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 | # Build a live range graph for tensors in one or more subgraphs. Used for tensor allocation as well as in the scheduler. |
| 18 | # Can work with either a pass packed subgraph or a scheduled subgraph. |
Louis Verhaard | 226ecaf | 2021-03-30 10:18:28 +0200 | [diff] [blame] | 19 | from typing import List |
| 20 | |
Tim Hall | d8339a7 | 2021-05-27 18:49:40 +0100 | [diff] [blame] | 21 | import numpy as np |
| 22 | |
Diego Russo | e8a1045 | 2020-04-21 17:39:10 +0100 | [diff] [blame] | 23 | from .nn_graph import PassPlacement |
Louis Verhaard | aee5d75 | 2020-09-30 09:01:52 +0200 | [diff] [blame] | 24 | from .operation import Op |
Tim Hall | d8339a7 | 2021-05-27 18:49:40 +0100 | [diff] [blame] | 25 | from .tensor import MemArea |
Patrik Gustavsson | eca2e95 | 2020-05-27 09:15:11 +0200 | [diff] [blame] | 26 | from .tensor import MemType |
Diego Russo | e8a1045 | 2020-04-21 17:39:10 +0100 | [diff] [blame] | 27 | from .tensor import Tensor |
Tim Hall | d8339a7 | 2021-05-27 18:49:40 +0100 | [diff] [blame] | 28 | from .tensor import TensorPurpose |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 29 | |
| 30 | |
| 31 | class LiveRange: |
Jacob Bohlin | 0628a8c | 2020-08-28 13:25:14 +0200 | [diff] [blame] | 32 | def __init__(self, tens, alignment): |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 33 | self.tensors = [] # Tensors that are assigned to the same LiveRange will be allocated to the same address |
| 34 | self.start_time = 99999999999 |
| 35 | self.end_time = -1 |
| 36 | self.size = 0 |
| 37 | self.name = "" |
Jacob Bohlin | 0628a8c | 2020-08-28 13:25:14 +0200 | [diff] [blame] | 38 | self.alignment = alignment |
Tim Hall | d8339a7 | 2021-05-27 18:49:40 +0100 | [diff] [blame] | 39 | self.mem_area = tens.mem_area if tens else MemArea.Unknown |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 40 | |
| 41 | if tens: |
| 42 | self.add_tensor(tens) |
| 43 | |
| 44 | def __str__(self): |
| 45 | return "<live_range.LiveRange: '%s' start_time=%s, end_time=%s>" % (self.name, self.start_time, self.end_time) |
| 46 | |
| 47 | __repr__ = __str__ |
| 48 | |
| 49 | def add_tensor(self, tens): |
| 50 | if self.size == 0: |
| 51 | self.size = tens.storage_size() |
| 52 | self.name = tens.name # LiveRange will be named after the first tensor added |
| 53 | else: |
| 54 | assert ( |
| 55 | self.size >= tens.storage_size() |
| 56 | ), "Tensors assigned to the same LiveRange need to fit the size of the LiveRange." |
| 57 | |
| 58 | self.tensors.append(tens) |
| 59 | |
Tim Hall | d8339a7 | 2021-05-27 18:49:40 +0100 | [diff] [blame] | 60 | def mark_usage(self, op_time, op_length=1): |
| 61 | op_time_start = max(op_time, 0) |
| 62 | op_time_end = op_time + op_length |
| 63 | if op_time_end <= op_time_start: |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 64 | return |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 65 | |
| 66 | self.start_time = min(self.start_time, op_time_start) |
| 67 | self.end_time = max(self.end_time, op_time_end) |
| 68 | |
Tim Hall | d8339a7 | 2021-05-27 18:49:40 +0100 | [diff] [blame] | 69 | def set_buffer_size(self, buffer_size): |
| 70 | self.size = buffer_size |
| 71 | self.mem_area = MemArea.Sram |
| 72 | |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 73 | def overlaps_ranges(self, other): |
| 74 | return max(self.start_time, other.start_time) < min(self.end_time, other.end_time) |
| 75 | |
| 76 | def overlaps_address(self, other): |
| 77 | # Returns the first pair of tensors in this LiveRange and 'other' which have |
| 78 | # overlapping addresses |
| 79 | for tens in self.tensors: |
