) -> SchedulerOpInfo: |
| 262 | """Returns schedule info about this SchedulerOperation based on how many ofm elements it should produce""" |
| 263 | ifm_shape = self.ifm.shape |
Jonas Ohlsson | 845e232 | 2022-03-01 12:39:55 +0100 | [diff] [blame] | 264 | ifm2_shape = self.ifm2.shape if self.ifm2 is not None else None |
Tim Hall | d8339a7 | 2021-05-27 18:49:40 +0100 | [diff] [blame] | 265 | ofm_shape = stripe |
| 266 | |
| 267 | if ofm_shape != self.ofm.shape: |
| 268 | # Striped Op - Need to calculate stripe input volume |
| 269 | stripe_input_w, stripe_input_h = self._get_stripe_input_requirement(stripe) |
| 270 | # Ensure stripe input volume is within the full IFM volume |
| 271 | stripe_input_h = min(stripe_input_h, self.ifm.shape.height) |
| 272 | stripe_input_w = min(stripe_input_w, self.ifm.shape.width) |
| 273 | ifm_shape = ifm_shape.with_hw(stripe_input_h, stripe_input_w) |
| 274 | |
| 275 | if self.ifm2: |
| 276 | stripe_input2_h = min(stripe_input_h, self.ifm2.shape.height) |
| 277 | stripe_input2_w = min(stripe_input_w, self.ifm2.shape.width) |
| 278 | ifm2_shape = ifm2_shape.with_hw(stripe_input2_h, stripe_input2_w) |
| 279 | |
| 280 | block_config = self._get_block_config(ifm_shape, ifm2_shape, self.uses_scalar, ofm_shape) |
| 281 | |
| 282 | scheduler_op_info = SchedulerOpInfo(block_config, 0, ifm_shape, ifm2_shape, ofm_shape) |
| 283 | if self.parent_op.weights: |
| 284 | # Default full-depth weight encoding with no buffering |
Tim Hall | d784af7 | 2021-06-08 21:25:57 +0100 | [diff] [blame] | 285 | ( |
| 286 | scheduler_op_info.npu_weights_tensor, |
| 287 | scheduler_op_info.npu_scales_tensor, |
| 288 | ) = weight_compressor.encode_weight_and_scale_tensor( |
Tim Hall | d8339a7 | 2021-05-27 18:49:40 +0100 | [diff] [blame] | 289 | self.arch, |
| 290 | self.parent_op, |
| 291 | self.parent_op.weights, |
| 292 | self.parent_op.bias, |
| 293 | self.kernel, |
| 294 | block_config, |
| 295 | [0, self.ofm.shape.depth], |
| 296 | ) |
| 297 | |
| 298 | self.parent_ps.block_config = block_config.old_style_representation() |
| 299 | return scheduler_op_info |
| 300 | |
| 301 | def _get_stripe_input_requirement(self, stripe_shape: Shape4D) -> Tuple[int, int]: |
| 302 | """Returns the amount of IFM required to produce the stripe with shape:'stripe_shape'""" |
| 303 | ofm_shape_to_produce = Block.from_shape(stripe_shape.as_list()) |
| 304 | |
Fredrik Svedberg | 3ff7a4a | 2021-09-29 10:08:04 +0200 | [diff] [blame] | 305 | return get_ifm_area_required(ofm_shape_to_produce, self.kernel, self.resampling_mode) |
Tim Hall | d8339a7 | 2021-05-27 18:49:40 +0100 | [diff] [blame] | 306 | |
Jonas Ohlsson | 845e232 | 2022-03-01 12:39:55 +0100 | [diff] [blame] | 307 | def _calculate_min_stripe_input(self) -> Tuple[int, int]: |
Tim Hall | d8339a7 | 2021-05-27 18:49:40 +0100 | [diff] [blame] | 308 | # Calculate the input volume required height and width for the smallest possible stripe (h,w = 1,1) |
| 309 | min_stripe = self.ofm.shape.with_hw(1, 1) |
| 310 | return self._get_stripe_input_requirement(min_stripe) |
| 311 | |
| 312 | def _get_block_config( |
| 313 | self, ifm_shape: Shape4D, ifm2_shape: Optional[Shape4D], uses_scalar: bool, ofm_shape: Shape4D |
Jonas Ohlsson | 845e232 | 2022-03-01 12:39:55 +0100 | [diff] [blame] | 314 | ) -> Optional[ArchitectureBlockConfig]: |
Tim Hall | d8339a7 | 2021-05-27 18:49:40 +0100 | [diff] [blame] | 315 | # Returns a block config and SHRAM layout |
| 316 | lut_banks = 2 if self.parent_op.activation_lut else 0 |
| 317 | return find_block_config( |
| 318 | self.arch, |
| 319 | self.op_type.npu_block_type, |
| 320 | ofm_shape, |
| 321 | ifm_shape, |
| 322 | ifm2_shape, |
| 323 | uses_scalar, |
| 324 | self.ifm.dtype.size_in_bits(), |
| 325 | self.kernel, |
| 326 | lut_banks, |
| 327 | self.parent_op.has_scaling(), |
| 328 | self.resampling_mode, |
| 329 | ) |
| 330 | |
| 331 | |
| 332 | class Connection: |
| 333 | """Scheduler internal representation of a Tensor that connects two SchedulerOperations |
| 334 | This class can be seen as an edge within the Scheduler Graph representation |
| 335 | """ |
| 336 | |
| 337 | def __init__(self, tensor: Tensor): |
| 338 | self.parent_tens = tensor |
| 339 | |
| 340 | # SchedulerOperation relationships |
| 341 | self.producers: List[SchedulerOperation] = [] |
| 342 | self.consumers: List[SchedulerOperation] = [] |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 343 | |
| 344 | def __str__(self): |
Tim Hall | d8339a7 | 2021-05-27 18:49:40 +0100 | [diff] [blame] | 345 | return f"<Connection {self.parent_tens.name}>" |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 346 | |
| 347 | __repr__ = __str__ |
| 348 | |
| 349 | |
Tim Hall | d8339a7 | 2021-05-27 18:49:40 +0100 | [diff] [blame] | 350 | class Schedule: |
| 351 | """Class that contains a solution of how to schedule an NPU subgraph and its cost""" |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 352 | |
Tim Hall | d8339a7 | 2021-05-27 18:49:40 +0100 | [diff] [blame] | 353 | def __init__(self, sg: Subgraph, label: str): |
| 354 | self.sg = sg |
| 355 | self.label = label |
| 356 | self.cost_map: Dict[SchedulerOperation, SchedulerOpInfo] = {} |
| 357 | self.cascades: Dict[int, CascadeInfo] = {} |
| 358 | self.fast_storage_peak_usage = 0 |
Jonas Ohlsson | 845e232 | 2022-03-01 12:39:55 +0100 | [diff] [blame] | 359 | self.memory_snapshot: Optional[List[int]] = None |
Tim Hall | d8339a7 | 2021-05-27 18:49:40 +0100 | [diff] [blame] | 360 | |
| 361 | @property |
| 362 | def name(self): |
| 363 | return f"{self.sg.name}_{self.label}" |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 364 | |
| 365 | |
Tim Hall | d8339a7 | 2021-05-27 18:49:40 +0100 | [diff] [blame] | 366 | class Scheduler: |
| 367 | """Main class of the Vela Scheduling""" |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 368 | |
Tim Hall | d8339a7 | 2021-05-27 18:49:40 +0100 | [diff] [blame] | 369 | def __init__(self, nng: Graph, sg: Subgraph, arch: ArchitectureFeatures, options: SchedulerOptions): |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 370 | self.nng = nng |
| 371 | self.sg = sg |
| 372 | self.arch = arch |
Ayaan Masood | b801dda | 2022-02-22 11:28:55 +0000 | [diff] [blame] | 373 | self.sched_ops: List[SchedulerOperation] = [] |
Jonas Ohlsson | 845e232 | 2022-03-01 12:39:55 +0100 | [diff] [blame] | 374 | self.max_schedule: Optional[Schedule] = None |
Tim Hall | d8339a7 | 2021-05-27 18:49:40 +0100 | [diff] [blame] | 375 | self.scheduler_options = options |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 376 | |
Johan Alfvén | 5e0ae55 | 2022-02-09 21:20:10 +0100 | [diff] [blame] | 377 | def avoid_nhcwb16_for_ofm(self, tens, ps, arch): |
| 378 | # Only run this check for opt strategy Size |
| 379 | if self.scheduler_options.optimization_strategy == OptimizationStrategy.Performance: |
| 380 | return False |
| 381 | |
| 382 | op = ps.primary_op |
| 383 | if not op.type.is_elementwise_op(): |
| 384 | return False |
| 385 | |
| 386 | depth = op.ofm_shapes[0][-1] |
| 387 | if (depth % 16) == 0: |
| 388 | return False |
| 389 | |
| 390 | # Check if overwriting the inputs can be allowed |
| 391 | OpShapeTens = namedtuple("OpShapeTens", ["op_shape", "tens"]) |
| 392 | outp = OpShapeTens(op.ofm_shapes[0], op.ofm) |
| 393 | inps = [] |
| 394 | if op.ifm is not None: |
| 395 | inps.append(OpShapeTens(op.ifm_shapes[0], op.ifm)) |
| 396 | if op.ifm2 is not None: |
| 397 | inps.append(OpShapeTens(op.ifm_shapes[1], op.ifm2)) |
| 398 | |
| 399 | # Find an input tensor that can be overwritten by the output |
