Patrik Gustavsson | 8f1f9aa | 2021-06-28 07:41:58 +0200 | [diff] [blame] | 1 | # Copyright (C) 2021 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 | # Description: |
| 17 | # Early optimisation of the TOSA based network graph, using the rewrite_graph module to do the traversal of the graph. |
Patrik Gustavsson | f366fb1 | 2021-09-07 13:30:29 +0200 | [diff] [blame] | 18 | import numpy as np |
| 19 | |
Patrik Gustavsson | 8f1f9aa | 2021-06-28 07:41:58 +0200 | [diff] [blame] | 20 | from . import rewrite_graph |
| 21 | from .api import NpuRoundingMode |
| 22 | from .data_type import DataType |
| 23 | from .debug_database import DebugDatabase |
Jonas Ohlsson | 0957e3e | 2021-09-01 15:57:21 +0200 | [diff] [blame] | 24 | from .graph_optimiser_util import bypass_memory_only_ops |
Patrik Gustavsson | c74682c | 2021-08-17 14:26:38 +0200 | [diff] [blame] | 25 | from .graph_optimiser_util import calc_explicit_padding |
Patrik Gustavsson | df99510 | 2021-08-23 15:33:59 +0200 | [diff] [blame] | 26 | from .graph_optimiser_util import convert_depthwise_to_conv |
Patrik Gustavsson | f436ada | 2021-09-14 14:56:48 +0200 | [diff] [blame] | 27 | from .graph_optimiser_util import convert_to_lut |
Patrik Gustavsson | f1580f0 | 2021-09-01 12:43:02 +0200 | [diff] [blame] | 28 | from .graph_optimiser_util import move_splitsliceread_to_consumer |
Patrik Gustavsson | 8f1f9aa | 2021-06-28 07:41:58 +0200 | [diff] [blame] | 29 | from .graph_optimiser_util import needed_total_padding |
| 30 | from .graph_optimiser_util import set_ifm_ofm_op_shapes |
| 31 | from .graph_optimiser_util import set_tensor_equivalence |
| 32 | from .operation import ExplicitScaling |
Patrik Gustavsson | 8f1f9aa | 2021-06-28 07:41:58 +0200 | [diff] [blame] | 33 | from .operation import Op |
Patrik Gustavsson | f1580f0 | 2021-09-01 12:43:02 +0200 | [diff] [blame] | 34 | from .operation_util import create_add_nop |
Patrik Gustavsson | c74682c | 2021-08-17 14:26:38 +0200 | [diff] [blame] | 35 | from .operation_util import create_avgpool_nop |
Patrik Gustavsson | f1580f0 | 2021-09-01 12:43:02 +0200 | [diff] [blame] | 36 | from .shape4d import Shape4D |
| 37 | from .tensor import create_const_tensor |
Patrik Gustavsson | e2bfa7e | 2021-09-08 15:04:11 +0200 | [diff] [blame] | 38 | from .tensor import create_equivalence_id |
Patrik Gustavsson | 8f1f9aa | 2021-06-28 07:41:58 +0200 | [diff] [blame] | 39 | |
| 40 | |
Patrik Gustavsson | c74682c | 2021-08-17 14:26:38 +0200 | [diff] [blame] | 41 | def replace_rescale_with_avg_pool(rescale_op): |
| 42 | assert rescale_op.type == Op.Rescale |
| 43 | |
| 44 | avgpool_op = create_avgpool_nop(rescale_op.name + "_avgpool") |
| 45 | rescale_op_clone = rescale_op.clone() |
| 46 | op = rescale_op |
| 47 | op.attrs = avgpool_op.attrs.copy() |
| 48 | op.type = Op.AvgPool |
| 49 | DebugDatabase.add_optimised(rescale_op_clone, op) |
| 50 | |
