Louis Verhaard | d700252 | 2021-01-20 17:23:54 +0100 | [diff] [blame] | 1 | # Copyright (C) 2020-2021 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 | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 16 | # Description: |
| 17 | # Wrapping function to do tensor address allocation. That is, assigning addresses to tensors based on what has been |
| 18 | # worked out from the allowable overlaps that are calculated by the live range analysis. |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 19 | import math |
Louis Verhaard | 226ecaf | 2021-03-30 10:18:28 +0200 | [diff] [blame] | 20 | from typing import List |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 21 | |
Diego Russo | ea6111a | 2020-04-14 18:41:58 +0100 | [diff] [blame] | 22 | import numpy as np |
| 23 | |
Louis Verhaard | d700252 | 2021-01-20 17:23:54 +0100 | [diff] [blame] | 24 | from . import hillclimb_allocation |
Diego Russo | ea6111a | 2020-04-14 18:41:58 +0100 | [diff] [blame] | 25 | from . import live_range |
| 26 | from . import numeric_util |
Jacob Bohlin | 0628a8c | 2020-08-28 13:25:14 +0200 | [diff] [blame] | 27 | from .errors import AllocationError |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 28 | from .greedy_allocation import allocate_live_ranges as greedy_allocate_live_ranges |
Louis Verhaard | 226ecaf | 2021-03-30 10:18:28 +0200 | [diff] [blame] | 29 | from .live_range import LiveRange |
Louis Verhaard | 9bfe0f8 | 2020-12-03 12:26:25 +0100 | [diff] [blame] | 30 | from .live_range import LiveRangeGraph |
Diego Russo | e8a1045 | 2020-04-21 17:39:10 +0100 | [diff] [blame] | 31 | from .nn_graph import TensorAllocator |
| 32 | from .tensor import MemArea |
Patrik Gustavsson | eca2e95 | 2020-05-27 09:15:11 +0200 | [diff] [blame] | 33 | from .tensor import MemType |
Jacob Bohlin | 0628a8c | 2020-08-28 13:25:14 +0200 | [diff] [blame] | 34 | from .tensor import Tensor |
Louis Verhaard | 0b8268a | 2020-08-05 16:11:29 +0200 | [diff] [blame] | 35 | from .tensor import TensorPurpose |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 36 | |
| 37 | |
Jacob Bohlin | 0628a8c | 2020-08-28 13:25:14 +0200 | [diff] [blame] | 38 | def linear_allocate_live_ranges(live_ranges, alloc_granularity=Tensor.AllocationQuantum): |
Louis Verhaard | 3c07c97 | 2020-05-07 08:12:58 +0200 | [diff] [blame] | 39 | # Allocates using increasing addresses. Duplicate constant tensors will be allocated to the same address |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 40 | total_sz = 0 |
| 41 | allocated_tensors = [] |
| 42 | |
Louis Verhaard | 3c07c97 | 2020-05-07 08:12:58 +0200 | [diff] [blame] | 43 | # just assign increasing addresses, except for duplicates |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 44 | for tens, lr in live_ranges.ranges.items(): |
| 45 | if tens in allocated_tensors: |
| 46 | continue |
| 47 | |
Louis Verhaard | 3c07c97 | 2020-05-07 08:12:58 +0200 | [diff] [blame] | 48 | address = total_sz |
| 49 | if tens.weight_compression_config is not None: |
| 50 | for allocated_tens in allocated_tensors: |
| 51 | if allocated_tens.weight_compression_config == tens.weight_compression_config: |
| 52 | address = allocated_tens.address |
| 53 | break |
Louis Verhaard | 0b8268a | 2020-08-05 16:11:29 +0200 | [diff] [blame] | 54 | if tens.purpose == TensorPurpose.LUT: |
| 55 | for allocated_tens in allocated_tensors: |
| 56 | if allocated_tens.equivalent(tens): |
| 57 | address = allocated_tens.address |
| 58 | break |
Louis Verhaard | 3c07c97 | 2020-05-07 08:12:58 +0200 | [diff] [blame] | 59 | lr.set_address(address) |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 60 | allocated_tensors += lr.tensors |
Louis Verhaard | 3c07c97 | 2020-05-07 08:12:58 +0200 | [diff] [blame] | 61 | if address == total_sz: |