| 400 | for inp in inps: |
| 401 | if ( |
| 402 | # check op input and output shapes allow overlapping |
| 403 | inp.op_shape == outp.op_shape |
| 404 | # check input tensor is valid |
| 405 | and inp.tens is not None |
| 406 | and inp.tens.shape != [] |
| 407 | # check input and output tensors are compatible |
| 408 | and inp.tens.format == outp.tens.format |
| 409 | and inp.tens.dtype == outp.tens.dtype |
| 410 | ): |
| 411 | if inp.tens.format == TensorFormat.NHWC: |
| 412 | return True |
| 413 | |
| 414 | return False |
| 415 | |
Tim Hall | d8339a7 | 2021-05-27 18:49:40 +0100 | [diff] [blame] | 416 | def create_scheduler_representation(self, arch: ArchitectureFeatures): |
| 417 | """Creates a Scheduler Graph representation""" |
| 418 | # Temporary dict for creating connections between the Operations |
| 419 | connections: Dict[Tensor, Connection] = {} |
| 420 | # Memory required for the largest FeatureMap that has to be full |
| 421 | min_memory_req = 0 |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 422 | for ps in self.sg.passes: |
Tim Hall | d8339a7 | 2021-05-27 18:49:40 +0100 | [diff] [blame] | 423 | if ps.primary_op: |
| 424 | # Set tensor format to NHCWB16 for output FeatureMaps, if possible |
Louis Verhaard | 0b9c9a3 | 2020-09-15 14:05:38 +0200 | [diff] [blame] | 425 | for output in ps.outputs: |
Jacob Bohlin | a5e8c1c | 2021-06-14 13:33:39 +0200 | [diff] [blame] | 426 | if output in self.sg.output_tensors or output.purpose != TensorPurpose.FeatureMap: |
Patrik Gustavsson | feeb06d | 2020-04-22 12:53:47 +0200 | [diff] [blame] | 427 | continue |
Johan Alfvén | 5e0ae55 | 2022-02-09 21:20:10 +0100 | [diff] [blame] | 428 | |
| 429 | if output.needs_linear_format: |
| 430 | continue |
| 431 | |
| 432 | if self.avoid_nhcwb16_for_ofm(output, ps, arch): |
| 433 | output.needs_linear_format = True |
| 434 | continue |
| 435 | |
| 436 | output.set_format(TensorFormat.NHCWB16, arch) |
Tim Hall | d8339a7 | 2021-05-27 18:49:40 +0100 | [diff] [blame] | 437 | |
| 438 | # Create SchedulerOperations |
| 439 | op = SchedulerOperation(ps, arch, self.nng) |
| 440 | op.index = len(self.sched_ops) |
| 441 | |
| 442 | # Make connections |
| 443 | if ps.ifm_tensor not in connections: |
| 444 | connections[ps.ifm_tensor] = Connection(ps.ifm_tensor) |
| 445 | if ps.ifm2_tensor and ps.ifm2_tensor not in connections: |
| 446 | connections[ps.ifm2_tensor] = Connection(ps.ifm2_tensor) |
| 447 | if ps.ofm_tensor not in connections: |
| 448 | connections[ps.ofm_tensor] = Connection(ps.ofm_tensor) |
| 449 | |
| 450 | op.add_ifm_connection(connections[ps.ifm_tensor]) |
| 451 | if ps.ifm2_tensor: |
| 452 | op.add_ifm2_connection(connections[ps.ifm2_tensor]) |
| 453 | op.add_ofm_connection(connections[ps.ofm_tensor]) |
| 454 | |
| 455 | # Set requirements on the ifm/ofm buffers |
| 456 | self.sched_ops.append(op) |
| 457 | if ps.ifm_tensor in self.sg.input_tensors: |
| 458 | # This Op consumes a subgraph input |
| 459 | op.requires_full_ifm = True |
| 460 | if ps.ifm2_tensor and ps.ifm2_tensor in self.sg.input_tensors: |
| 461 | # This Op consumes a subgraph input |
| 462 | op.requires_full_ifm2 = True |
| 463 | if ps.ofm_tensor in self.sg.output_tensors: |
| 464 | # This Op produces a subgraph output |
| 465 | op.requires_full_ofm = True |
| 466 | if ps.ifm_tensor.needs_linear_format: |
| 467 | op.requires_full_ifm = True |
| 468 | if ps.ifm2_tensor and ps.ifm2_tensor.needs_linear_format: |
| 469 | op.requires_full_ifm2 = True |
| 470 | if ps.ofm_tensor.needs_linear_format or ps.primary_op.memory_function == Op.ConcatSliceWrite: |
| 471 | op.requires_full_ofm = True |
| 472 | if len(ps.primary_op.outputs) > 1 or len(ps.primary_op.outputs[0].consumer_list) > 1: |
| 473 | # Op has multiple outputs or consumers - requires full OFM |
| 474 | op.requires_full_ofm = True |
| 475 | |
| 476 | # Check memory requirements if this Op requires any full FeatureMaps |
| 477 | op_memory_req = 0 |
| 478 | if op.requires_full_ifm: |
| 479 | op_memory_req += op.ifm_size_in_bytes() |
| 480 | if op.requires_full_ifm2: |
| 481 | op_memory_req += op.ifm2_size_in_bytes() |
| 482 | if op.requires_full_ofm: |
| 483 | op_memory_req += op.ofm_size_in_bytes() |
| 484 | |
| 485 | min_memory_req = max(op_memory_req, min_memory_req) |
| 486 | |
| 487 | # Theoretical minimum required memory - used to guide the cascade building |
| 488 | self.min_memory_req = min_memory_req |
| 489 | |
| 490 | def create_initial_schedule(self) -> Schedule: |
| 491 | """Creates an initial schedule with no cascading or buffering of any kind""" |
| 492 | schedule = Schedule(self.sg, "MAX") |
Tim Hall | d8339a7 | 2021-05-27 18:49:40 +0100 | [diff] [blame] | 493 | for op in self.sched_ops: |
| 494 | cost = op.create_scheduler_info(self.nng, op.ofm.shape) |
| 495 | cost.cycles = self.estimate_op_performance(op, cost.block_config, op.ofm.shape.depth) |
| 496 | schedule.cost_map[op] = cost |
| 497 | |
| 498 | return schedule |
| 499 | |
| 500 | def update_op_memory_snapshot(self, schedule: Schedule): |
| 501 | memories_list = [(self.arch.fast_storage_mem_area, set((MemType.Scratch, MemType.Scratch_fast)))] |
| 502 | |
| 503 | # Collect live ranges from tensors |
| 504 | lr_graph = live_range.LiveRangeGraph() |
| 505 | for mem_area, mem_type_set in memories_list: |
| 506 | live_range.extract_live_ranges_from_cascaded_passes( |
Jonas Ohlsson | d857507 | 2022-03-30 10:30:25 +0200 | [diff] [blame] | 507 | self.nng.get_root_subgraph(), |
| 508 | mem_area, |
| 509 | mem_type_set, |
| 510 | lr_graph, |
| 511 | Tensor.AllocationQuantum, |
Tim Hall | d8339a7 | 2021-05-27 18:49:40 +0100 | [diff] [blame] | 512 | ) |
| 513 | |
| 514 | # Populate time-array with memory used by live ranges |
| 515 | temporal_usage = lr_graph.get_temporal_memory_usage(self.arch.fast_storage_mem_area) |
| 516 | schedule.memory_snapshot = temporal_usage |
| 517 | |
| 518 | # Set the peak memory usage |
| 519 | schedule.fast_storage_peak_usage = max(temporal_usage, default=0) |
| 520 | |
| 521 | def estimate_op_performance(self, op: SchedulerOperation, block_config, ofm_depth): |
| 522 | query = npu_performance.PerformanceQuery(op.op_type.npu_block_type) |
| 523 | query.ifm_shape = op.ifm.shape |
| 524 | query.ifm_memory_area = op.ifm.mem_area |
| 525 | query.ifm_bits = op.ifm.dtype.size_in_bits() |
| 526 | query.ifm_format = op.ifm.format |
| 527 | query.ifm2_shape = op.ifm2 and op.ifm2.shape |
| 528 | query.ifm2_memory_area = op.ifm2 and op.ifm2.mem_area |
| 529 | query.ifm2_bits = op.ifm2 and op.ifm2.dtype.size_in_bits() |
| 530 | query.ifm2_format = op.ifm2 and op.ifm2.format |
| 531 | query.ofm_shape = op.ofm.shape.with_depth(ofm_depth) |
| 532 | query.ofm_memory_area = op.ofm.mem_area |
| 533 | query.ofm_bits = op.ofm.dtype.size_in_bits() |
| 534 | query.ofm_format = op.ofm.format |
| 535 | if op.parent_op.bias: |
| 536 | query.const_shape = Shape4D(1, 1, 1, op.ofm.shape.depth) |
| 537 | query.const_memory_area = self.arch.fast_storage_mem_area |
| 538 | |
| 539 | query.kernel = op.kernel |
| 540 | query.config = block_config |
| 541 | |
| 542 | return npu_performance.measure_cycle_cost(self.arch, op.op_type, op.activation and op.activation.op_type, query) |
| 543 | |
Tim Hall | 789e6f3 | 2021-06-17 17:02:31 +0100 | [diff] [blame] | 544 | def propose_schedule_buffering(self, ref_schedule: Schedule, staging_limit_bytes): |
Tim Hall | d8339a7 | 2021-05-27 18:49:40 +0100 | [diff] [blame] | 545 | """Create a buffered schedule""" |
| 546 | buffered_schedule = Schedule(self.sg, f"{ref_schedule.label}_BUFFERED") |
