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. |
| 16 | |
| 17 | |
| 18 | # Description: |
| 19 | # Contains the main sequencing of the compiler. |
| 20 | |
| 21 | from . import graph_optimiser |
| 22 | from . import mark_tensors |
| 23 | from . import insert_dma |
| 24 | from . import pass_packing |
| 25 | from . import scheduler |
| 26 | from . import tensor_allocation |
| 27 | from . import npu_performance |
| 28 | import time |
| 29 | |
| 30 | from . import high_level_command_stream |
| 31 | from . import high_level_command_stream_generator |
| 32 | from . import register_command_stream_generator |
| 33 | from . import extract_npu_subgraphs |
| 34 | from . import npu_serialisation |
| 35 | from . import weight_compressor |
| 36 | from . import live_range |
| 37 | from .tensor import MemArea |
| 38 | from .nn_graph import TensorAllocator, PassPlacement |
| 39 | from .rewrite_graph import verify_graph_health, verify_subgraph_health |
| 40 | |
| 41 | |
| 42 | class CompilerOptions: |
| 43 | """Set of options to change compiler behaviour - verbosity, targets, turning off passes. |
| 44 | |
| 45 | Note the difference between ArchitectureFeatures and CompilerOptions |
| 46 | - ArchitectureFeatures is for changing the Ethos-U55 and system architecture |
| 47 | - CompilerOptions is for changing the behaviour of the compiler |
| 48 | """ |
| 49 | |
| 50 | def __init__( |
| 51 | self, |
| 52 | verbose_graph=False, |
| 53 | verbose_quantization=False, |
| 54 | verbose_packing=False, |
| 55 | verbose_tensor_purpose=False, |
| 56 | verbose_tensor_format=False, |
| 57 | verbose_allocation=False, |
| 58 | verbose_high_level_command_stream=False, |
| 59 | verbose_register_command_stream=False, |
| 60 | verbose_operators=False, |
| 61 | show_minimum_possible_allocation=False, |
| 62 | show_cpu_operations=False, |
| 63 | tensor_allocator=TensorAllocator.Greedy, |
| 64 | timing=False, |
| 65 | output_dir="outputs", |
| 66 | ): |
| 67 | |
| 68 | self.verbose_graph = verbose_graph |
| 69 | self.verbose_quantization = verbose_quantization |
| 70 | self.verbose_packing = verbose_packing |
| 71 | self.verbose_tensor_purpose = verbose_tensor_purpose |
| 72 | self.verbose_tensor_format = verbose_tensor_format |
| 73 | self.verbose_allocation = verbose_allocation |
| 74 | self.verbose_high_level_command_stream = verbose_high_level_command_stream |
| 75 | self.verbose_register_command_stream = verbose_register_command_stream |
| 76 | self.verbose_operators = verbose_operators |
| 77 | self.show_minimum_possible_allocation = show_minimum_possible_allocation |
| 78 | self.show_cpu_operations = show_cpu_operations |
| 79 | self.tensor_allocator = tensor_allocator |
| 80 | self.timing = timing |
| 81 | self.output_dir = output_dir |
| 82 | |
| 83 | def __str__(self): |
| 84 | return type(self).__name__ + ": " + str(self.__dict__) |
| 85 | |
| 86 | __repr__ = __str__ |
| 87 | |
| 88 | |
| 89 | def compiler_driver(nng, arch, options, scheduler_options): |
| 90 | assert verify_graph_health(nng) |
| 91 | nng = graph_optimiser.optimise_graph_a(nng, arch, options.verbose_graph) |
| 92 | assert verify_graph_health(nng) |
| 93 | |
| 94 | if options.verbose_quantization: |
| 95 | nng.print_graph_with_tensor_quantization() |
| 96 | |
| 97 | nng = graph_optimiser.optimise_graph_b(nng, arch, options.verbose_graph) |
| 98 | assert verify_graph_health(nng) |
| 99 | |
| 100 | nng = mark_tensors.mark_tensor_purpose(nng, arch, options.verbose_tensor_purpose) |
| 101 | assert verify_graph_health(nng) |
| 102 | nng = insert_dma.insert_dma_commands(nng, arch, options.verbose_graph) |
| 103 | assert verify_graph_health(nng) |
| 104 | pass_packing.pack_into_passes(nng, arch, options.verbose_packing) |
| 105 | assert verify_graph_health(nng) |
| 106 | |
| 107 | extract_npu_subgraphs.extract_npu_subgraphs(nng, arch) |
| 108 | |
| 109 | mark_tensors.mark_tensor_format(nng, arch, options.verbose_tensor_format) |
| 110 | assert verify_graph_health(nng) |