| 80 | for other_tens in other.tensors: |
| 81 | if max(tens.address, other_tens.address) < min( |
| 82 | tens.address + self.size, other_tens.address + other.size |
| 83 | ): |
| 84 | return True, tens, other_tens |
| 85 | |
| 86 | return False, None, None |
| 87 | |
| 88 | def __lt__(self, other): |
| 89 | if self.start_time != other.start_time: |
| 90 | return self.start_time < other.start_time |
| 91 | if self.end_time != other.end_time: |
| 92 | return self.end_time < other.end_time |
| 93 | if self.size != other.size: |
| 94 | return self.size < other.size |
| 95 | return self.name < other.name |
| 96 | |
| 97 | def set_address(self, address): |
Jacob Bohlin | 1a66697 | 2020-09-11 10:04:15 +0200 | [diff] [blame] | 98 | # Set address of all tensors in LiveRange |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 99 | for tens in self.tensors: |
Jacob Bohlin | 1a66697 | 2020-09-11 10:04:15 +0200 | [diff] [blame] | 100 | tens.address = address |
| 101 | |
| 102 | return address |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 103 | |
| 104 | def get_alignment(self): |
Jacob Bohlin | 0628a8c | 2020-08-28 13:25:14 +0200 | [diff] [blame] | 105 | return self.alignment |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 106 | |
Jacob Bohlin | 0628a8c | 2020-08-28 13:25:14 +0200 | [diff] [blame] | 107 | def set_alignment(self, alignment): |
| 108 | self.alignment = max(self.alignment, alignment) |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 109 | |
| 110 | |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 111 | class LiveRangeGraph: |
| 112 | def __init__(self): |
Louis Verhaard | 226ecaf | 2021-03-30 10:18:28 +0200 | [diff] [blame] | 113 | self.lrs: List[LiveRange] = [] # List of all created ranges |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 114 | self.ranges = {} # tens -> range |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 115 | self.ignore_tensors = set() |
| 116 | self.processed_subgraphs = set() |
| 117 | self.current_time = 0 |
Tim Hall | d8339a7 | 2021-05-27 18:49:40 +0100 | [diff] [blame] | 118 | self.end_time = None |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 119 | |
Jacob Bohlin | 0628a8c | 2020-08-28 13:25:14 +0200 | [diff] [blame] | 120 | def get_or_create_range(self, tens, alignment=Tensor.AllocationQuantum): |
Jacob Bohlin | 1a66697 | 2020-09-11 10:04:15 +0200 | [diff] [blame] | 121 | # Return the live range of the tensor (or any of its clones) |
| 122 | for existing_tensor, rng in self.ranges.items(): |
| 123 | if tens.equivalent(existing_tensor): |
Jacob Bohlin | 0628a8c | 2020-08-28 13:25:14 +0200 | [diff] [blame] | 124 | rng.set_alignment(alignment) |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 125 | return rng |
| 126 | |
| 127 | # No live range found for the tensor, create a new one |
Jacob Bohlin | 0628a8c | 2020-08-28 13:25:14 +0200 | [diff] [blame] | 128 | rng = LiveRange(tens, alignment) |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 129 | self.ranges[tens] = rng |
Louis Verhaard | 226ecaf | 2021-03-30 10:18:28 +0200 | [diff] [blame] | 130 | self.lrs.append(rng) |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 131 | return rng |
| 132 | |
| 133 | def fuse_ranges(self, in_tens, out_tens): |
| 134 | live_range = self.get_or_create_range(in_tens) |
| 135 | assert out_tens not in self.ranges, out_tens |
| 136 | live_range.add_tensor(out_tens) |
| 137 | self.ranges[out_tens] = live_range |
| 138 | return live_range |
| 139 | |
Tim Hall | d8339a7 | 2021-05-27 18:49:40 +0100 | [diff] [blame] | 140 | def update_endtime(self): |
| 141 | self.end_time = 0 |
| 142 | for rng in self.ranges.values(): |