Tim Hall | d8339a7 | 2021-05-27 18:49:40 +0100 | [diff] [blame] | 547 | |
| 548 | prev_op = None |
| 549 | for sched_op in self.sched_ops: |
| 550 | if sched_op not in ref_schedule.cost_map: |
| 551 | # sched_op is not part of this sub-schedule - skip |
| 552 | continue |
| 553 | |
| 554 | self.propose_operator_buffering(sched_op, prev_op, buffered_schedule, ref_schedule, staging_limit_bytes) |
| 555 | prev_op = sched_op |
| 556 | |
| 557 | return buffered_schedule |
| 558 | |
| 559 | def propose_operator_buffering( |
| 560 | self, |
| 561 | sched_op: SchedulerOperation, |
Jonas Ohlsson | 845e232 | 2022-03-01 12:39:55 +0100 | [diff] [blame] | 562 | prev_op: Optional[SchedulerOperation], |
Tim Hall | d8339a7 | 2021-05-27 18:49:40 +0100 | [diff] [blame] | 563 | buffered_schedule: Schedule, |
| 564 | ref_schedule: Schedule, |
| 565 | staging_limit_bytes, |
| 566 | ): |
| 567 | # Mild recursion might mean this Op has already been seen |
| 568 | if sched_op in buffered_schedule.cost_map: |
| 569 | return |
| 570 | |
| 571 | # Take the reference schedule as default costings for this schedule |
| 572 | ref_cost = ref_schedule.cost_map[sched_op] |
| 573 | cost = copy.copy(ref_cost) |
| 574 | cost.slack_buffering_cycles = ref_cost.cycles.op_cycles |
| 575 | memory_snapshot = ref_schedule.memory_snapshot |
| 576 | ref_memory_usage = memory_snapshot[ref_cost.time_index] if ref_cost.time_index < len(memory_snapshot) else 0 |
| 577 | cost.slack_buffering_memory = staging_limit_bytes - ref_memory_usage |
| 578 | buffered_schedule.cost_map[sched_op] = cost |
| 579 | |
| 580 | # Attempt weight buffering on anything with a weights tensor |
| 581 | if sched_op.parent_op.weights: |
| 582 | self.propose_weight_buffering( |
| 583 | sched_op.parent_op.weights, |
| 584 | sched_op.parent_op.bias, |
| 585 | sched_op, |
| 586 | prev_op, |
| 587 | buffered_schedule, |
| 588 | ref_schedule, |
| 589 | cost.slack_buffering_memory, |
| 590 | ) |
| 591 | |
| 592 | return cost |
| 593 | |
| 594 | def weights_needs_dma(self, weight_tensor): |
| 595 | if weight_tensor and weight_tensor.mem_type not in (MemType.Scratch, MemType.Scratch_fast): |
| 596 | # Weights are in permanent storage |
| 597 | # Only when permanent storage differs from feature map storage, there is a point moving the data |
| 598 | if ( |
| 599 | weight_tensor.mem_area in (MemArea.Dram, MemArea.OffChipFlash) |
| 600 | and self.arch.permanent_storage_mem_area != self.arch.fast_storage_mem_area |
| 601 | ): |
| 602 | return True |
| 603 | return False |
| 604 | |
| 605 | def propose_weight_buffering( |
| 606 | self, |
| 607 | weight_tensor, |
| 608 | scale_tensor, |
| 609 | sched_op: SchedulerOperation, |
| 610 | prev_op: SchedulerOperation, |
| 611 | buffered_schedule: Schedule, |
| 612 | ref_schedule: Schedule, |
| 613 | buffer_limit_bytes, |
| 614 | ): |
| 615 | cost = buffered_schedule.cost_map[sched_op] |
| 616 | prev_cost = buffered_schedule.cost_map.get(prev_op) |
| 617 | ref_cost = ref_schedule.cost_map[sched_op] |
| 618 | assert cost and ref_cost |
| 619 | |
| 620 | needs_dma = self.weights_needs_dma(weight_tensor) |
| 621 | |
| 622 | ofm_full_depth_slices = [0, ref_cost.stripe.depth] |
| 623 | |
| 624 | # Encode weights for the full depth |
Tim Hall | d784af7 | 2021-06-08 21:25:57 +0100 | [diff] [blame] | 625 | full_weights, full_scales = weight_compressor.encode_weight_and_scale_tensor( |
Tim Hall | d8339a7 | 2021-05-27 18:49:40 +0100 | [diff] [blame] | 626 | self.arch, |
| 627 | sched_op.parent_op, |
| 628 | weight_tensor, |
| 629 | scale_tensor, |
| 630 | sched_op.kernel, |
| 631 | cost.block_config, |
| 632 | ofm_full_depth_slices, |
| 633 | ) |
| 634 | full_weights_bytes = len(full_weights.buffer) |
| 635 | cost.ofm_depth_slices = ofm_full_depth_slices |
| 636 | |
| 637 | # No buffering required - take all the weights from permanent storage |
| 638 | if sched_op.op_type == Op.FullyConnected or not needs_dma: |
| 639 | cost.npu_weights_tensor = full_weights |
Tim Hall | d784af7 | 2021-06-08 21:25:57 +0100 | [diff] [blame] | 640 | cost.npu_scales_tensor = full_scales |
Tim Hall | d8339a7 | 2021-05-27 18:49:40 +0100 | [diff] [blame] | 641 | return |
| 642 | |
Jonas Ohlsson | 845e232 | 2022-03-01 12:39:55 +0100 | [diff] [blame] | 643 | encoded_weights: Optional[NpuWeightTensor] = full_weights |
Tim Hall | d784af7 | 2021-06-08 21:25:57 +0100 | [diff] [blame] | 644 | encoded_scales = full_scales |
Tim Hall | d8339a7 | 2021-05-27 18:49:40 +0100 | [diff] [blame] | 645 | |
| 646 | # How many NPU cycles are available under the previously executing |
| 647 | # operator and SRAM unused for performing buffered DMA transfers |
| 648 | slack_cycles = prev_cost.slack_buffering_cycles if prev_cost else 0 |
| 649 | slack_memory = prev_cost.slack_buffering_memory if prev_cost else 0 |
| 650 | |
| 651 | # Force full depth for cascaded Ops |
| 652 | if ref_cost.cascade != 0: |
| 653 | weight_tensor_purpose = TensorSubPurpose.Standard |
| 654 | weight_buffer_size = full_weights_bytes |
| 655 | # Update the memory snapshot to reflect the added size of the weights |
| 656 | ref_schedule.memory_snapshot[ref_cost.time_index] += weight_buffer_size |
| 657 | else: |
| 658 | # Estimate the buffering cycle time for the full set of weights |
| 659 | full_transfer_cycles = npu_performance.measure_mem2mem_cycles( |
| 660 | self.arch, weight_tensor.mem_area, self.arch.fast_storage_mem_area, full_weights_bytes |
| 661 | ) |
| 662 | cost.full_weight_transfer_cycles = full_transfer_cycles |
| 663 | |
| 664 | # Calculate the amount of prebuffering necessary (or what is possible with limited |
| 665 | # double buffer buffer size) |
| 666 | half_buffer_limit = buffer_limit_bytes // 2 |
| 667 | if full_transfer_cycles > slack_cycles: |
| 668 | prebuffer_ratio = slack_cycles / full_transfer_cycles |
| 669 | prebuffer_bytes = min(prebuffer_ratio * full_weights_bytes, half_buffer_limit) |
| 670 | else: |
| 671 | prebuffer_bytes = min(full_weights_bytes, half_buffer_limit) |
Tim Hall | 789e6f3 | 2021-06-17 17:02:31 +0100 | [diff] [blame] | 672 | |
| 673 | prebuffer_ratio = prebuffer_bytes / full_weights_bytes |
Tim Hall | d8339a7 | 2021-05-27 18:49:40 +0100 | [diff] [blame] | 674 | |
| 675 | # Have to split the weights if the initial buffering can't store |
| 676 | # all of the compressed weights |
| 677 | if prebuffer_bytes < full_weights_bytes: |
Tim Hall | 789e6f3 | 2021-06-17 17:02:31 +0100 | [diff] [blame] | 678 | block_depth = cost.block_config.ofm_block.depth |
Tim Hall | d8339a7 | 2021-05-27 18:49:40 +0100 | [diff] [blame] | 679 | |
Tim Hall | 789e6f3 | 2021-06-17 17:02:31 +0100 | [diff] [blame] | 680 | # Choose initial prebuffering depth (already buffer clamped) |
| 681 | prebuffer_depth = ref_cost.stripe.depth * prebuffer_ratio |
Tim Hall | d8339a7 | 2021-05-27 18:49:40 +0100 | [diff] [blame] | 682 | prebuffer_depth = int(max(16, round_down(prebuffer_depth, ArchitectureFeatures.OFMSplitDepth))) |
| 683 | |
Tim Hall | 789e6f3 | 2021-06-17 17:02:31 +0100 | [diff] [blame] | 684 | # Calculate cycles executed during the prebuffer |
| 685 | pre_op_cycles = self.estimate_op_performance(sched_op, cost.block_config, prebuffer_depth) |
| 686 | buffering_depth = ref_cost.stripe.depth * (pre_op_cycles.op_cycles / full_transfer_cycles) |
Tim Hall | d8339a7 | 2021-05-27 18:49:40 +0100 | [diff] [blame] | 687 | |
Tim Hall | 789e6f3 | 2021-06-17 17:02:31 +0100 | [diff] [blame] | 688 | # Choose initial buffering depth and clamp to the double buffering limit |
| 689 | buffering_depth = round_up(buffering_depth, block_depth) |