| 62 | total_sz += numeric_util.round_up(int(math.ceil(lr.size)), alloc_granularity) |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 63 | |
Jacob Bohlin | 0628a8c | 2020-08-28 13:25:14 +0200 | [diff] [blame] | 64 | verify_alignment(live_ranges, alloc_granularity) |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 65 | return total_sz |
| 66 | |
| 67 | |
Louis Verhaard | d700252 | 2021-01-20 17:23:54 +0100 | [diff] [blame] | 68 | def hillclimb_allocate_live_ranges(live_ranges: LiveRangeGraph, alloc_granularity: int) -> int: |
| 69 | # Allocates using the hill climb allocator |
Louis Verhaard | 226ecaf | 2021-03-30 10:18:28 +0200 | [diff] [blame] | 70 | addresses = hillclimb_allocation.allocate_live_ranges(live_ranges.lrs) |
Louis Verhaard | 9bfe0f8 | 2020-12-03 12:26:25 +0100 | [diff] [blame] | 71 | # The result is a list containing the allocated addresses |
| 72 | total_sz = 0 |
Louis Verhaard | 226ecaf | 2021-03-30 10:18:28 +0200 | [diff] [blame] | 73 | for lr, address in zip(live_ranges.lrs, addresses): |
Louis Verhaard | 9bfe0f8 | 2020-12-03 12:26:25 +0100 | [diff] [blame] | 74 | total_sz = max(total_sz, address + lr.size) |
| 75 | lr.set_address(address) |
| 76 | verify_allocation(live_ranges, alloc_granularity) |
| 77 | return total_sz |
| 78 | |
| 79 | |
| 80 | def verify_alignment(live_ranges: LiveRangeGraph, alignment: int): |
Louis Verhaard | 226ecaf | 2021-03-30 10:18:28 +0200 | [diff] [blame] | 81 | for lr in live_ranges.lrs: |
Jacob Bohlin | 0628a8c | 2020-08-28 13:25:14 +0200 | [diff] [blame] | 82 | for tens in lr.tensors: |
| 83 | if not all(op and op.run_on_npu for op in tens.ops + tens.consumer_list): |
| 84 | # This is a CPU tensor, verify alignment |
| 85 | if tens.address % alignment != 0: |
Michael McGeagh | 7a6f843 | 2020-12-02 15:29:22 +0000 | [diff] [blame] | 86 | raise AllocationError(f"Tensor '{tens.name}' not aligned to {alignment} bytes") |
Jacob Bohlin | 0628a8c | 2020-08-28 13:25:14 +0200 | [diff] [blame] | 87 | |
| 88 | |
Louis Verhaard | 9bfe0f8 | 2020-12-03 12:26:25 +0100 | [diff] [blame] | 89 | def verify_allocation(live_ranges: LiveRangeGraph, alignment: int): |
Louis Verhaard | 226ecaf | 2021-03-30 10:18:28 +0200 | [diff] [blame] | 90 | verify_alignment(live_ranges, alignment) |
| 91 | nr_time_slots = 1 + max(lr.end_time for lr in live_ranges.lrs) |
| 92 | # Contains active live ranges at each timestamp |
| 93 | lrs_at_time = [[] for i in range(nr_time_slots)] |
| 94 | for lr in live_ranges.lrs: |
| 95 | for t in range(lr.start_time, lr.end_time + 1): |
| 96 | lrs_at_time[t].append(lr) |
| 97 | for t in range(nr_time_slots): |
| 98 | for ix, n in enumerate(lrs_at_time[t]): |
| 99 | for m in lrs_at_time[t][ix + 1 :]: |
Louis Verhaard | 9bfe0f8 | 2020-12-03 12:26:25 +0100 | [diff] [blame] | 100 | overlap, tens_n, tens_m = n.overlaps_address(m) |
| 101 | if overlap and not (tens_n.equivalent(tens_m) and tens_n.address == tens_m.address): |
| 102 | raise AllocationError( |
Michael McGeagh | 7a6f843 | 2020-12-02 15:29:22 +0000 | [diff] [blame] | 103 | f"Overlapping buffers: {n.name}: {tens_n.address} -> {tens_n.address + n.size}" |
| 104 | f" and {m.name}: {tens_m.address} -> {tens_m.address + m.size}" |
Louis Verhaard | 9bfe0f8 | 2020-12-03 12:26:25 +0100 | [diff] [blame] | 105 | ) |
| 106 | |
| 107 | |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 108 | def mark_sram_used_for_cascaded_passes(sg, lrs): |
| 109 | end_pos = max(ps.time for ps in sg.cascaded_passes) + 2 |
| 110 | mem_usage = np.zeros(end_pos, dtype=np.int64) |
| 111 | |
| 112 | for tens, rng in lrs.ranges.items(): |
| 113 | storage_size = tens.storage_size() |
| 114 | mem_usage[rng.start_time : rng.end_time] += storage_size |
