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. |
Diego Russo | e8a1045 | 2020-04-21 17:39:10 +0100 | [diff] [blame] | 19 | from .nn_graph import PassPlacement |
Louis Verhaard | aee5d75 | 2020-09-30 09:01:52 +0200 | [diff] [blame] | 20 | from .operation import Op |
Patrik Gustavsson | eca2e95 | 2020-05-27 09:15:11 +0200 | [diff] [blame] | 21 | from .tensor import MemType |
Diego Russo | e8a1045 | 2020-04-21 17:39:10 +0100 | [diff] [blame] | 22 | from .tensor import Tensor |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 23 | |
| 24 | |
| 25 | class LiveRange: |
Jacob Bohlin | 0628a8c | 2020-08-28 13:25:14 +0200 | [diff] [blame] | 26 | def __init__(self, tens, alignment): |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 27 | self.tensors = [] # Tensors that are assigned to the same LiveRange will be allocated to the same address |
| 28 | self.start_time = 99999999999 |
| 29 | self.end_time = -1 |
| 30 | self.size = 0 |
| 31 | self.name = "" |
Jacob Bohlin | 0628a8c | 2020-08-28 13:25:14 +0200 | [diff] [blame] | 32 | self.alignment = alignment |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 33 | |
| 34 | if tens: |
| 35 | self.add_tensor(tens) |
| 36 | |
| 37 | def __str__(self): |
| 38 | return "<live_range.LiveRange: '%s' start_time=%s, end_time=%s>" % (self.name, self.start_time, self.end_time) |
| 39 | |
| 40 | __repr__ = __str__ |
| 41 | |
| 42 | def add_tensor(self, tens): |
| 43 | if self.size == 0: |
| 44 | self.size = tens.storage_size() |
| 45 | self.name = tens.name # LiveRange will be named after the first tensor added |
| 46 | else: |
| 47 | assert ( |
| 48 | self.size >= tens.storage_size() |
| 49 | ), "Tensors assigned to the same LiveRange need to fit the size of the LiveRange." |
| 50 | |
| 51 | self.tensors.append(tens) |
| 52 | |
| 53 | def mark_usage(self, op_time): |
| 54 | if op_time == -1: |
| 55 | return |
| 56 | op_time_start = op_time |
| 57 | op_time_end = op_time + 1 |
| 58 | |
| 59 | self.start_time = min(self.start_time, op_time_start) |
| 60 | self.end_time = max(self.end_time, op_time_end) |
| 61 | |
| 62 | def overlaps_ranges(self, other): |
| 63 | return max(self.start_time, other.start_time) < min(self.end_time, other.end_time) |
| 64 | |
| 65 | def overlaps_address(self, other): |
| 66 | # Returns the first pair of tensors in this LiveRange and 'other' which have |
| 67 | # overlapping addresses |
| 68 | for tens in self.tensors: |
| 69 | for other_tens in other.tensors: |
| 70 | if max(tens.address, other_tens.address) < min( |
| 71 | tens.address + self.size, other_tens.address + other.size |
| 72 | ): |
| 73 | return True, tens, other_tens |
| 74 | |
| 75 | return False, None, None |
| 76 | |
| 77 | def __lt__(self, other): |
| 78 | if self.start_time != other.start_time: |
| 79 | return self.start_time < other.start_time |
| 80 | if self.end_time != other.end_time: |
| 81 | return self.end_time < other.end_time |
| 82 | if self.size != other.size: |
| 83 | return self.size < other.size |
| 84 | return self.name < other.name |
| 85 | |
| 86 | def set_address(self, address): |
Jacob Bohlin | 1a66697 | 2020-09-11 10:04:15 +0200 | [diff] [blame] | 87 | # Set address of all tensors in LiveRange |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 88 | for tens in self.tensors: |
Jacob Bohlin | 1a66697 | 2020-09-11 10:04:15 +0200 | [diff] [blame] | 89 | tens.address = address |
| 90 | |
| 91 | return address |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 92 | |
| 93 | def get_alignment(self): |
Jacob Bohlin | 0628a8c | 2020-08-28 13:25:14 +0200 | [diff] [blame] | 94 | return self.alignment |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 95 | |
Jacob Bohlin | 0628a8c | 2020-08-28 13:25:14 +0200 | [diff] [blame] | 96 | def set_alignment(self, alignment): |