| 690 | buffering_bytes = (buffering_depth / ref_cost.stripe.depth) * full_weights_bytes |
| 691 | if buffering_bytes > half_buffer_limit: |
| 692 | buffering_depth = (half_buffer_limit / full_weights_bytes) * ref_cost.stripe.depth |
| 693 | |
| 694 | while True: |
| 695 | # Attempt to buffer whole blocks |
Johan Alfvén | cce7f2d | 2022-04-08 10:47:09 +0200 | [diff] [blame] | 696 | if buffering_depth > block_depth: |
Tim Hall | 789e6f3 | 2021-06-17 17:02:31 +0100 | [diff] [blame] | 697 | buffering_depth = round_down(buffering_depth, block_depth) |
| 698 | else: |
| 699 | buffering_depth = round_down(buffering_depth, ArchitectureFeatures.OFMSplitDepth) |
| 700 | buffering_depth = int(max(buffering_depth, ArchitectureFeatures.OFMSplitDepth)) |
Tim Hall | d8339a7 | 2021-05-27 18:49:40 +0100 | [diff] [blame] | 701 | |
| 702 | # Create list of depth slices |
| 703 | depth_slices = [0] |
| 704 | if prebuffer_depth < ref_cost.stripe.depth: |
| 705 | depth_slices += list(range(prebuffer_depth, ref_cost.stripe.depth, buffering_depth)) |
| 706 | depth_slices.append(ref_cost.stripe.depth) |
| 707 | |
| 708 | # Encode weights based depth slices |
| 709 | cost.ofm_depth_slices = depth_slices |
Tim Hall | d784af7 | 2021-06-08 21:25:57 +0100 | [diff] [blame] | 710 | encoded_weights, encoded_scales = weight_compressor.encode_weight_and_scale_tensor( |
Tim Hall | d8339a7 | 2021-05-27 18:49:40 +0100 | [diff] [blame] | 711 | self.arch, |
| 712 | sched_op.parent_op, |
| 713 | weight_tensor, |
| 714 | scale_tensor, |
| 715 | sched_op.kernel, |
| 716 | cost.block_config, |
| 717 | cost.ofm_depth_slices, |
| 718 | ) |
Jonas Ohlsson | 845e232 | 2022-03-01 12:39:55 +0100 | [diff] [blame] | 719 | assert encoded_weights is not None |
Tim Hall | d8339a7 | 2021-05-27 18:49:40 +0100 | [diff] [blame] | 720 | # Chosen buffering might not fit at all, iterate until it does |
| 721 | # or until the minimum usable slice size is reached |
| 722 | if ( |
Tim Hall | b5df773 | 2022-05-04 16:20:43 +0100 | [diff] [blame^] | 723 | encoded_weights.max_range_bytes <= half_buffer_limit |
Tim Hall | d8339a7 | 2021-05-27 18:49:40 +0100 | [diff] [blame] | 724 | or prebuffer_depth == ArchitectureFeatures.OFMSplitDepth |
| 725 | ): |
| 726 | break |
| 727 | |
Tim Hall | 789e6f3 | 2021-06-17 17:02:31 +0100 | [diff] [blame] | 728 | if buffering_depth > prebuffer_depth: |
| 729 | buffering_depth = round_up(buffering_depth // 2, ArchitectureFeatures.OFMSplitDepth) |
| 730 | else: |
| 731 | prebuffer_depth = round_up(prebuffer_depth // 2, ArchitectureFeatures.OFMSplitDepth) |
Tim Hall | d8339a7 | 2021-05-27 18:49:40 +0100 | [diff] [blame] | 732 | |
| 733 | # Calculate cycles required to run the last op for use as future slack |
| 734 | tail_cycles = self.estimate_op_performance( |
| 735 | sched_op, cost.block_config, depth_slices[-1] - depth_slices[-2] |
| 736 | ) |
| 737 | cost.slack_buffering_cycles = tail_cycles.op_cycles |
| 738 | |
| 739 | # Determine whether the weights need to be double buffered |
Tim Hall | b5df773 | 2022-05-04 16:20:43 +0100 | [diff] [blame^] | 740 | weight_buffer_size = min(len(encoded_weights.buffer), encoded_weights.max_range_bytes) |
Tim Hall | d8339a7 | 2021-05-27 18:49:40 +0100 | [diff] [blame] | 741 | |
| 742 | # Only buffer weights if there's still space left for the buffer |
| 743 | if weight_buffer_size <= buffer_limit_bytes: |
| 744 | assert weight_buffer_size % 16 == 0 |
| 745 | # Determine whether to double buffer or single buffer |
Tim Hall | b5df773 | 2022-05-04 16:20:43 +0100 | [diff] [blame^] | 746 | if (weight_buffer_size * 2 <= buffer_limit_bytes) and (weight_buffer_size < len(encoded_weights.buffer)): |
| 747 | weight_buffer_size = weight_buffer_size * 2 |
Tim Hall | d8339a7 | 2021-05-27 18:49:40 +0100 | [diff] [blame] | 748 | weight_tensor_purpose = TensorSubPurpose.DoubleBuffer |
| 749 | else: |
| 750 | weight_tensor_purpose = TensorSubPurpose.Standard |
| 751 | |
Tim Hall | b5df773 | 2022-05-04 16:20:43 +0100 | [diff] [blame^] | 752 | cost.buffered_weight_tensor = self.buffer_tensor( |
| 753 | encoded_weights, weight_tensor_purpose, weight_buffer_size, weight_tensor.name |
| 754 | ) |
Tim Hall | d8339a7 | 2021-05-27 18:49:40 +0100 | [diff] [blame] | 755 | if ref_cost.cascade == 0: |
Tim Hall | b5df773 | 2022-05-04 16:20:43 +0100 | [diff] [blame^] | 756 | # Determine if the lifetime can be extended and pre-buffer weights under the previous operation |
| 757 | cost.buffered_weight_tensor.pre_buffer = weight_buffer_size < slack_memory |
Tim Hall | d8339a7 | 2021-05-27 18:49:40 +0100 | [diff] [blame] | 758 | |
| 759 | cost.slack_buffering_memory -= weight_buffer_size |
| 760 | else: |
| 761 | # Don't slice or buffer - use the whole depth from persistent storage |
| 762 | cost.ofm_depth_slices = ofm_full_depth_slices |
| 763 | encoded_weights = full_weights |
Tim Hall | d784af7 | 2021-06-08 21:25:57 +0100 | [diff] [blame] | 764 | encoded_scales = full_scales |
Tim Hall | d8339a7 | 2021-05-27 18:49:40 +0100 | [diff] [blame] | 765 | |
| 766 | cost.npu_weights_tensor = encoded_weights |
Tim Hall | d784af7 | 2021-06-08 21:25:57 +0100 | [diff] [blame] | 767 | cost.npu_scales_tensor = encoded_scales |
Tim Hall | d8339a7 | 2021-05-27 18:49:40 +0100 | [diff] [blame] | 768 | |
Jacob Bohlin | eee9e5d | 2021-08-17 17:44:45 +0200 | [diff] [blame] | 769 | def buffer_tensor(self, src_tensor: Tensor, sub_purpose: TensorSubPurpose, buffer_size: int, name: str) -> Tensor: |
Tim Hall | b5df773 | 2022-05-04 16:20:43 +0100 | [diff] [blame^] | 770 | buffered_weight_tensor = Tensor([1, 1, 1, buffer_size], DataType.uint8, name + "_buffer") |
Jacob Bohlin | eee9e5d | 2021-08-17 17:44:45 +0200 | [diff] [blame] | 771 | buffered_weight_tensor.src_tensor = src_tensor |
| 772 | buffered_weight_tensor.mem_area = self.arch.fast_storage_mem_area |
| 773 | buffered_weight_tensor.mem_type = MemType.Scratch_fast |
| 774 | buffered_weight_tensor.purpose = TensorPurpose.Weights |
| 775 | buffered_weight_tensor.sub_purpose = sub_purpose |
| 776 | return buffered_weight_tensor |
| 777 | |
Tim Hall | d8339a7 | 2021-05-27 18:49:40 +0100 | [diff] [blame] | 778 | def propose_minimal_schedule(self) -> Schedule: |
| 779 | """Proposes scheduling parameters where every operator is subdivided into the smallest stripe that satisfies the |
| 780 | next operators stride""" |
| 781 | min_schedule = Schedule(self.sg, "MIN") |
| 782 | cost_map = min_schedule.cost_map |
| 783 | |
| 784 | # Keep track of the previous Op - which consumes the current Op's OFM |
Jonas Ohlsson | 845e232 | 2022-03-01 12:39:55 +0100 | [diff] [blame] | 785 | prev_op: Optional[SchedulerOperation] = None |
Tim Hall | d8339a7 | 2021-05-27 18:49:40 +0100 | [diff] [blame] | 786 | for sched_op in reversed(self.sched_ops): |
| 787 | min_stripe_height = prev_op.kernel.stride.y if prev_op else 1 |
| 788 | min_stripe = sched_op.ofm.shape.with_height(min_stripe_height) |
| 789 | |
| 790 | cost = sched_op.create_scheduler_info(self.nng, min_stripe) |
| 791 | cost.cycles = self.estimate_op_performance(sched_op, cost.block_config, sched_op.ofm.shape.depth) |
| 792 | cost_map[sched_op] = cost |
| 793 | |
| 794 | prev_op = sched_op |
| 795 | |
| 796 | return min_schedule |
| 797 | |
| 798 | def propose_schedule_striping(self, final_stripe: Shape4D, label: str, ref_schedule: Schedule) -> Schedule: |
| 799 | """Proposes new striping for a schedule. The stripe is derived from the ifm requirements of the next Op down""" |
| 800 | ref_cost = ref_schedule.cost_map |
| 801 | |
| 802 | striped_schedule = Schedule(self.sg, label) |