| 115 | |
| 116 | for cps in sg.cascaded_passes: |
| 117 | sram_used = max(mem_usage[cps.time], mem_usage[cps.time + 1]) |
| 118 | cps.sram_used = sram_used |
| 119 | for ps in cps.passes: |
| 120 | ps.sram_used = sram_used |
| 121 | |
| 122 | |
Tim Hall | 64556f3 | 2021-05-17 22:57:46 +0100 | [diff] [blame^] | 123 | def print_allocation(lrs, mem_area, mem_type_set, tensor_allocator, sg, actual_mem_usage_for_alloc): |
| 124 | print("\n" + "#" * 80) |
| 125 | sg_placement = ( |
| 126 | sg.placement.name |
| 127 | if mem_type_set.intersection((MemType.Permanent_NPU, MemType.Permanent_CPU,)) |
| 128 | else "Cpu and Npu" |
| 129 | ) |
| 130 | print( |
| 131 | f"Tensor Allocation for mem_area {mem_area.name}, of mem_type_set (" |
| 132 | f'{", ".join(f"{mem_type.name}" for mem_type in mem_type_set)}' |
| 133 | f"), using allocator {tensor_allocator}, in {sg_placement} subgraph:" |
| 134 | ) |
| 135 | |
| 136 | memory_hist = memory_usage_histogram(lrs.lrs) |
| 137 | min_mem_usage_for_alloc = max(memory_hist) |
| 138 | print("Start Time - End Time: Start Addr - End Addr: Tensor Size: Memory Usage: Tensor Purpose: Tensor Name") |
| 139 | for start_time, end_time, size, start_addr, end_addr, purpose, name in sorted( |
| 140 | (lr.start_time, lr.end_time, lr.size, tens.address, tens.address + lr.size, tens.purpose, tens.name,) |
| 141 | for tens, lr in lrs.ranges.items() |
| 142 | ): |
| 143 | print( |
| 144 | f"{start_time:10d} - {end_time:10d}: {start_addr:#10x} - {end_addr:#10x}: {size:11d}:" |
| 145 | f" {memory_hist[start_time]:12d}: {purpose.display_name():15s}: {name:s}" |
| 146 | ) |
| 147 | |
| 148 | alloc_overhead_fraction = (actual_mem_usage_for_alloc - min_mem_usage_for_alloc) / min_mem_usage_for_alloc |
| 149 | print( |
| 150 | f"Allocation Peak Tensor Size: {min_mem_usage_for_alloc:9d} ({min_mem_usage_for_alloc:#10x})" |
| 151 | f" Bytes {min_mem_usage_for_alloc/1024.0:8.2f} KiB" |
| 152 | ) |
| 153 | print( |
| 154 | f"Allocation Peak Memory Usage: {actual_mem_usage_for_alloc:9d} ({actual_mem_usage_for_alloc:#10x})" |
| 155 | f" Bytes {actual_mem_usage_for_alloc/1024.0:8.2f} KiB" |
| 156 | ) |
| 157 | print( |
| 158 | f"Allocation Overhead: {actual_mem_usage_for_alloc-min_mem_usage_for_alloc:9d}" |
| 159 | f" Bytes ({100*alloc_overhead_fraction:.2f} %)" |
| 160 | ) |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 161 | |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 162 | |
Tim Hall | 64556f3 | 2021-05-17 22:57:46 +0100 | [diff] [blame^] | 163 | def memory_usage_histogram(lrs: List[LiveRange]): |
| 164 | histogram = [0] * (1 + max(lr.end_time for lr in lrs)) |
Louis Verhaard | 226ecaf | 2021-03-30 10:18:28 +0200 | [diff] [blame] | 165 | for lr in lrs: |
erik.andersson@arm.com | 3438c92 | 2021-03-24 10:32:09 +0100 | [diff] [blame] | 166 | for t in range(lr.start_time, lr.end_time + 1): |
Tim Hall | 64556f3 | 2021-05-17 22:57:46 +0100 | [diff] [blame^] | 167 | histogram[t] += lr.size |
erik.andersson@arm.com | 3438c92 | 2021-03-24 10:32:09 +0100 | [diff] [blame] | 168 | |
Tim Hall | 64556f3 | 2021-05-17 22:57:46 +0100 | [diff] [blame^] | 169 | return histogram |
erik.andersson@arm.com | 3438c92 | 2021-03-24 10:32:09 +0100 | [diff] [blame] | 170 | |
| 171 | |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 172 | def allocate_tensors( |
| 173 | nng, |
| 174 | sg, |
| 175 | arch, |
| 176 | mem_area, |
Patrik Gustavsson | eca2e95 | 2020-05-27 09:15:11 +0200 | [diff] [blame] | 177 | mem_type_set, |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 178 | tensor_allocator=TensorAllocator.Greedy, |
| 179 | verbose_allocation=False, |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 180 | lr_graph=None, |