| 143 | self.end_time = max(self.end_time, rng.end_time) |
| 144 | return self.end_time + 1 |
| 145 | |
| 146 | def get_temporal_memory_usage(self, target_mem_area): |
| 147 | if not self.end_time: |
| 148 | self.update_endtime() |
| 149 | usage = np.zeros(self.end_time, dtype=np.int32) |
| 150 | for rng in self.ranges.values(): |
| 151 | if rng.mem_area == target_mem_area: |
| 152 | # End time is inclusive |
| 153 | usage[rng.start_time : rng.end_time + 1] += rng.size |
| 154 | |
| 155 | return usage |
| 156 | |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 157 | |
Patrik Gustavsson | a151f59 | 2020-10-16 13:59:52 +0200 | [diff] [blame] | 158 | def tensor_should_be_ignored(lr_graph, tens, target_mem_area, target_mem_type_set): |
| 159 | if tens.mem_area != target_mem_area or tens.mem_type not in target_mem_type_set: |
| 160 | return True |
| 161 | if tens in lr_graph.ignore_tensors: |
| 162 | return True |
| 163 | if tens.name.endswith("reshape_shape_npu"): |
| 164 | # Reshape tensor, no need to allocate |
| 165 | lr_graph.ignore_tensors.add(tens) |
| 166 | return True |
| 167 | return False |
| 168 | |
| 169 | |
| 170 | # Tries merging of ifm/ofm live ranges for memory only ops and elementwise ops |
| 171 | def merge_op_ranges(sg, lr_graph, target_mem_area, target_mem_type_set): |
| 172 | for ps in sg.passes: |
| 173 | if ps.placement == PassPlacement.MemoryOnly: |
| 174 | # For memory only passes, e.g. Reshape. Add input and output tensor to the same LiveRange |
| 175 | input_tensor = ps.inputs[0] |
| 176 | output_tensor = ps.outputs[0] |
| 177 | if not tensor_should_be_ignored(lr_graph, input_tensor, target_mem_area, target_mem_type_set) and not ( |
| 178 | tensor_should_be_ignored(lr_graph, output_tensor, target_mem_area, target_mem_type_set) |
| 179 | ): |
| 180 | lr_graph.fuse_ranges(input_tensor, output_tensor) |
| 181 | elif ps.is_element_wise: |
| 182 | merge_elementwise_op_ranges(ps, lr_graph, target_mem_area, target_mem_type_set) |
| 183 | |
| 184 | |
| 185 | # Tries to merge ifm/ofm live of elementwise op |
| 186 | def merge_elementwise_op_ranges(ps, lr_graph, target_mem_area, target_mem_type_set): |
| 187 | elem_op = None |
| 188 | for op in ps.ops: |
| 189 | if op.type.is_elementwise_op(): |
| 190 | assert elem_op is None |
| 191 | elem_op = op |
| 192 | |
| 193 | if elem_op is not None and not tensor_should_be_ignored( |
| 194 | lr_graph, elem_op.ofm, target_mem_area, target_mem_type_set |
| 195 | ): |
| 196 | # Check if overwriting the inputs can be allowed |
| 197 | if elem_op.type not in (Op.SHL, Op.SHR): |
| 198 | inps = [] |
| 199 | if ( |
| 200 | elem_op.ifm is not None |
| 201 | and elem_op.ifm.shape != [] |
| 202 | and elem_op.ifm.mem_area == target_mem_area |
| 203 | and elem_op.ifm.mem_type in target_mem_type_set |
| 204 | ): |
| 205 | inps.append(elem_op.ifm) |
| 206 | if ( |
| 207 | elem_op.ifm2 is not None |
| 208 | and elem_op.ifm2.shape != [] |
| 209 | and elem_op.ifm2.mem_area == target_mem_area |
| 210 | and elem_op.ifm.mem_type in target_mem_type_set |
| 211 | ): |
| 212 | inps.append(elem_op.ifm2) |
| 213 | |
| 214 | if len(inps) > 0: |
Patrik Gustavsson | 2349d42 | 2020-12-01 16:02:29 +0100 | [diff] [blame] | 215 | for i, inp in enumerate(inps): |
Patrik Gustavsson | a151f59 | 2020-10-16 13:59:52 +0200 | [diff] [blame] | 216 | # check input format, dtype, broadcasting or if there are more input consumers |
| 217 | if ( |
| 218 | inp.format == elem_op.ofm.format |
| 219 | and inp.dtype == elem_op.ofm.dtype |
Patrik Gustavsson | 2349d42 | 2020-12-01 16:02:29 +0100 | [diff] [blame] | 220 | and elem_op.ifm_shapes[i] == elem_op.ofm_shapes[0] |