| 803 | stripe = final_stripe |
| 804 | for sched_op in reversed(self.sched_ops): |
| 805 | if sched_op not in ref_cost: |
| 806 | # sched_op is not part of the sub-schedule - skip |
| 807 | continue |
| 808 | |
| 809 | # Create a cost entry with the new stripe |
| 810 | cost = sched_op.create_scheduler_info(self.nng, stripe) |
| 811 | |
Tim Hall | b5df773 | 2022-05-04 16:20:43 +0100 | [diff] [blame^] | 812 | if ref_cost[sched_op].buffered_weight_tensor: |
Jacob Bohlin | eee9e5d | 2021-08-17 17:44:45 +0200 | [diff] [blame] | 813 | # If the weights are buffered in the reference schedule they should be in the new proposal |
| 814 | weight_tensor = cost.npu_weights_tensor |
Tim Hall | b5df773 | 2022-05-04 16:20:43 +0100 | [diff] [blame^] | 815 | cost.buffered_weight_tensor = self.buffer_tensor( |
| 816 | weight_tensor, TensorSubPurpose.Standard, len(weight_tensor.buffer), weight_tensor.name |
Jacob Bohlin | eee9e5d | 2021-08-17 17:44:45 +0200 | [diff] [blame] | 817 | ) |
Tim Hall | d8339a7 | 2021-05-27 18:49:40 +0100 | [diff] [blame] | 818 | |
| 819 | # Estimate performance |
| 820 | cost.cycles = self.estimate_op_performance(sched_op, cost.block_config, sched_op.ofm.shape.depth) |
| 821 | striped_schedule.cost_map[sched_op] = cost |
| 822 | |
| 823 | # Calculate the preceeding Op's stripe |
| 824 | stripe = sched_op.ifm.shape.with_height(stripe.height * sched_op.kernel.stride.y) |
| 825 | |
| 826 | return striped_schedule |
| 827 | |
| 828 | def estimate_schedule_memory_usage(self, schedule: Schedule, non_local_mem_usage: dict): |
| 829 | """Estimates the memory usage of a schedule""" |
| 830 | cost = schedule.cost_map |
| 831 | cascades = schedule.cascades |
| 832 | peak_mem_usage = 0 |
| 833 | for sched_op in self.sched_ops: |
| 834 | if sched_op not in cost: |
| 835 | # sched_op is not part of the sub-schedule - skip |
| 836 | continue |
| 837 | |
| 838 | if cost[sched_op].cascade: |
| 839 | # This Op is part of a cascade - use the cascade's memory usage |
| 840 | cascade_info = cascades[cost[sched_op].cascade] |
| 841 | # Non-local memory usage is already included in the cascade_info |
| 842 | peak_mem_usage = max(cascade_info.mem_usage, peak_mem_usage) |
| 843 | else: |
| 844 | # This Op is not part of a cascade - calculate the memory usage |
Tim Hall | b5df773 | 2022-05-04 16:20:43 +0100 | [diff] [blame^] | 845 | op_weight_buffer = 0 |
| 846 | if cost[sched_op].buffered_weight_tensor: |
| 847 | op_weight_buffer = cost[sched_op].buffered_weight_tensor.storage_size() |
Tim Hall | d8339a7 | 2021-05-27 18:49:40 +0100 | [diff] [blame] | 848 | |
| 849 | op_mem_usage = ( |
| 850 | sched_op.ifm_size_in_bytes() |
| 851 | + sched_op.ofm_size_in_bytes() |
| 852 | + op_weight_buffer |
| 853 | + non_local_mem_usage.get(sched_op, 0) |
| 854 | ) |
| 855 | peak_mem_usage = max(op_mem_usage, peak_mem_usage) |
| 856 | |
| 857 | return peak_mem_usage |
| 858 | |
| 859 | def optimize_sub_schedule( |
| 860 | self, cascade_info: CascadeInfo, ref_schedule: Schedule, max_template: Schedule, memory_limit: int |
| 861 | ) -> Schedule: |
| 862 | """Extracts the Ops covered by the given cascade and creates a sub-schedule. The sub-schedule is optimized by |
| 863 | proposing weight buffering and then continously proposing new stripe sizes""" |
| 864 | ref_cost = ref_schedule.cost_map |
| 865 | # Extract the ops that are part of this sub-schedule |
| 866 | start = cascade_info.start |
| 867 | end = cascade_info.end |
| 868 | sub_schedule_ops = self.sched_ops[start : end + 1] |
| 869 | # Create a sub-schedule that contains only the costs for the Ops that are part of the sub-schedule |
| 870 | sub_schedule = Schedule(self.sg, f"SUB_{start}_{end}") |
| 871 | for sched_op in sub_schedule_ops: |
| 872 | sub_schedule.cost_map[sched_op] = ref_cost[sched_op] |
| 873 | |
| 874 | sub_schedule.cascades[end] = cascade_info |
| 875 | # Use the memory snapshot from the reference schedule |
| 876 | sub_schedule.memory_snapshot = ref_schedule.memory_snapshot |
| 877 | |
| 878 | # Calculate memory usage that is live during the sub-schedule but not part of it |
| 879 | time_for_cascade = ref_cost[sub_schedule_ops[0]].time_index |
| 880 | mem_usage_parallel_to_sub_schedule = ref_schedule.memory_snapshot[time_for_cascade] - cascade_info.mem_usage |
| 881 | # If the first Op's IFM has other consumers it has to live throughout the whole sub-schedule whether it's |
| 882 | # included in a cascade or not |
| 883 | persistent_initial_ifm = ( |
| 884 | sub_schedule_ops[0].ifm_size_in_bytes() if len(sub_schedule_ops[0].ifm.connection.consumers) > 1 else 0 |
| 885 | ) |
| 886 | # Calculate non-local-mem-usage per Operator |
| 887 | non_local_mem_usage = {} |
| 888 | for idx, sched_op in enumerate(sub_schedule_ops): |
| 889 | non_local_mem_usage[sched_op] = mem_usage_parallel_to_sub_schedule |
| 890 | if idx != 0: |
| 891 | non_local_mem_usage[sched_op] += persistent_initial_ifm |
| 892 | |
| 893 | cascade_builder = CascadeBuilder(sub_schedule_ops, self.arch.is_spilling_enabled(), non_local_mem_usage) |
| 894 | |
| 895 | # Start by adding buffering |
Tim Hall | 789e6f3 | 2021-06-17 17:02:31 +0100 | [diff] [blame] | 896 | buffered_sub_schedule = self.propose_schedule_buffering( |
| 897 | sub_schedule, self.scheduler_options.optimization_sram_limit |
| 898 | ) |
Tim Hall | d8339a7 | 2021-05-27 18:49:40 +0100 | [diff] [blame] | 899 | # Copy the cascades over from the unbuffered-schedule |
| 900 | buffered_sub_schedule.cascades = sub_schedule.cascades |
| 901 | |
| 902 | # Generate the possible stripings for the final Op in the sub-schedule |
| 903 | final_ofm_shape = sub_schedule_ops[-1].ofm.shape |
| 904 | possible_stripes = [ |
| 905 | final_ofm_shape.with_height(stripe_h) for stripe_h in range(1, final_ofm_shape.height // 2 + 1) |
| 906 | ] |
| 907 | |
| 908 | # Propose different striping - the possible stripes are proposed similarly to a binary search |
Jacob Bohlin | fad7204 | 2021-08-24 21:51:41 +0200 | [diff] [blame] | 909 | best_schedule = None |
Tim Hall | d8339a7 | 2021-05-27 18:49:40 +0100 | [diff] [blame] | 910 | iteration = 0 |
| 911 | while len(possible_stripes) > 1: |
| 912 | proposed_stripe = possible_stripes[len(possible_stripes) // 2] |
| 913 | proposed_schedule = self.propose_schedule_striping( |
| 914 | proposed_stripe, f"OPTIMIZED_{iteration}", buffered_sub_schedule |
| 915 | ) |
| 916 | |
| 917 | cascade_builder.build_cascades(proposed_schedule, max_template, memory_limit) |
| 918 | |
| 919 | # Check if proposal fits |
| 920 | proposed_schedule_mem_usage = self.estimate_schedule_memory_usage(proposed_schedule, non_local_mem_usage) |
| 921 | if (proposed_schedule_mem_usage) <= memory_limit: |
| 922 | # Remove all possible stripes smaller than this |
| 923 | possible_stripes = possible_stripes[len(possible_stripes) // 2 :] |
| 924 | best_schedule = proposed_schedule |
| 925 | if not proposed_schedule.cascades: |
| 926 | # No cascading required - early exit |
| 927 | break |
| 928 | else: |
| 929 | # Proposal doesn't fit within the limit - remove all possible stripes larger than this |
| 930 | possible_stripes = possible_stripes[: len(possible_stripes) // 2] |
| 931 | |
| 932 | iteration += 1 |
| 933 | |
| 934 | return best_schedule |
| 935 | |
| 936 | def optimize_schedule( |
Jonas Ohlsson | d857507 | 2022-03-30 10:30:25 +0200 | [diff] [blame] | 937 | self, |
| 938 | schedule: Schedule, |
| 939 | max_sched: Schedule, |
| 940 | max_template: Schedule, |
| 941 | options: SchedulerOptions, |
Tim Hall | d8339a7 | 2021-05-27 18:49:40 +0100 | [diff] [blame] | 942 | ) -> Schedule: |