| 97 | self.alignment = max(self.alignment, alignment) |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 98 | |
| 99 | |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 100 | class LiveRangeGraph: |
| 101 | def __init__(self): |
| 102 | self.ranges = {} # tens -> range |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 103 | self.ignore_tensors = set() |
| 104 | self.processed_subgraphs = set() |
| 105 | self.current_time = 0 |
| 106 | |
Jacob Bohlin | 0628a8c | 2020-08-28 13:25:14 +0200 | [diff] [blame] | 107 | def get_or_create_range(self, tens, alignment=Tensor.AllocationQuantum): |
Jacob Bohlin | 1a66697 | 2020-09-11 10:04:15 +0200 | [diff] [blame] | 108 | # Return the live range of the tensor (or any of its clones) |
| 109 | for existing_tensor, rng in self.ranges.items(): |
| 110 | if tens.equivalent(existing_tensor): |
Jacob Bohlin | 0628a8c | 2020-08-28 13:25:14 +0200 | [diff] [blame] | 111 | rng.set_alignment(alignment) |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 112 | return rng |
| 113 | |
| 114 | # No live range found for the tensor, create a new one |
Jacob Bohlin | 0628a8c | 2020-08-28 13:25:14 +0200 | [diff] [blame] | 115 | rng = LiveRange(tens, alignment) |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 116 | self.ranges[tens] = rng |
| 117 | return rng |
| 118 | |
| 119 | def fuse_ranges(self, in_tens, out_tens): |
| 120 | live_range = self.get_or_create_range(in_tens) |
| 121 | assert out_tens not in self.ranges, out_tens |
| 122 | live_range.add_tensor(out_tens) |
| 123 | self.ranges[out_tens] = live_range |
| 124 | return live_range |
| 125 | |
| 126 | |
Patrik Gustavsson | a151f59 | 2020-10-16 13:59:52 +0200 | [diff] [blame] | 127 | def tensor_should_be_ignored(lr_graph, tens, target_mem_area, target_mem_type_set): |
| 128 | if tens.mem_area != target_mem_area or tens.mem_type not in target_mem_type_set: |
| 129 | return True |
| 130 | if tens in lr_graph.ignore_tensors: |
| 131 | return True |
| 132 | if tens.name.endswith("reshape_shape_npu"): |
| 133 | # Reshape tensor, no need to allocate |
| 134 | lr_graph.ignore_tensors.add(tens) |
| 135 | return True |
| 136 | return False |
| 137 | |
| 138 | |
| 139 | # Tries merging of ifm/ofm live ranges for memory only ops and elementwise ops |
| 140 | def merge_op_ranges(sg, lr_graph, target_mem_area, target_mem_type_set): |
| 141 | for ps in sg.passes: |
| 142 | if ps.placement == PassPlacement.MemoryOnly: |
| 143 | # For memory only passes, e.g. Reshape. Add input and output tensor to the same LiveRange |
| 144 | input_tensor = ps.inputs[0] |
| 145 | output_tensor = ps.outputs[0] |
| 146 | if not tensor_should_be_ignored(lr_graph, input_tensor, target_mem_area, target_mem_type_set) and not ( |
| 147 | tensor_should_be_ignored(lr_graph, output_tensor, target_mem_area, target_mem_type_set) |
| 148 | ): |
| 149 | lr_graph.fuse_ranges(input_tensor, output_tensor) |
| 150 | elif ps.is_element_wise: |
| 151 | merge_elementwise_op_ranges(ps, lr_graph, target_mem_area, target_mem_type_set) |
| 152 | |
| 153 | |
| 154 | # Tries to merge ifm/ofm live of elementwise op |
| 155 | def merge_elementwise_op_ranges(ps, lr_graph, target_mem_area, target_mem_type_set): |
| 156 | elem_op = None |
| 157 | for op in ps.ops: |
| 158 | if op.type.is_elementwise_op(): |
| 159 | assert elem_op is None |
| 160 | elem_op = op |
| 161 | |
| 162 | if elem_op is not None and not tensor_should_be_ignored( |
| 163 | lr_graph, elem_op.ofm, target_mem_area, target_mem_type_set |
| 164 | ): |
| 165 | # Check if overwriting the inputs can be allowed |
| 166 | if elem_op.type not in (Op.SHL, Op.SHR): |
| 167 | inps = [] |
| 168 | if ( |
| 169 | elem_op.ifm is not None |
| 170 | and elem_op.ifm.shape != [] |
| 171 | and elem_op.ifm.mem_area == target_mem_area |
| 172 | and elem_op.ifm.mem_type in target_mem_type_set |
| 173 | ): |
| 174 | inps.append(elem_op.ifm) |