| 111 | if options.timing: |
| 112 | start = time.time() |
| 113 | |
| 114 | # Run the scheduler |
| 115 | scheduler.schedule_passes(nng, arch, scheduler_options) |
| 116 | |
| 117 | if options.timing: |
| 118 | stop = time.time() |
| 119 | print("Scheduling took %f s" % (stop - start)) |
| 120 | start = time.time() |
| 121 | |
| 122 | # Update the compressed weights now that we have determined the |
| 123 | # block config, and calc and pack the scales and biases |
| 124 | weight_compressor.update_pass_weight_and_scale_tensors(nng, arch) |
| 125 | |
| 126 | # Memory area for all non-constant tensors (Cpu and Npu) |
| 127 | non_const_mem_area = MemArea.Sram |
| 128 | |
| 129 | # LiveRanges for constant tensors for all Npu subgraphs |
| 130 | permanent_storage = arch.permanent_storage_mem_area |
| 131 | lr_graph_flash = live_range.LiveRangeGraph() |
| 132 | |
| 133 | # Placeholders for scratch and flash tensors that are common for all Npu subgraphs |
| 134 | scratch_tens = None |
| 135 | flash_tens = None |
| 136 | |
| 137 | # Calculate live ranges for all constant Npu tensors, in permanent storage |
| 138 | for sg in nng.subgraphs: |
| 139 | if sg.placement == PassPlacement.Npu: |
| 140 | lr_graph_flash = live_range.extract_live_ranges_from_cascaded_passes( |
| 141 | sg, permanent_storage, ignore_subgraph_input_output_tensors=True, lr_graph=lr_graph_flash |
| 142 | ) |
| 143 | |
| 144 | # Allocate all Npu constant tensors to the first Npu subgraph since it is |
| 145 | # processed first during serialization into tensors |
| 146 | first_npu_sg = nng.subgraphs[1] |
| 147 | assert first_npu_sg.placement == PassPlacement.Npu |
| 148 | tensor_allocation.allocate_tensors( |
| 149 | nng, |
| 150 | first_npu_sg, |
| 151 | arch, |
| 152 | permanent_storage, |
| 153 | scheduler_options.use_ifm_ofm_overlap, |
| 154 | options.tensor_allocator, |
| 155 | options.verbose_allocation, |
| 156 | options.show_minimum_possible_allocation, |
| 157 | lr_graph_flash, |
| 158 | ) |
| 159 | |
| 160 | # Allocate all non-constant tensors to the root, i.e. Cpu, subgraph. This step |
| 161 | # will start at the root subgraph's input and traverse from top to bottom. When |
| 162 | # it comes across an Npu-op it will extract live ranges for it's corresponding |
| 163 | # Npu subgraph and add them to the root's live range graph. Finally, all of the |
| 164 | # non-constant tensors are allocated together |
| 165 | root_sg = nng.get_root_subgraph() |
| 166 | tensor_allocation.allocate_tensors( |
| 167 | nng, |
| 168 | root_sg, |
| 169 | arch, |
| 170 | non_const_mem_area, |
| 171 | scheduler_options.use_ifm_ofm_overlap, |
| 172 | options.tensor_allocator, |
| 173 | options.verbose_allocation, |
| 174 | options.show_minimum_possible_allocation, |
| 175 | ) |
| 176 | |
| 177 | # Generate command streams and serialise Npu-ops into tensors |
| 178 | for sg in nng.subgraphs: |
| 179 | high_level_command_stream_generator.generate_high_level_command_stream( |
| 180 | nng, sg, arch, options.verbose_high_level_command_stream |
| 181 | ) |
| 182 | register_command_stream_generator.generate_register_command_stream( |
| 183 | nng, sg, arch, options.verbose_register_command_stream |
| 184 | ) |
| 185 | scratch_tens, flash_tens = npu_serialisation.serialise_npu_subgraph_into_tensors( |
| 186 | nng, sg, arch, scratch_tens, flash_tens |
| 187 | ) |
| 188 | |
| 189 | npu_serialisation.rewrite_npu_call_ops(nng, root_sg, arch) |
| 190 | |
| 191 | # Allocate all Cpu constant tensors, this is done last because the Npu-ops |
| 192 | # have to be serialized into flash and scratch tensors first |
| 193 | tensor_allocation.allocate_tensors( |
| 194 | nng, |
| 195 | root_sg, |
| 196 | arch, |
| 197 | permanent_storage, |
| 198 | scheduler_options.use_ifm_ofm_overlap, |
| 199 | options.tensor_allocator, |
| 200 | options.verbose_allocation, |
| 201 | options.show_minimum_possible_allocation, |
| 202 | ) |
| 203 | |
| 204 | npu_performance.calc_performance_for_network(nng, arch) |