| 943 | """Extracts sub-schedules based on the cascades and optimizes them and applies them to the final schedule""" |
| 944 | sram_limit = options.optimization_sram_limit |
| 945 | if max_sched.fast_storage_peak_usage < sram_limit and not self.arch.is_spilling_enabled(): |
| 946 | # Maximum performance schedule fits within the SRAM target |
| 947 | return max_sched |
| 948 | |
Jacob Bohlin | fad7204 | 2021-08-24 21:51:41 +0200 | [diff] [blame] | 949 | # Iterate over a copy of the cascades since they may change during the loop |
| 950 | for cascade_info in list(schedule.cascades.values()): |
Tim Hall | d8339a7 | 2021-05-27 18:49:40 +0100 | [diff] [blame] | 951 | # Optimize the sub-schedule in this cascade |
| 952 | opt_sub_schedule = self.optimize_sub_schedule(cascade_info, schedule, max_template, sram_limit) |
Jacob Bohlin | fad7204 | 2021-08-24 21:51:41 +0200 | [diff] [blame] | 953 | if opt_sub_schedule: |
| 954 | # Remove the existing cascade |
| 955 | del schedule.cascades[cascade_info.end] |
| 956 | # Update the sub-schedule Op and cascade costs to the full schedule |
| 957 | schedule.cost_map.update(opt_sub_schedule.cost_map) |
| 958 | schedule.cascades.update(opt_sub_schedule.cascades) |
Tim Hall | d8339a7 | 2021-05-27 18:49:40 +0100 | [diff] [blame] | 959 | |
| 960 | # Update memory snapshot |
| 961 | self.sg.schedule = schedule |
| 962 | self.update_op_memory_snapshot(schedule) |
| 963 | # Propose schedule buffering to the optimized schedule |
Tim Hall | 789e6f3 | 2021-06-17 17:02:31 +0100 | [diff] [blame] | 964 | optimized_sched = self.propose_schedule_buffering(schedule, self.scheduler_options.optimization_sram_limit) |
Tim Hall | d8339a7 | 2021-05-27 18:49:40 +0100 | [diff] [blame] | 965 | # Copy the cascade's metadata from the unbuffered schedule |
| 966 | optimized_sched.cascades = schedule.cascades |
| 967 | return optimized_sched |
| 968 | |
| 969 | def apply_schedule(self, sched: Schedule): |
| 970 | """Applies the given schedule as a final solution""" |
| 971 | for sched_op in self.sched_ops: |
| 972 | op_info = sched.cost_map[sched_op] |
| 973 | cascade_info = sched.cascades.get(op_info.cascade, None) |
| 974 | if cascade_info and sched_op in cascade_info.buffers: |
| 975 | buffer_tens = sched_op.ifm.connection.parent_tens |
| 976 | # Apply memory area and type |
| 977 | buffer_tens.mem_area = self.arch.fast_storage_mem_area |
| 978 | buffer_tens.mem_type = MemType.Scratch_fast |
| 979 | # Apply Rolling buffer |
| 980 | buffer_tens.set_format(TensorFormat.NHCWB16, self.arch) |
| 981 | buffer_tens.set_new_sub_purpose(TensorSubPurpose.RollingBufferY, cascade_info.buffers[sched_op].height) |
| 982 | |
| 983 | sched_op.parent_ps.block_config = op_info.block_config.old_style_representation() |
| 984 | |
| 985 | # Ensure that the src_tensor reference is set correctly |
Tim Hall | b5df773 | 2022-05-04 16:20:43 +0100 | [diff] [blame^] | 986 | if op_info.buffered_weight_tensor: |
| 987 | op_info.buffered_weight_tensor.src_tensor = op_info.npu_weights_tensor |
Tim Hall | d8339a7 | 2021-05-27 18:49:40 +0100 | [diff] [blame] | 988 | |
erik.andersson@arm.com | de6cb64 | 2022-02-02 14:03:15 +0100 | [diff] [blame] | 989 | def use_fast_storage_for_feature_maps(self, schedule, staging_limit): |
| 990 | scratched_fms = {} |
| 991 | max_mem_usage = [] |
| 992 | base_mem_usage = [] |
| 993 | fast_storage_type = MemType.Scratch_fast |
| 994 | fast_storage_mem_area = self.arch.fast_storage_mem_area |
Tim Hall | d8339a7 | 2021-05-27 18:49:40 +0100 | [diff] [blame] | 995 | |
| 996 | # Force all OFMs to fast-storage |
| 997 | for sched_op in self.sched_ops: |
| 998 | cost = schedule.cost_map[sched_op] |
erik.andersson@arm.com | de6cb64 | 2022-02-02 14:03:15 +0100 | [diff] [blame] | 999 | if cost.cascade == 0 and sched_op.get_dependants(): |
| 1000 | ofm_tens = sched_op.ofm.connection.parent_tens |
| 1001 | if not any(cons is None for cons in ofm_tens.consumer_list): |
| 1002 | if ofm_tens not in scratched_fms: |
| 1003 | scratched_fms[ofm_tens] = (ofm_tens.mem_area, ofm_tens.mem_type) |
| 1004 | ofm_tens.mem_area = fast_storage_mem_area |
| 1005 | ofm_tens.mem_type = fast_storage_type |
Tim Hall | d8339a7 | 2021-05-27 18:49:40 +0100 | [diff] [blame] | 1006 | |
| 1007 | # Collect live ranges from tensors |
erik.andersson@arm.com | de6cb64 | 2022-02-02 14:03:15 +0100 | [diff] [blame] | 1008 | memories_list = [(fast_storage_mem_area, set((MemType.Scratch, MemType.Scratch_fast)))] |
Tim Hall | d8339a7 | 2021-05-27 18:49:40 +0100 | [diff] [blame] | 1009 | lr_graph = live_range.LiveRangeGraph() |
| 1010 | for mem_area, mem_type_set in memories_list: |
| 1011 | live_range.extract_live_ranges_from_cascaded_passes( |
Jonas Ohlsson | d857507 | 2022-03-30 10:30:25 +0200 | [diff] [blame] | 1012 | self.nng.get_root_subgraph(), |
| 1013 | mem_area, |
| 1014 | mem_type_set, |
| 1015 | lr_graph, |
| 1016 | Tensor.AllocationQuantum, |
Tim Hall | d8339a7 | 2021-05-27 18:49:40 +0100 | [diff] [blame] | 1017 | ) |
erik.andersson@arm.com | de6cb64 | 2022-02-02 14:03:15 +0100 | [diff] [blame] | 1018 | max_mem_usage = lr_graph.get_temporal_memory_usage(fast_storage_mem_area) |
Tim Hall | d8339a7 | 2021-05-27 18:49:40 +0100 | [diff] [blame] | 1019 | |
erik.andersson@arm.com | de6cb64 | 2022-02-02 14:03:15 +0100 | [diff] [blame] | 1020 | # If true, everything fits and we can proceed |
| 1021 | if max(max_mem_usage) <= staging_limit: |
| 1022 | return |
| 1023 | |
| 1024 | # Build up the base memory usage by removing the |
| 1025 | # mem_usage of the lrs we previously moved to fast-storage |
| 1026 | base_mem_usage = np.array(max_mem_usage) |
| 1027 | curr_lrs = [] |
Tim Hall | d8339a7 | 2021-05-27 18:49:40 +0100 | [diff] [blame] | 1028 | for lr in lr_graph.lrs: |
erik.andersson@arm.com | de6cb64 | 2022-02-02 14:03:15 +0100 | [diff] [blame] | 1029 | for tens in lr.tensors: |
| 1030 | if scratched_fms.get(tens): |
| 1031 | curr_lrs.append(lr) |
| 1032 | base_mem_usage[lr.start_time : lr.end_time + 1] -= lr.size |
| 1033 | break |
| 1034 | |
| 1035 | competing_lrs = [] |
| 1036 | for lr in curr_lrs: |
| 1037 | base_usage = max(base_mem_usage[lr.start_time : lr.end_time + 1]) |
| 1038 | # If true, the lr will never fit and may thus be evicted |
| 1039 | if base_usage + lr.size > staging_limit: |
| 1040 | FastStorageComponentAllocator.evict(lr, max_mem_usage, scratched_fms) |
| 1041 | continue |
| 1042 | # Since max_mem_usage is the memory usage with all FMs still in fast-storage, |
| 1043 | # the memory limit cannot be exceeded if max_mem_usage does not. |
| 1044 | # Thus, the affected lrs can remain in fast-storage if the following is true |
| 1045 | if max(max_mem_usage[lr.start_time : lr.end_time + 1]) <= staging_limit: |
| 1046 | FastStorageComponentAllocator.keep(lr, base_mem_usage, staging_limit) |
| 1047 | else: |
| 1048 | competing_lrs.append(lr) |
| 1049 | sz = len(competing_lrs) |
| 1050 | # All lrs and their tensors have been handled if sz is zero, we may thus return |
| 1051 | if sz == 0: |
| 1052 | return |
| 1053 | |
| 1054 | competing_lrs = sorted(competing_lrs, key=lambda lr: (lr.start_time, lr.end_time + 1, lr.size)) |
| 1055 | start = 0 |
| 1056 | start_time = competing_lrs[0].start_time |
| 1057 | end_time = competing_lrs[0].end_time |
| 1058 | component_allocator = FastStorageComponentAllocator(base_mem_usage, max_mem_usage, staging_limit) |
| 1059 | # Build up components and then allocate each separately |
| 1060 | for i, lr in enumerate(competing_lrs): |