| 175 | if ( |
| 176 | elem_op.ifm2 is not None |
| 177 | and elem_op.ifm2.shape != [] |
| 178 | and elem_op.ifm2.mem_area == target_mem_area |
| 179 | and elem_op.ifm.mem_type in target_mem_type_set |
| 180 | ): |
| 181 | inps.append(elem_op.ifm2) |
| 182 | |
| 183 | if len(inps) > 0: |
Patrik Gustavsson | 2349d42 | 2020-12-01 16:02:29 +0100 | [diff] [blame] | 184 | for i, inp in enumerate(inps): |
Patrik Gustavsson | a151f59 | 2020-10-16 13:59:52 +0200 | [diff] [blame] | 185 | # check input format, dtype, broadcasting or if there are more input consumers |
| 186 | if ( |
| 187 | inp.format == elem_op.ofm.format |
| 188 | and inp.dtype == elem_op.ofm.dtype |
Patrik Gustavsson | 2349d42 | 2020-12-01 16:02:29 +0100 | [diff] [blame] | 189 | and elem_op.ifm_shapes[i] == elem_op.ofm_shapes[0] |
Patrik Gustavsson | a151f59 | 2020-10-16 13:59:52 +0200 | [diff] [blame] | 190 | and (len(inp.consumer_list) == 1 and len(inp.ops) == 1) |
| 191 | ): |
| 192 | lr_graph.fuse_ranges(inp, elem_op.ofm) |
| 193 | break |
| 194 | |
| 195 | |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 196 | def extract_live_ranges_from_passes( |
Michael McGeagh | 6f72526 | 2020-12-03 15:21:36 +0000 | [diff] [blame] | 197 | 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] | 198 | ): |
| 199 | lr_graph = LiveRangeGraph() |
| 200 | |
| 201 | if ignore_subgraph_input_output_tensors: |
| 202 | lr_graph.ignore_tensors.update(sg.input_tensors) |
| 203 | lr_graph.ignore_tensors.update(sg.output_tensors) |
| 204 | |
Michael McGeagh | 6f72526 | 2020-12-03 15:21:36 +0000 | [diff] [blame] | 205 | if target_mem_type_set is None: |
| 206 | target_mem_type_set = set((MemType.Scratch, MemType.Scratch_fast)) |
| 207 | |
Patrik Gustavsson | a151f59 | 2020-10-16 13:59:52 +0200 | [diff] [blame] | 208 | # Try to merge live ranges of operations in the NPU subgraphs |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 209 | if sg.placement == PassPlacement.Npu: |
Patrik Gustavsson | fad90c2 | 2020-11-03 13:07:40 +0100 | [diff] [blame] | 210 | 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] | 211 | |
| 212 | for idx, ps in enumerate(sg.passes): |
| 213 | ps.time = 2 * idx |
| 214 | |
| 215 | time_for_pass = ps.time |
| 216 | |
Patrik Gustavsson | a151f59 | 2020-10-16 13:59:52 +0200 | [diff] [blame] | 217 | for tens in ps.inputs + ps.intermediates + ps.outputs: |
Patrik Gustavsson | fad90c2 | 2020-11-03 13:07:40 +0100 | [diff] [blame] | 218 | 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] | 219 | continue |
| 220 | rng = lr_graph.get_or_create_range(tens) |
| 221 | rng.mark_usage(time_for_pass) |
| 222 | |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 223 | end_time = len(sg.passes) * 2 |
| 224 | for tens in sg.output_tensors: |
Patrik Gustavsson | fad90c2 | 2020-11-03 13:07:40 +0100 | [diff] [blame] | 225 | 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] | 226 | continue |
| 227 | rng = lr_graph.get_or_create_range(tens) |
| 228 | rng.mark_usage(end_time) |
| 229 | |
| 230 | return lr_graph |
| 231 | |
| 232 | |
| 233 | def extract_live_ranges_from_cascaded_passes( |
| 234 | sg, |
| 235 | target_mem_area, |
Patrik Gustavsson | eca2e95 | 2020-05-27 09:15:11 +0200 | [diff] [blame] | 236 | target_mem_type_set, |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 237 | ignore_subgraph_input_output_tensors=False, |
| 238 | lr_graph=None, |
Tim Hall | b9b515c | 2020-11-01 21:27:19 +0000 | [diff] [blame] | 239 | cpu_tensor_alignment=Tensor.AllocationQuantum, |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 240 | ): |
Diego Russo | ea6111a | 2020-04-14 18:41:58 +0100 | [diff] [blame] | 241 | if lr_graph is None: |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 242 | lr_graph = LiveRangeGraph() |
| 243 | |
| 244 | if sg in lr_graph.processed_subgraphs: |