| 51 | return op |
| 52 | |
| 53 | |
| 54 | def calc_skirt(kernel, input_shape, explicit_padding): |
Patrik Gustavsson | 8f1f9aa | 2021-06-28 07:41:58 +0200 | [diff] [blame] | 55 | k_w, k_h = kernel.dilated_wh() |
| 56 | s_x, s_y = kernel.stride |
| 57 | ypad = needed_total_padding(int(input_shape.height), int(s_y), int(k_h)) |
| 58 | xpad = needed_total_padding(int(input_shape.width), int(s_x), int(k_w)) |
Patrik Gustavsson | c74682c | 2021-08-17 14:26:38 +0200 | [diff] [blame] | 59 | |
| 60 | top, left, bottom, right = explicit_padding |
| 61 | top_pad, bottom_pad = calc_explicit_padding(int(input_shape.height), int(s_y), int(k_h), int(top), int(bottom)) |
| 62 | left_pad, right_pad = calc_explicit_padding(int(input_shape.width), int(s_x), int(k_w), int(left), int(right)) |
Patrik Gustavsson | 8f1f9aa | 2021-06-28 07:41:58 +0200 | [diff] [blame] | 63 | |
| 64 | padding = (top_pad, left_pad, bottom_pad, right_pad) |
| 65 | skirt = (top_pad, left_pad, ypad - top_pad, xpad - left_pad) |
| 66 | return padding, skirt |
| 67 | |
| 68 | |
| 69 | def add_padding_fields(op, arch, nng): |
| 70 | if op.run_on_npu: |
Patrik Gustavsson | c74682c | 2021-08-17 14:26:38 +0200 | [diff] [blame] | 71 | if "explicit_padding" in op.attrs: |
Patrik Gustavsson | 8f1f9aa | 2021-06-28 07:41:58 +0200 | [diff] [blame] | 72 | input_shape = op.ifm_shapes[0] |
| 73 | |
| 74 | if op.type == Op.Conv2DBackpropInputSwitchedBias: |
| 75 | # TODO not yet supported, but there will be need for separate handling |
| 76 | assert False |
| 77 | else: |
Patrik Gustavsson | c74682c | 2021-08-17 14:26:38 +0200 | [diff] [blame] | 78 | padding, skirt = calc_skirt(op.kernel, input_shape, op.attrs.get("explicit_padding")) |
Patrik Gustavsson | 8f1f9aa | 2021-06-28 07:41:58 +0200 | [diff] [blame] | 79 | |
| 80 | op.attrs["explicit_padding"] = padding |
| 81 | op.attrs["skirt"] = skirt |
| 82 | |
| 83 | return op |
| 84 | |
| 85 | |
Patrik Gustavsson | f366fb1 | 2021-09-07 13:30:29 +0200 | [diff] [blame] | 86 | # Counts leading zeroes for a (int32) |
| 87 | def count_leading_zeros(a): |
| 88 | lz = int(32) |
| 89 | if a != 0: |
| 90 | mask = 1 << (32 - 1) |
| 91 | lz = 0 |
| 92 | while (mask & a) == 0: |
| 93 | mask = mask >> 1 |
| 94 | lz = lz + 1 |
| 95 | return lz |
| 96 | |
| 97 | |
| 98 | def calc_scaling_avgpool(op, arch, nng): |
| 99 | if op.type == Op.AvgPool: |
| 100 | top, left, _, _ = op.attrs["explicit_padding"] |
| 101 | # TODO Only support for when global scaling can be used. |
| 102 | # That is when there is no padding |
| 103 | assert top == 0 and left == 0 |
| 104 | assert op.explicit_scaling is None |
| 105 | multiplier = [] |
| 106 | shift = [] |
| 107 | |
| 108 | kernel_wh = op.kernel.elements_wh() |
| 109 | k = 32 - count_leading_zeros(kernel_wh - 1) |
| 110 | numerator = np.int64(((1 << 30) + 1) << k) |
| 111 | multiplier.append(numerator // kernel_wh) |