Patrik Gustavsson | a151f59 | 2020-10-16 13:59:52 +0200 | [diff] [blame] | 221 | and (len(inp.consumer_list) == 1 and len(inp.ops) == 1) |
| 222 | ): |
| 223 | lr_graph.fuse_ranges(inp, elem_op.ofm) |
| 224 | break |
| 225 | |
| 226 | |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 227 | def extract_live_ranges_from_passes( |
Michael McGeagh | 6f72526 | 2020-12-03 15:21:36 +0000 | [diff] [blame] | 228 | sg, target_mem_area, target_mem_type_set=None, ignore_subgraph_input_output_tensors=False, |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 229 | ): |
| 230 | lr_graph = LiveRangeGraph() |
| 231 | |
| 232 | if ignore_subgraph_input_output_tensors: |
| 233 | lr_graph.ignore_tensors.update(sg.input_tensors) |
| 234 | lr_graph.ignore_tensors.update(sg.output_tensors) |
| 235 | |
Michael McGeagh | 6f72526 | 2020-12-03 15:21:36 +0000 | [diff] [blame] | 236 | if target_mem_type_set is None: |
| 237 | target_mem_type_set = set((MemType.Scratch, MemType.Scratch_fast)) |
| 238 | |
Patrik Gustavsson | a151f59 | 2020-10-16 13:59:52 +0200 | [diff] [blame] | 239 | # Try to merge live ranges of operations in the NPU subgraphs |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 240 | if sg.placement == PassPlacement.Npu: |
Patrik Gustavsson | fad90c2 | 2020-11-03 13:07:40 +0100 | [diff] [blame] | 241 | merge_op_ranges(sg, lr_graph, target_mem_area, target_mem_type_set) |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 242 | |
| 243 | for idx, ps in enumerate(sg.passes): |
| 244 | ps.time = 2 * idx |
| 245 | |
| 246 | time_for_pass = ps.time |
| 247 | |
Patrik Gustavsson | a151f59 | 2020-10-16 13:59:52 +0200 | [diff] [blame] | 248 | for tens in ps.inputs + ps.intermediates + ps.outputs: |
Patrik Gustavsson | fad90c2 | 2020-11-03 13:07:40 +0100 | [diff] [blame] | 249 | if tensor_should_be_ignored(lr_graph, tens, target_mem_area, target_mem_type_set): |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 250 | continue |
| 251 | rng = lr_graph.get_or_create_range(tens) |
| 252 | rng.mark_usage(time_for_pass) |
| 253 | |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 254 | end_time = len(sg.passes) * 2 |
| 255 | for tens in sg.output_tensors: |
Patrik Gustavsson | fad90c2 | 2020-11-03 13:07:40 +0100 | [diff] [blame] | 256 | if tensor_should_be_ignored(lr_graph, tens, target_mem_area, target_mem_type_set): |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 257 | continue |
| 258 | rng = lr_graph.get_or_create_range(tens) |
| 259 | rng.mark_usage(end_time) |
| 260 | |
| 261 | return lr_graph |
| 262 | |
| 263 | |
| 264 | def extract_live_ranges_from_cascaded_passes( |
| 265 | sg, |
| 266 | target_mem_area, |
Patrik Gustavsson | eca2e95 | 2020-05-27 09:15:11 +0200 | [diff] [blame] | 267 | target_mem_type_set, |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 268 | ignore_subgraph_input_output_tensors=False, |
| 269 | lr_graph=None, |
Tim Hall | b9b515c | 2020-11-01 21:27:19 +0000 | [diff] [blame] | 270 | cpu_tensor_alignment=Tensor.AllocationQuantum, |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 271 | ): |
Diego Russo | ea6111a | 2020-04-14 18:41:58 +0100 | [diff] [blame] | 272 | if lr_graph is None: |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 273 | lr_graph = LiveRangeGraph() |
| 274 | |
| 275 | if sg in lr_graph.processed_subgraphs: |
| 276 | # if subgraph has been processed already, return the lr_graph as is |
| 277 | return lr_graph |
| 278 | |
| 279 | if ignore_subgraph_input_output_tensors: |
| 280 | lr_graph.ignore_tensors.update(sg.input_tensors) |
| 281 | lr_graph.ignore_tensors.update(sg.output_tensors) |
| 282 | |