| 1061 | if lr.start_time <= end_time and i - start < component_allocator.max_exhaustive_size: |
| 1062 | start_time = min(start_time, lr.start_time) |
| 1063 | end_time = max(end_time, lr.end_time) |
| 1064 | else: |
| 1065 | component_allocator.allocate_component( |
| 1066 | component_allocator, |
| 1067 | competing_lrs[start:i], |
| 1068 | max_mem_usage, |
| 1069 | base_mem_usage, |
| 1070 | staging_limit, |
| 1071 | scratched_fms, |
| 1072 | ) |
| 1073 | start = i |
| 1074 | start_time = lr.start_time |
| 1075 | end_time = lr.end_time |
| 1076 | component_allocator.allocate_component( |
| 1077 | component_allocator, competing_lrs[start:sz], max_mem_usage, base_mem_usage, staging_limit, scratched_fms |
| 1078 | ) |
Tim Hall | d8339a7 | 2021-05-27 18:49:40 +0100 | [diff] [blame] | 1079 | |
| 1080 | def move_constant_data(self): |
| 1081 | """Determine if data, can be moved from permanent storage to another memory area. A move |
| 1082 | will generate a DMA command in the high-level command stream""" |
| 1083 | for sched_op in self.sched_ops: |
| 1084 | parent_op = sched_op.parent_op |
| 1085 | is_lut_used = any(inp.purpose == TensorPurpose.LUT for inp in parent_op.inputs) |
| 1086 | max_ifm_shram_avail = ( |
| 1087 | (self.arch.available_shram_banks(is_lut_used) - self.arch.shram_reserved_output_banks) |
| 1088 | * self.arch.shram_bank_size |
| 1089 | // 2 |
| 1090 | ) |
| 1091 | |
| 1092 | for idx, tens in enumerate(parent_op.inputs): |
| 1093 | if tens.mem_type not in (MemType.Scratch, MemType.Scratch_fast): |
| 1094 | # Tensor is in permanent storage |
| 1095 | # Only when permanent storage differs from feature map storage, there is a point moving the data |
| 1096 | if ( |
| 1097 | tens.mem_area in self.arch.permanent_storage_mem_area |
| 1098 | and self.arch.permanent_storage_mem_area != self.arch.feature_map_storage_mem_area |
| 1099 | ) or tens.purpose == TensorPurpose.LUT: |
| 1100 | if tens.purpose == TensorPurpose.LUT or ( |
Patrik Gustavsson | 94292fe | 2021-09-02 08:22:58 +0200 | [diff] [blame] | 1101 | # For elementwise broadcast |
Tim Hall | d8339a7 | 2021-05-27 18:49:40 +0100 | [diff] [blame] | 1102 | tens.purpose == TensorPurpose.FeatureMap |
| 1103 | and sched_op.op_type.is_binary_elementwise_op() |
| 1104 | and tens.shape != [] |
| 1105 | and sched_op.ifm.shape != sched_op.ofm.shape |
Patrik Gustavsson | 94292fe | 2021-09-02 08:22:58 +0200 | [diff] [blame] | 1106 | and parent_op.write_shape is None |
Tim Hall | d8339a7 | 2021-05-27 18:49:40 +0100 | [diff] [blame] | 1107 | and tens.storage_size() > max_ifm_shram_avail |
| 1108 | ): |
| 1109 | only_vector_product_consumers = all( |
| 1110 | oper and oper.type.npu_block_type == NpuBlockType.VectorProduct |
| 1111 | for oper in tens.consumers() |
| 1112 | ) |
| 1113 | |
| 1114 | if (not only_vector_product_consumers) or tens.purpose == TensorPurpose.LUT: |
| 1115 | new_tens = tens.clone_into_fast_storage(self.arch) |
| 1116 | if tens.purpose == TensorPurpose.LUT: |
| 1117 | new_tens.mem_area = MemArea.Shram |
| 1118 | |
| 1119 | new_tens.consumer_list.append(parent_op) |
| 1120 | parent_op.inputs[idx] = new_tens |
Dwight Lidman | 352607c | 2021-09-29 17:00:09 +0200 | [diff] [blame] | 1121 | # If the index is out of range, IFM and IFM2 are the same tensor |
| 1122 | # and pass inputs don't have duplicates |
| 1123 | if idx < len(sched_op.parent_ps.inputs): |
| 1124 | sched_op.parent_ps.inputs[idx] = new_tens |
Tim Hall | d8339a7 | 2021-05-27 18:49:40 +0100 | [diff] [blame] | 1125 | |
| 1126 | def print_schedule(self, schedule: Schedule): |
| 1127 | print(f"Schedule: '{schedule.name}'") |
| 1128 | for sched_op in self.sched_ops: |
| 1129 | if sched_op not in schedule.cost_map: |
| 1130 | # Sub-schedule printing |
| 1131 | continue |
| 1132 | |
| 1133 | op_info = schedule.cost_map[sched_op] |
| 1134 | print(f"\t{sched_op.index}: Operation {sched_op.name} - OFM {sched_op.ofm.shape}") |
| 1135 | print(f"\t\tType: {sched_op.op_type}") |
| 1136 | print(f"\t\tKernel: {sched_op.kernel}") |
| 1137 | print(f"{op_info}") |
| 1138 | mem_usage = ( |
| 1139 | schedule.memory_snapshot[op_info.time_index] |
| 1140 | if op_info.time_index < len(schedule.memory_snapshot) |
| 1141 | else 0 |
| 1142 | ) |
| 1143 | print(f"\t\tSRAM Used: {mem_usage} bytes") |
| 1144 | |
Jonas Ohlsson | 25e700c | 2022-03-04 14:58:56 +0100 | [diff] [blame] | 1145 | print("\tCascades:") |
Tim Hall | d8339a7 | 2021-05-27 18:49:40 +0100 | [diff] [blame] | 1146 | for i, cascade in enumerate(schedule.cascades.values()): |
| 1147 | print(f"\t\t{i}: {cascade.start} -> {cascade.end}, size: {cascade.mem_usage}") |
Patrik Gustavsson | feeb06d | 2020-04-22 12:53:47 +0200 | [diff] [blame] | 1148 | |
Andreas Nevalainen | 27d36f0 | 2020-11-19 11:27:50 +0100 | [diff] [blame] | 1149 | |
Tim Hall | d8339a7 | 2021-05-27 18:49:40 +0100 | [diff] [blame] | 1150 | def _update_tensor_allocation(nng: Graph, arch: ArchitectureFeatures, options): |
| 1151 | """ |
| 1152 | Creates live ranges and runs tensor allocator for the current schedule |
| 1153 | (i.e. sg.schedule for all subgraphs), returns the maximum memory usage |
| 1154 | and updates SchedulerOpInfo.mem_usage for all operations in the schedule. |
| 1155 | """ |
| 1156 | root_sg = nng.get_root_subgraph() |
| 1157 | |
| 1158 | alloc_list = [] |
| 1159 | if arch.is_spilling_enabled(): |
| 1160 | mem_alloc_scratch_fast = (arch.fast_storage_mem_area, set((MemType.Scratch_fast,))) |
| 1161 | mem_alloc_scratch = (arch.feature_map_storage_mem_area, set((MemType.Scratch,))) |
| 1162 | # Order is important |
| 1163 | alloc_list.append(mem_alloc_scratch_fast) |
| 1164 | alloc_list.append(mem_alloc_scratch) |
| 1165 | else: |
| 1166 | mem_alloc_scratch = (arch.feature_map_storage_mem_area, set((MemType.Scratch, MemType.Scratch_fast))) |
| 1167 | alloc_list.append(mem_alloc_scratch) |
| 1168 | |
| 1169 | for mem_area, mem_type_set in alloc_list: |
| 1170 | tensor_allocation.allocate_tensors( |
| 1171 | nng, |
| 1172 | root_sg, |
| 1173 | arch, |
| 1174 | mem_area, |
| 1175 | mem_type_set, |
| 1176 | tensor_allocator=options.tensor_allocator, |
| 1177 | verbose_allocation=options.verbose_allocation, |
| 1178 | cpu_tensor_alignment=options.cpu_tensor_alignment, |
| 1179 | ) |
| 1180 | |
| 1181 | |
erik.andersson@arm.com | de6cb64 | 2022-02-02 14:03:15 +0100 | [diff] [blame] | 1182 | class FastStorageComponentAllocator: |
| 1183 | def __init__(self, base_mem_usage, max_mem_usage, staging_limit): |
| 1184 | self.base_mem_usage = base_mem_usage |
| 1185 | self.max_mem_usage = list(max_mem_usage) |
| 1186 | self.staging_limit = staging_limit |
| 1187 | self.lrs = [] |
| 1188 | self.evicted = [] |
| 1189 | self.curr_evicted = [] |
| 1190 | self.remaining_total_size = [] |
| 1191 | self.best_allocated_size = 0 |
| 1192 | self.max_exhaustive_size = 20 |
| 1193 | |
| 1194 | def allocate_exhaustive(self, ix, alloc_size): |
| 1195 | if ix >= len(self.lrs): |
| 1196 | if alloc_size > self.best_allocated_size: |
| 1197 | self.best_allocated_size = alloc_size |
Louis Verhaard | 5c8f1e5 | 2022-02-23 14:13:07 +0100 | [diff] [blame] | 1198 | self.evicted = self.curr_evicted.copy() |
erik.andersson@arm.com | de6cb64 | 2022-02-02 14:03:15 +0100 | [diff] [blame] | 1199 | return |
| 1200 | |
| 1201 | lr = self.lrs[ix] |
| 1202 | for t in range(lr.start_time, lr.end_time): |
| 1203 | assert self.base_mem_usage[t] <= self.max_mem_usage[t] |
| 1204 | base_usage = max(self.base_mem_usage[lr.start_time : lr.end_time + 1]) |
| 1205 | can_fit = base_usage + lr.size <= self.staging_limit |
| 1206 | always_fits = can_fit |
| 1207 | |
