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 | # Generate a high-level command stream from a scheduled subgraph with CascadedPasses. |
| 20 | # |
| 21 | # Also used during scheduling to work out allowable IFM/OFM overlap, this functionality can be accessed using |
| 22 | # calc_allowed_ofm_ifm_overlap_for_cascaded_pass(). |
| 23 | |
| 24 | from .nn_graph import SchedulingStrategy, PassPlacement |
| 25 | import numpy as np |
| 26 | from .operation import NpuBlockType |
| 27 | from .high_level_command_stream import Box, CommandType, Command, NpuStripe, DMA |
| 28 | |
| 29 | |
| 30 | def need_dma(tens): |
| 31 | return len(tens.ops) == 1 and tens.ops[0].type == "DMA" |
| 32 | |
| 33 | |
| 34 | def dma_weights_if_necessary(ps, box, weight_tensor): |
| 35 | if need_dma(weight_tensor): |
| 36 | dma_op = weight_tensor.ops[0] |
| 37 | in_tensor = dma_op.inputs[0] |
| 38 | yield DMA(in_tensor, weight_tensor, box) |
| 39 | |
| 40 | |
| 41 | def generate_high_level_command_stream_for_pass(strat, passes, block_configs, idx): |
| 42 | is_first = idx == 0 |
| 43 | is_last = idx == len(passes) - 1 |
| 44 | ps = passes[idx] |
| 45 | block_config = block_configs[idx] |
| 46 | |
| 47 | ifm_tensor = ps.ifm_tensor |
| 48 | ifm2_tensor = ps.ifm2_tensor |
| 49 | ofm_tensor = ps.ofm_tensor |
| 50 | weight_tensor = ps.weight_tensor |
| 51 | scale_tensor = ps.scale_tensor |
| 52 | |
| 53 | ofm_start = [0] * len(ofm_tensor.shape) |
| 54 | ofm_end = list(ofm_tensor.shape) |
| 55 | |
| 56 | strides = None |
| 57 | skirt = None |
| 58 | if ps.primary_op is not None: |
| 59 | strides = ps.primary_op.attrs.get("strides", None) |
| 60 | skirt = ps.primary_op.attrs.get("skirt", None) |
| 61 | |
| 62 | npu_block_type = ps.npu_block_type |
| 63 | |
| 64 | concat_axis = 0 |
| 65 | concat_offset = 0 |
| 66 | |
| 67 | split_offsets = [None, None] # offset for [ifm, ifm2] |
| 68 | |
| 69 | # Fusable activation functions |
| 70 | activation_ops = set(("Sigmoid", "Tanh", "Relu", "Relu6", "ReluN1To1")) |
| 71 | |
| 72 | for op in ps.ops: |
| 73 | if op.type == "ConcatSliceWrite": |
| 74 | concat_axis = op.attrs["concat_axis"] |
| 75 | concat_start = op.attrs["concat_start"] |
| 76 | concat_end = op.attrs["concat_end"] |
| 77 | |
| 78 | ofm_start[concat_axis] = concat_start |
| 79 | ofm_end[concat_axis] = concat_end |
| 80 | concat_offset = concat_start |
| 81 | ps.primary_op.attrs["fused_memory_function"] = op.type |
| 82 | elif op.type in activation_ops: |
| 83 | ps.primary_op.attrs["fused_activation_function"] = op.type |
| 84 | |
| 85 | # The ops list has to be reversed here since the Pass Packing is done in reverse |
| 86 | ifm_idx = 0 |
| 87 | for op in reversed(ps.ops): |
| 88 | if op.type == "SplitSliceRead": |
| 89 | split_offsets[ifm_idx] = op.attrs["split_start"] |
| 90 | ps.primary_op.attrs["fused_memory_function"] = op.type |
| 91 | ifm_idx += 1 |
| 92 | |
| 93 | if strat == SchedulingStrategy.WeightStream: |
| 94 | ofm_step = block_config[-1] |
| 95 | ofm_stop = ofm_end[-1] |
| 96 | if weight_tensor is None or not need_dma(weight_tensor): |
| 97 | ofm_step = ofm_stop |
| 98 | for start in range(ofm_start[-1], ofm_stop, ofm_step): |
| 99 | end = min(start + ofm_step, ofm_stop) |
| 100 | ofm_start[-1] = start |
| 101 | ofm_end[-1] = end |
| 102 | ofm_box = Box(ofm_start, ofm_end) |
| 103 | ifm_box = None |
| 104 | ifm2_box = None |
| 105 | |
| 106 | if ifm_tensor.shape != []: |
| 107 | ifm_box, _, _ = ofm_box.transform_with_strides_and_skirt( |
| 108 | strides, skirt, ifm_tensor.shape, npu_block_type, concat_axis, concat_offset, split_offsets[0] |
| 109 | ) |
| 110 | else: |
| 111 | ifm_box = Box([], []) |
| 112 | if ifm2_tensor is not None and ifm2_tensor.shape != []: |
| 113 | ifm2_box, _, _ = ofm_box.transform_with_strides_and_skirt( |