Patrik Gustavsson | a151f59 | 2020-10-16 13:59:52 +0200 | [diff] [blame] | 283 | # Try to merge live ranges of operations in the NPU subgraphs |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 284 | if sg.placement == PassPlacement.Npu: |
Patrik Gustavsson | a151f59 | 2020-10-16 13:59:52 +0200 | [diff] [blame] | 285 | merge_op_ranges(sg, lr_graph, target_mem_area, target_mem_type_set) |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 286 | |
| 287 | for cps in sg.cascaded_passes: |
| 288 | cps.time = lr_graph.current_time |
| 289 | |
| 290 | time_for_pass = cps.time |
| 291 | |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 292 | for tens in cps.inputs: |
Patrik Gustavsson | a151f59 | 2020-10-16 13:59:52 +0200 | [diff] [blame] | 293 | if tensor_should_be_ignored(lr_graph, tens, target_mem_area, target_mem_type_set): |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 294 | continue |
Tim Hall | b9b515c | 2020-11-01 21:27:19 +0000 | [diff] [blame] | 295 | rng = lr_graph.get_or_create_range(tens, cpu_tensor_alignment) |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 296 | rng.mark_usage(time_for_pass) |
| 297 | |
| 298 | cps_primary_op = cps.passes[0].primary_op |
Patrik Gustavsson | eca2e95 | 2020-05-27 09:15:11 +0200 | [diff] [blame] | 299 | |
Louis Verhaard | aee5d75 | 2020-09-30 09:01:52 +0200 | [diff] [blame] | 300 | if ( |
| 301 | cps_primary_op |
| 302 | and cps_primary_op.type == Op.CustomNpuOp |
| 303 | and MemType.Permanent_CPU not in target_mem_type_set |
| 304 | ): |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 305 | # If the primary-op is an NpuOp that means this is where an Npu subgraph |
| 306 | # is called. Go into said subgraph and extract live ranges before continuing. |
Jacob Bohlin | 0628a8c | 2020-08-28 13:25:14 +0200 | [diff] [blame] | 307 | # Use default allocation alignment of 16 for Npu tensors |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 308 | npu_sg = cps_primary_op.attrs["subgraph"] |
Tim Hall | d8339a7 | 2021-05-27 18:49:40 +0100 | [diff] [blame] | 309 | lr_graph = _extract_live_ranges_from_schedule(npu_sg, target_mem_area, target_mem_type_set, lr_graph) |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 310 | # Set the new time after handling the Npu subgraph |
| 311 | time_for_pass = lr_graph.current_time |
| 312 | cps.time = time_for_pass |
| 313 | |
Patrik Gustavsson | a151f59 | 2020-10-16 13:59:52 +0200 | [diff] [blame] | 314 | for tens in cps.intermediates + cps.outputs: |
| 315 | if tensor_should_be_ignored(lr_graph, tens, target_mem_area, target_mem_type_set): |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 316 | continue |
Tim Hall | b9b515c | 2020-11-01 21:27:19 +0000 | [diff] [blame] | 317 | rng = lr_graph.get_or_create_range(tens, cpu_tensor_alignment) |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 318 | rng.mark_usage(time_for_pass) |
| 319 | |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 320 | lr_graph.current_time += 2 |
| 321 | |
| 322 | end_time = 0 |
| 323 | for rng in lr_graph.ranges.values(): |
| 324 | # Find the maximum end time of all live-ranges in the graph |
| 325 | end_time = max(end_time, rng.end_time) |
| 326 | |
| 327 | for tens in sg.output_tensors: |
Patrik Gustavsson | a151f59 | 2020-10-16 13:59:52 +0200 | [diff] [blame] | 328 | if tensor_should_be_ignored(lr_graph, tens, target_mem_area, target_mem_type_set): |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 329 | continue |
Tim Hall | b9b515c | 2020-11-01 21:27:19 +0000 | [diff] [blame] | 330 | rng = lr_graph.get_or_create_range(tens, cpu_tensor_alignment) |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 331 | rng.mark_usage(end_time) |
| 332 | |