| 245 | # if subgraph has been processed already, return the lr_graph as is |
| 246 | return lr_graph |
| 247 | |
| 248 | if ignore_subgraph_input_output_tensors: |
| 249 | lr_graph.ignore_tensors.update(sg.input_tensors) |
| 250 | lr_graph.ignore_tensors.update(sg.output_tensors) |
| 251 | |
Patrik Gustavsson | a151f59 | 2020-10-16 13:59:52 +0200 | [diff] [blame] | 252 | # Try to merge live ranges of operations in the NPU subgraphs |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 253 | if sg.placement == PassPlacement.Npu: |
Patrik Gustavsson | a151f59 | 2020-10-16 13:59:52 +0200 | [diff] [blame] | 254 | 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] | 255 | |
| 256 | for cps in sg.cascaded_passes: |
| 257 | cps.time = lr_graph.current_time |
| 258 | |
| 259 | time_for_pass = cps.time |
| 260 | |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 261 | for tens in cps.inputs: |
Patrik Gustavsson | a151f59 | 2020-10-16 13:59:52 +0200 | [diff] [blame] | 262 | 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] | 263 | continue |
Tim Hall | b9b515c | 2020-11-01 21:27:19 +0000 | [diff] [blame] | 264 | rng = lr_graph.get_or_create_range(tens, cpu_tensor_alignment) |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 265 | rng.mark_usage(time_for_pass) |
| 266 | |
| 267 | cps_primary_op = cps.passes[0].primary_op |
Patrik Gustavsson | eca2e95 | 2020-05-27 09:15:11 +0200 | [diff] [blame] | 268 | |
Louis Verhaard | aee5d75 | 2020-09-30 09:01:52 +0200 | [diff] [blame] | 269 | if ( |
| 270 | cps_primary_op |
| 271 | and cps_primary_op.type == Op.CustomNpuOp |
| 272 | and MemType.Permanent_CPU not in target_mem_type_set |
| 273 | ): |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 274 | # If the primary-op is an NpuOp that means this is where an Npu subgraph |
| 275 | # is called. Go into said subgraph and extract live ranges before continuing. |
Jacob Bohlin | 0628a8c | 2020-08-28 13:25:14 +0200 | [diff] [blame] | 276 | # Use default allocation alignment of 16 for Npu tensors |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 277 | npu_sg = cps_primary_op.attrs["subgraph"] |
| 278 | lr_graph = extract_live_ranges_from_cascaded_passes( |
Patrik Gustavsson | fad90c2 | 2020-11-03 13:07:40 +0100 | [diff] [blame] | 279 | npu_sg, target_mem_area, target_mem_type_set, False, lr_graph, |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 280 | ) |
| 281 | # Set the new time after handling the Npu subgraph |
| 282 | time_for_pass = lr_graph.current_time |
| 283 | cps.time = time_for_pass |
| 284 | |
Patrik Gustavsson | a151f59 | 2020-10-16 13:59:52 +0200 | [diff] [blame] | 285 | for tens in cps.intermediates + cps.outputs: |
| 286 | 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] | 287 | continue |
Tim Hall | b9b515c | 2020-11-01 21:27:19 +0000 | [diff] [blame] | 288 | rng = lr_graph.get_or_create_range(tens, cpu_tensor_alignment) |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 289 | rng.mark_usage(time_for_pass) |
| 290 | |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 291 | lr_graph.current_time += 2 |
| 292 | |
| 293 | end_time = 0 |
| 294 | for rng in lr_graph.ranges.values(): |
| 295 | # Find the maximum end time of all live-ranges in the graph |
| 296 | end_time = max(end_time, rng.end_time) |
| 297 | |
| 298 | for tens in sg.output_tensors: |
Patrik Gustavsson | a151f59 | 2020-10-16 13:59:52 +0200 | [diff] [blame] | 299 | 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] | 300 | continue |
Tim Hall | b9b515c | 2020-11-01 21:27:19 +0000 | [diff] [blame] | 301 | rng = lr_graph.get_or_create_range(tens, cpu_tensor_alignment) |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 302 | rng.mark_usage(end_time) |
| 303 | |
| 304 | # Add subgraph to set of processed subgraphs |
| 305 | lr_graph.processed_subgraphs.add(sg) |
| 306 | return lr_graph |