| 114 | strides, skirt, ifm2_tensor.shape, npu_block_type, concat_axis, concat_offset, split_offsets[1] |
| 115 | ) |
| 116 | else: |
| 117 | ifm2_box = Box([], []) |
| 118 | |
| 119 | weight_box = None |
| 120 | if weight_tensor is not None: |
| 121 | weight_oc_start = start |
| 122 | weight_oc_end = end |
| 123 | if concat_axis - len(weight_tensor.shape) == -1: |
| 124 | weight_oc_start -= concat_offset |
| 125 | weight_oc_end -= concat_offset |
| 126 | |
| 127 | weight_box = Box.make_weight_box( |
| 128 | weight_tensor.shape, |
| 129 | npu_block_type, |
| 130 | weight_oc_start, |
| 131 | weight_oc_end, |
| 132 | weight_tensor.weight_transpose_depthwise, |
| 133 | ) |
| 134 | yield from dma_weights_if_necessary(ps, weight_box, weight_tensor) |
| 135 | |
| 136 | yield NpuStripe( |
| 137 | ps, |
| 138 | block_config, |
| 139 | is_first, |
| 140 | is_last, |
| 141 | True, |
| 142 | True, |
| 143 | ifm_tensor, |
| 144 | ifm_box, |
| 145 | ofm_tensor, |
| 146 | ofm_box, |
| 147 | weight_tensor, |
| 148 | weight_box, |
| 149 | scale_tensor, |
| 150 | concat_axis, |
| 151 | concat_offset, |
| 152 | ifm2_tensor=ifm2_tensor, |
| 153 | ifm2_box=ifm2_box, |
| 154 | ) |
| 155 | |
| 156 | elif strat == SchedulingStrategy.IfmStream: |
| 157 | y_step = block_config[0] |
| 158 | y_start = 0 |
| 159 | y_dim = 1 |
| 160 | if len(ofm_tensor.shape) >= 3: |
| 161 | y_start = ofm_start[-3] |
| 162 | y_dim = ofm_end[-3] |
| 163 | if idx > 0: |
| 164 | ifm_y_present = 0 |
| 165 | prev_pass = passes[idx - 1] |
| 166 | prev_pass_gen = generate_high_level_command_stream_for_pass(strat, passes, block_configs, idx - 1) |
| 167 | else: |
| 168 | ifm_y_present = 1 |
| 169 | if len(ifm_tensor.shape) >= 3: |
| 170 | ifm_y_present = ifm_tensor.shape[-3] |
| 171 | prev_pass_gen = [] |
| 172 | prev_pass = None |
| 173 | |
| 174 | if len(passes) == 1: |
| 175 | # no cascading, can just issue one big stripe |
| 176 | # but only if we've done allocation and OFM does not overlap IFM |
| 177 | if ifm_tensor.address != -1 and ofm_tensor.address != -1: |
| 178 | if ( |
| 179 | ifm_tensor.address + ifm_tensor.storage_size() <= ofm_tensor.address |
| 180 | or ofm_tensor.address + ofm_tensor.storage_size() <= ifm_tensor.address |
| 181 | ): |
| 182 | y_step = y_dim |
| 183 | |
| 184 | weight_box = None |
| 185 | |
| 186 | for start in range(y_start, y_dim, y_step): |
| 187 | end = min(start + y_step, y_dim) |
| 188 | if len(ofm_tensor.shape) >= 3: |
| 189 | ofm_start[-3] = start |
| 190 | ofm_end[-3] = end |
| 191 | ofm_box = Box(ofm_start, ofm_end) |
| 192 | |
| 193 | k_height = 1 |
| 194 | if npu_block_type == NpuBlockType.Pooling: |
| 195 | if ps.primary_op is not None: |
| 196 | k_height = ps.primary_op.attrs["ksize"][1] |
| 197 | else: |
| 198 | if weight_tensor is not None: |
| 199 | k_height = weight_tensor.shape[0] |
| 200 | |
| 201 | ifm_box, pad_top, pad_bottom = ofm_box.transform_with_strides_and_skirt( |
| 202 | strides, skirt, ifm_tensor.shape, npu_block_type, concat_axis, concat_offset, split_offsets[0], k_height |
| 203 | ) |
| 204 | |
| 205 | ifm_y_needed = 1 |
| 206 | if len(ifm_box.end_coord) >= 3: |
| 207 | ifm_y_needed = ifm_box.end_coord[-3] |
| 208 | if ifm_y_present < ifm_y_needed: |
| 209 | for prev_cmd in prev_pass_gen: |
| 210 | yield prev_cmd |
| 211 | rng = prev_cmd.get_ofm_y_range_for_pass(prev_pass) |
| 212 | if rng is not None: |
| 213 | ifm_y_present = max(ifm_y_present, rng[1]) |
| 214 | if ifm_y_present >= ifm_y_needed: |
| 215 | break |
| 216 | |
| 217 | if weight_tensor is not None and weight_box is None: |
| 218 | weight_box = Box.make_weight_box( |
| 219 | weight_tensor.shape, npu_block_type, weights_transposed=weight_tensor.weight_transpose_depthwise |
| 220 | ) |