| 333 | # Add subgraph to set of processed subgraphs |
| 334 | lr_graph.processed_subgraphs.add(sg) |
| 335 | return lr_graph |
Tim Hall | d8339a7 | 2021-05-27 18:49:40 +0100 | [diff] [blame] | 336 | |
| 337 | |
| 338 | def create_linear_live_range_graph(sg, target_mem_area, target_mem_type_set, lr_graph): |
| 339 | assert lr_graph is not None |
| 340 | sg_time = lr_graph.current_time |
| 341 | for ps in sg.passes: |
| 342 | for tens in ps.inputs + ps.outputs + ps.intermediates: |
| 343 | if tens.purpose == TensorPurpose.Weights or tensor_should_be_ignored( |
| 344 | lr_graph, tens, target_mem_area, target_mem_type_set |
| 345 | ): |
| 346 | continue |
| 347 | |
| 348 | rng = lr_graph.get_or_create_range(tens) |
| 349 | rng.mark_usage(sg_time) |
| 350 | |
| 351 | for sched_op, op_info in sg.schedule.cost_map.items(): |
| 352 | if op_info.npu_weights_tensor and not ( |
| 353 | tensor_should_be_ignored(lr_graph, op_info.npu_weights_tensor, target_mem_area, target_mem_type_set) |
| 354 | ): |
| 355 | rng = lr_graph.get_or_create_range(op_info.npu_weights_tensor) |
| 356 | rng.mark_usage(sg_time) |
| 357 | |
| 358 | lr_graph.current_time += 1 |
| 359 | return lr_graph |
| 360 | |
| 361 | |
| 362 | def _extract_live_ranges_from_schedule(sg, target_mem_area, target_mem_type_set, lr_graph): |
| 363 | time_for_cascade = {} |
| 364 | for sched_op in sg.sched_ops: |
| 365 | op_info = sg.schedule.cost_map[sched_op] |
| 366 | cascade = op_info.cascade |
| 367 | cascade_info = sg.schedule.cascades.get(cascade, None) |
| 368 | |
| 369 | time_to_set = time_for_cascade.get(cascade, lr_graph.current_time) |
| 370 | |
| 371 | op_info.time_index = time_to_set |
| 372 | |
| 373 | # Mark usage for all tensors related to this Pass |
| 374 | ps = sched_op.parent_ps |
| 375 | for tens in ps.inputs + ps.outputs + ps.intermediates: |
| 376 | if ( |
| 377 | target_mem_area == MemArea.Sram |
| 378 | and cascade_info |
| 379 | and tens == ps.ifm_tensor |
| 380 | and sched_op in cascade_info.buffers |
| 381 | ): |
| 382 | # This tensor is a rolling buffer in a cascade and the size of the LiveRange needs to be modified |
| 383 | # for enabling temporal memory snapshots without modifying the original Tensor |
| 384 | rng = lr_graph.get_or_create_range(tens) |
| 385 | rng.set_buffer_size(cascade_info.buffers[sched_op].elements() * sched_op.ifm.dtype.size_in_bytes()) |
| 386 | elif ( |
| 387 | tens.purpose == TensorPurpose.Weights |
| 388 | or tens.purpose == TensorPurpose.FSBias |
| 389 | or tens.mem_type not in target_mem_type_set |
| 390 | or tens.mem_area != target_mem_area |
| 391 | ): |
| 392 | continue |
| 393 | |
| 394 | else: |
| 395 | rng = lr_graph.get_or_create_range(tens) |
| 396 | |
| 397 | rng.mark_usage(time_to_set) |
| 398 | |
| 399 | weight_tens = op_info.buffered_weight_tensor |
| 400 | if weight_tens and weight_tens.mem_type in target_mem_type_set and weight_tens.mem_area == target_mem_area: |
| 401 | rng = lr_graph.get_or_create_range(weight_tens) |
| 402 | if weight_tens.pre_buffer: |
| 403 | rng.mark_usage(time_to_set - 1, 2) |
| 404 | else: |
| 405 | rng.mark_usage(time_to_set) |
| 406 | |
| 407 | if time_to_set == lr_graph.current_time: |
| 408 | lr_graph.current_time += 2 |
| 409 | |
| 410 | if cascade != 0: |
| 411 | time_for_cascade[cascade] = time_to_set |
| 412 | |
| 413 | end_time = lr_graph.update_endtime() |
| 414 | |
| 415 | for tens in sg.output_tensors: |
| 416 | if tens.mem_type not in target_mem_type_set or tens.mem_area != target_mem_area: |
| 417 | continue |
| 418 | rng = lr_graph.get_or_create_range(tens) |
| 419 | rng.mark_usage(end_time) |
| 420 | |
| 421 | return lr_graph |