| 1208 | if can_fit: |
| 1209 | max_usage = max(self.max_mem_usage[lr.start_time : lr.end_time + 1]) |
| 1210 | always_fits = max_usage <= self.staging_limit |
| 1211 | |
| 1212 | if can_fit or always_fits: |
| 1213 | self.curr_evicted[ix] = False |
| 1214 | self.base_mem_usage = self.update_mem_usage(self.base_mem_usage, lr, True) |
| 1215 | self.allocate_exhaustive(ix + 1, alloc_size + lr.size) |
| 1216 | self.base_mem_usage = self.update_mem_usage(self.base_mem_usage, lr, False) |
| 1217 | |
| 1218 | if not always_fits: |
| 1219 | self.curr_evicted[ix] = True |
| 1220 | self.max_mem_usage = self.update_mem_usage(self.max_mem_usage, lr, False) |
| 1221 | self.allocate_exhaustive(ix + 1, alloc_size) |
| 1222 | self.max_mem_usage = self.update_mem_usage(self.max_mem_usage, lr, True) |
| 1223 | |
| 1224 | @staticmethod |
| 1225 | def update_mem_usage(mem_usage, lr, increase): |
| 1226 | for t in range(lr.start_time, lr.end_time + 1): |
| 1227 | mem_usage[t] += lr.size if increase else -lr.size |
| 1228 | assert mem_usage[t] >= 0 |
| 1229 | return mem_usage |
| 1230 | |
| 1231 | @staticmethod |
| 1232 | def evict(lr, max_mem_usage, scratched_fms): |
| 1233 | for t in range(lr.start_time, lr.end_time + 1): |
| 1234 | max_mem_usage[t] -= lr.size |
| 1235 | for tens in lr.tensors: |
| 1236 | if tens in scratched_fms: |
| 1237 | tens.mem_area = scratched_fms[tens][0] |
| 1238 | tens.mem_type = scratched_fms[tens][1] |
| 1239 | |
| 1240 | @staticmethod |
| 1241 | def keep(lr, base_mem_usage, staging_limit): |
| 1242 | for t in range(lr.start_time, lr.end_time + 1): |
| 1243 | base_mem_usage[t] += lr.size |
| 1244 | assert base_mem_usage[t] <= staging_limit |
| 1245 | |
| 1246 | def allocate_component(self, allocator, lrs, max_mem, min_mem, staging_limit, scratched_fms): |
| 1247 | sz = len(lrs) |
| 1248 | allocator.lrs = lrs |
| 1249 | allocator.evicted = [0] * len(lrs) |
| 1250 | allocator.curr_evicted = [0] * sz |
| 1251 | allocator.best_allocated_size = -1 |
| 1252 | # Recursively evaluate all permutations of allocations of the lrs found in the component |
| 1253 | allocator.allocate_exhaustive(0, 0) |
| 1254 | |
| 1255 | # Optimal allocation has been found, move lrs accordingly |
| 1256 | for i, e in enumerate(allocator.evicted): |
| 1257 | if e: |
| 1258 | self.evict(lrs[i], max_mem, scratched_fms) |
| 1259 | else: |
| 1260 | self.keep(lrs[i], min_mem, staging_limit) |
| 1261 | |
| 1262 | |
Tim Hall | d8339a7 | 2021-05-27 18:49:40 +0100 | [diff] [blame] | 1263 | def schedule_passes(nng: Graph, arch: ArchitectureFeatures, options, scheduler_options: SchedulerOptions): |
| 1264 | """Entry point for the Scheduler""" |
| 1265 | # Initialize CPU subgraphs |
| 1266 | schedulers = dict() |
| 1267 | # Initialize schedulers with max schedule. Only schedule NPU subgraphs |
Andreas Nevalainen | 27d36f0 | 2020-11-19 11:27:50 +0100 | [diff] [blame] | 1268 | for sg in nng.subgraphs: |
Tim Hall | d8339a7 | 2021-05-27 18:49:40 +0100 | [diff] [blame] | 1269 | if sg.placement != PassPlacement.Npu: |
| 1270 | # Create cascaded passes for CPU Ops |
| 1271 | cascaded_passes = [] |
| 1272 | for idx, ps in enumerate(sg.passes): |
| 1273 | cps = CascadedPass( |
Jonas Ohlsson | d857507 | 2022-03-30 10:30:25 +0200 | [diff] [blame] | 1274 | ps.name, |
| 1275 | SchedulingStrategy.WeightStream, |
| 1276 | ps.inputs, |
| 1277 | [], |
| 1278 | ps.outputs, |
| 1279 | [ps], |
| 1280 | ps.placement, |
| 1281 | False, |
Tim Hall | d8339a7 | 2021-05-27 18:49:40 +0100 | [diff] [blame] | 1282 | ) |
Andreas Nevalainen | 27d36f0 | 2020-11-19 11:27:50 +0100 | [diff] [blame] | 1283 | |
Tim Hall | d8339a7 | 2021-05-27 18:49:40 +0100 | [diff] [blame] | 1284 | cps.time = idx |
| 1285 | ps.cascade = cps |
| 1286 | cascaded_passes.append(cps) |
Andreas Nevalainen | 27d36f0 | 2020-11-19 11:27:50 +0100 | [diff] [blame] | 1287 | |
Tim Hall | d8339a7 | 2021-05-27 18:49:40 +0100 | [diff] [blame] | 1288 | sg.cascaded_passes = cascaded_passes |
| 1289 | else: |
| 1290 | # Npu subgraph - create schedule |
| 1291 | scheduler = Scheduler(nng, sg, arch, scheduler_options) |
| 1292 | schedulers[sg] = scheduler |
Andreas Nevalainen | 897cc14 | 2020-10-28 15:42:08 +0100 | [diff] [blame] | 1293 | |
Tim Hall | d8339a7 | 2021-05-27 18:49:40 +0100 | [diff] [blame] | 1294 | scheduler.create_scheduler_representation(arch) |
| 1295 | sg.sched_ops = scheduler.sched_ops |
| 1296 | scheduler.move_constant_data() |
Andreas Nevalainen | 897cc14 | 2020-10-28 15:42:08 +0100 | [diff] [blame] | 1297 | |
Tim Hall | d8339a7 | 2021-05-27 18:49:40 +0100 | [diff] [blame] | 1298 | # Create the Max schedule template |
| 1299 | max_schedule_template = scheduler.create_initial_schedule() |
| 1300 | scheduler.max_schedule = max_schedule_template |
Andreas Nevalainen | 897cc14 | 2020-10-28 15:42:08 +0100 | [diff] [blame] | 1301 | |
Tim Hall | d8339a7 | 2021-05-27 18:49:40 +0100 | [diff] [blame] | 1302 | # Create the optimimised Max schedule |
| 1303 | sg.schedule = max_schedule_template |
| 1304 | scheduler.update_op_memory_snapshot(max_schedule_template) |
Tim Hall | 789e6f3 | 2021-06-17 17:02:31 +0100 | [diff] [blame] | 1305 | opt_max_schedule = scheduler.propose_schedule_buffering(max_schedule_template, 1 << 32) |
Tim Hall | d8339a7 | 2021-05-27 18:49:40 +0100 | [diff] [blame] | 1306 | sg.schedule = opt_max_schedule |
| 1307 | scheduler.update_op_memory_snapshot(opt_max_schedule) |
Andreas Nevalainen | 897cc14 | 2020-10-28 15:42:08 +0100 | [diff] [blame] | 1308 | |
Tim Hall | d8339a7 | 2021-05-27 18:49:40 +0100 | [diff] [blame] | 1309 | # Create Min schedule |
| 1310 | min_schedule = scheduler.propose_minimal_schedule() |
| 1311 | initial_sram_limit = scheduler_options.optimization_sram_limit |
| 1312 | if scheduler_options.optimization_strategy == OptimizationStrategy.Size: |
| 1313 | initial_sram_limit = scheduler.min_memory_req |
Andreas Nevalainen | 897cc14 | 2020-10-28 15:42:08 +0100 | [diff] [blame] | 1314 | |
Tim Hall | d8339a7 | 2021-05-27 18:49:40 +0100 | [diff] [blame] | 1315 | cascade_builder = CascadeBuilder(scheduler.sched_ops, arch.is_spilling_enabled()) |
| 1316 | cascade_builder.build_cascades(min_schedule, max_schedule_template, initial_sram_limit) |
| 1317 | sg.schedule = min_schedule |
| 1318 | scheduler.update_op_memory_snapshot(min_schedule) |
Andreas Nevalainen | 897cc14 | 2020-10-28 15:42:08 +0100 | [diff] [blame] | 1319 | |
Tim Hall | d8339a7 | 2021-05-27 18:49:40 +0100 | [diff] [blame] | 1320 | if scheduler_options.optimization_strategy == OptimizationStrategy.Performance: |
| 1321 | # Create an optimized schedule |
| 1322 | sg.schedule = scheduler.optimize_schedule( |
| 1323 | min_schedule, opt_max_schedule, max_schedule_template, scheduler_options |
| 1324 | ) |
| 1325 | scheduler.update_op_memory_snapshot(sg.schedule) |
Andreas Nevalainen | 897cc14 | 2020-10-28 15:42:08 +0100 | [diff] [blame] | 1326 | |
Tim Hall | d8339a7 | 2021-05-27 18:49:40 +0100 | [diff] [blame] | 1327 | scheduler.apply_schedule(sg.schedule) |
| 1328 | scheduler.use_fast_storage_for_feature_maps(sg.schedule, scheduler_options.optimization_sram_limit) |
Andreas Nevalainen | 897cc14 | 2020-10-28 15:42:08 +0100 | [diff] [blame] | 1329 | |
Tim Hall | d8339a7 | 2021-05-27 18:49:40 +0100 | [diff] [blame] | 1330 | if scheduler_options.verbose_schedule: |
| 1331 | scheduler.print_schedule(sg.schedule) |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 1332 | |
Tim Hall | d8339a7 | 2021-05-27 18:49:40 +0100 | [diff] [blame] | 1333 | # Evaluate schedule |
| 1334 | _update_tensor_allocation(nng, arch, options) |