| 221 | yield from dma_weights_if_necessary(ps, weight_box, weight_tensor) |
| 222 | |
| 223 | # Check if first/last stripe in pass |
| 224 | is_first_h_stripe = start == y_start |
| 225 | is_last_h_stripe = (start + y_step) >= y_dim |
| 226 | |
| 227 | stripe = NpuStripe( |
| 228 | ps, |
| 229 | block_config, |
| 230 | is_first, |
| 231 | is_last, |
| 232 | is_first_h_stripe, |
| 233 | is_last_h_stripe, |
| 234 | ifm_tensor, |
| 235 | ifm_box, |
| 236 | ofm_tensor, |
| 237 | ofm_box, |
| 238 | weight_tensor, |
| 239 | weight_box, |
| 240 | scale_tensor, |
| 241 | concat_axis, |
| 242 | concat_offset, |
| 243 | None, |
| 244 | None, |
| 245 | pad_top, |
| 246 | pad_bottom, |
| 247 | ) |
| 248 | yield stripe |
| 249 | else: |
| 250 | assert 0, "unknown scheduling strategy" |
| 251 | |
| 252 | |
| 253 | def generate_high_level_command_stream_for_pass_list(strat, passes, block_configs): |
| 254 | if strat == SchedulingStrategy.WeightStream: |
| 255 | for idx in range(len(passes)): |
| 256 | yield from generate_high_level_command_stream_for_pass(strat, passes, block_configs, idx) |
| 257 | elif strat == SchedulingStrategy.IfmStream: |
| 258 | yield from generate_high_level_command_stream_for_pass(strat, passes, block_configs, len(passes) - 1) |
| 259 | else: |
| 260 | assert 0, "Unknown streaming strategy" |
| 261 | |
| 262 | |
| 263 | def generate_high_level_command_stream_for_cascaded_pass(cps): |
| 264 | yield from generate_high_level_command_stream_for_pass_list( |
| 265 | cps.strategy, cps.passes, [ps.block_config for ps in cps.passes] |
| 266 | ) |
| 267 | |
| 268 | |
| 269 | def generate_high_level_command_stream(nng, sg, arch, verbose_high_level_command_stream): |
| 270 | res = [] |
| 271 | for cps in sg.cascaded_passes: |
| 272 | if cps.placement == PassPlacement.Npu: |
| 273 | res += list(generate_high_level_command_stream_for_cascaded_pass(cps)) |
| 274 | |
| 275 | sg.high_level_command_stream = res |
| 276 | if verbose_high_level_command_stream: |
| 277 | sg.print_high_level_command_stream() |
| 278 | |
| 279 | |
| 280 | def calc_allowed_ofm_ifm_overlap_for_pass_list(strat, passes, block_configs): |
| 281 | highest_ofm_write = 0 |
| 282 | if not passes[0].ifm_tensor or not passes[-1].ofm_tensor: |
| 283 | return 0 |
| 284 | |
| 285 | ifm_read = passes[0].ifm_tensor.storage_size |
| 286 | min_overlap = 999999999999999999999 |
| 287 | ofm_size = passes[-1].ofm_tensor.storage_size() |
| 288 | if strat == SchedulingStrategy.WeightStream: |
| 289 | return 0 |
| 290 | for cmd in generate_high_level_command_stream_for_pass_list(strat, passes, block_configs): |
| 291 | if cmd.is_npu_pass_command(): |
| 292 | if cmd.is_first: |
| 293 | ifm_read = cmd.ifm_tensor.address_offset_for_coordinate(cmd.ifm_box.start_coord, is_top_box=False) |
| 294 | if ifm_read is None: |
| 295 | return 0 |
| 296 | if cmd.is_last: |
| 297 | write_offset = cmd.ofm_tensor.address_offset_for_coordinate(cmd.ofm_box.end_coord, is_top_box=True) |
| 298 | if write_offset is None: |
| 299 | return 0 |
| 300 | highest_ofm_write = max(write_offset, highest_ofm_write) |
| 301 | |
| 302 | if cmd.is_first or cmd.is_last: |
| 303 | overlap_required = max(highest_ofm_write - min(ifm_read, ofm_size), 0) |
| 304 | can_overwrite = ofm_size - overlap_required |
| 305 | min_overlap = min(min_overlap, can_overwrite) |
| 306 | |
| 307 | if cmd.is_first: |
| 308 | ifm_read = cmd.ifm_tensor.address_offset_for_coordinate(cmd.ifm_box.end_coord, is_top_box=True) |
| 309 | |
| 310 | min_overlap = max(min_overlap, 0) |
| 311 | return min_overlap |
| 312 | |
| 313 | |
| 314 | def calc_allowed_ofm_ifm_overlap_for_cascaded_pass(cps): |
| 315 | return calc_allowed_ofm_ifm_overlap_for_pass_list(cps.strategy, cps.passes, [ps.block_config for ps in cps.passes]) |