Tim Hall | b9b515c | 2020-11-01 21:27:19 +0000 | [diff] [blame] | 181 | cpu_tensor_alignment=Tensor.AllocationQuantum, |
Louis Verhaard | 0b9c9a3 | 2020-09-15 14:05:38 +0200 | [diff] [blame] | 182 | max_size=None, |
| 183 | dry_test=False, |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 184 | ): |
Louis Verhaard | 0b9c9a3 | 2020-09-15 14:05:38 +0200 | [diff] [blame] | 185 | # Allocates addresses to tensors, returns False if tensors could not be fit within max_size |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 186 | ignore_subgraph_input_output_tensors = False |
| 187 | lrs = live_range.extract_live_ranges_from_cascaded_passes( |
| 188 | sg, |
| 189 | mem_area, |
Patrik Gustavsson | eca2e95 | 2020-05-27 09:15:11 +0200 | [diff] [blame] | 190 | mem_type_set, |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 191 | ignore_subgraph_input_output_tensors=ignore_subgraph_input_output_tensors, |
| 192 | lr_graph=lr_graph, |
Tim Hall | b9b515c | 2020-11-01 21:27:19 +0000 | [diff] [blame] | 193 | cpu_tensor_alignment=cpu_tensor_alignment, |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 194 | ) |
| 195 | |
| 196 | if lrs.ranges: |
| 197 | tens_alloc = tensor_allocator |
| 198 | if tens_alloc == TensorAllocator.Greedy: |
Tim Hall | 64556f3 | 2021-05-17 22:57:46 +0100 | [diff] [blame^] | 199 | total_sz = greedy_allocate_live_ranges(sg, arch, lrs, mem_area, cpu_tensor_alignment) |
Louis Verhaard | 9bfe0f8 | 2020-12-03 12:26:25 +0100 | [diff] [blame] | 200 | verify_allocation(lrs, cpu_tensor_alignment) |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 201 | elif tens_alloc == TensorAllocator.LinearAlloc: |
Tim Hall | b9b515c | 2020-11-01 21:27:19 +0000 | [diff] [blame] | 202 | total_sz = linear_allocate_live_ranges(lrs, cpu_tensor_alignment) |
Louis Verhaard | d700252 | 2021-01-20 17:23:54 +0100 | [diff] [blame] | 203 | elif tens_alloc == TensorAllocator.HillClimb: |
| 204 | total_sz = hillclimb_allocate_live_ranges(lrs, cpu_tensor_alignment) |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 205 | else: |
| 206 | assert 0 |
Louis Verhaard | 0b9c9a3 | 2020-09-15 14:05:38 +0200 | [diff] [blame] | 207 | alloc_ok = max_size is None or total_sz <= max_size |
| 208 | if dry_test or not alloc_ok: |
| 209 | # Dry test or allocation failed; undo allocation |
| 210 | for lr in lrs.ranges.values(): |
| 211 | lr.set_address(None) |
| 212 | return alloc_ok |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 213 | |
Patrik Gustavsson | eca2e95 | 2020-05-27 09:15:11 +0200 | [diff] [blame] | 214 | if sg.memory_used.get(mem_area, 0) == 0: |
| 215 | sg.memory_used[mem_area] = total_sz |
| 216 | else: |
| 217 | sg.memory_used[mem_area] += total_sz |
| 218 | |
| 219 | # Keep track of how much should be used for scratch or permanent storage for NPU |
| 220 | for mem_type in mem_type_set: |
| 221 | if sg.memory_used_per_type.get(mem_type, 0) == 0: |
| 222 | sg.memory_used_per_type[mem_type] = total_sz |
| 223 | else: |
| 224 | sg.memory_used_per_type[mem_type] += total_sz |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 225 | |
Tim Hall | 64556f3 | 2021-05-17 22:57:46 +0100 | [diff] [blame^] | 226 | if verbose_allocation: |
| 227 | print_allocation(lrs, mem_area, mem_type_set, tensor_allocator, sg, total_sz) |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 228 | |
| 229 | if mem_area == MemArea.Sram: |
| 230 | # Mark Sram usage for all subgraphs |
| 231 | for sg_ in nng.subgraphs: |
| 232 | mark_sram_used_for_cascaded_passes(sg_, lrs) |
| 233 | |
| 234 | if sg == nng.get_root_subgraph(): |
| 235 | nng.memory_used = sg.memory_used |
Diqing Zhong | db5124c | 2021-01-11 12:52:48 +0100 | [diff] [blame] | 236 | |
Louis Verhaard | 0b9c9a3 | 2020-09-15 14:05:38 +0200 | [diff] [blame] | 237 | return True |