| 112 | shift.append(30 + k) |
| 113 | |
| 114 | op.rounding_mode = NpuRoundingMode.NATURAL |
| 115 | op.explicit_scaling = ExplicitScaling(False, shift, multiplier) |
| 116 | return op |
| 117 | |
| 118 | |
Patrik Gustavsson | df99510 | 2021-08-23 15:33:59 +0200 | [diff] [blame] | 119 | def remove_const_transpose(op, arch, nng): |
| 120 | if op.type == Op.Transpose: |
| 121 | removed = False |
| 122 | if len(op.ifm.ops) == 1: |
| 123 | prev_op = op.ifm.ops[0] |
| 124 | if prev_op.type == Op.Const: |
| 125 | # Transpose the Tensor and data and remove Transpose |
| 126 | # TODO move to Tensor? |
| 127 | reorder = op.attrs["perms"] |
| 128 | shape = op.ifm.shape.copy() |
| 129 | tens = op.ifm |
| 130 | |
| 131 | tens.shape = [shape[idx] for idx in reorder] |
| 132 | tens.bandwidth_shape = tens.shape |
| 133 | tens.storage_shape = tens.shape |
| 134 | |
| 135 | if tens.values is not None: |
| 136 | tens.values = tens.values.transpose(reorder) |
| 137 | |
| 138 | op.ofm.values = tens.values |
| 139 | # Bypass the Transpose op |
| 140 | prev_op.set_output_tensor(op.ofm) |
| 141 | DebugDatabase.add_optimised(op, prev_op) |
| 142 | removed = True |
| 143 | |
| 144 | if not removed: |
Patrik Gustavsson | f1580f0 | 2021-09-01 12:43:02 +0200 | [diff] [blame] | 145 | print("Warning: Cannot remove Transpose, and handling of Transpose is not supported") |
Patrik Gustavsson | df99510 | 2021-08-23 15:33:59 +0200 | [diff] [blame] | 146 | assert False |
| 147 | |
| 148 | return op |
| 149 | |
| 150 | |
Patrik Gustavsson | f1580f0 | 2021-09-01 12:43:02 +0200 | [diff] [blame] | 151 | # TODO can we change to add for both TFLite and TOSA? |
| 152 | def insert_add_copy_op_after_tens(tens): |
| 153 | tens_cons_list_copy = tens.consumer_list.copy() |
| 154 | copy_tens = tens.clone() |
| 155 | |
| 156 | name = tens.name + "_add" |
| 157 | ifm2 = create_const_tensor( |
| 158 | name + "_zero_scalar", |
| 159 | [1], |
| 160 | copy_tens.dtype, |
| 161 | [0], |
| 162 | copy_tens.dtype.as_numpy_type(), |
| 163 | quantization=copy_tens.quantization, |
| 164 | ) |
| 165 | copy_op = create_add_nop(name) |
| 166 | copy_op.add_input_tensor(tens) |
| 167 | copy_op.add_input_tensor(ifm2) |
| 168 | copy_op.set_output_tensor(copy_tens) |
| 169 | copy_op.set_ifm_ofm_shapes() |
| 170 | copy_op.run_on_npu = True |
| 171 | |
| 172 | # Set copy_ifm consumers |
| 173 | for tens_cons in tens_cons_list_copy: |
| 174 | if tens_cons is not None: |
| 175 | for ifm_idx, cons_inp in enumerate(tens_cons.inputs): |
| 176 | if cons_inp == tens: |
| 177 | tens_cons.set_input_tensor(copy_tens, ifm_idx) |
| 178 | |
| 179 | DebugDatabase.add_optimised(tens.ops[0], copy_op) |
| 180 | |
| 181 | |
| 182 | def fix_sg_input_output_tosa(op, arch, nng): |
| 183 | if not op.run_on_npu or op.type != Op.Reshape: |
| 184 | return op |
| 185 | |
| 186 | # For the Reshape operators we want to remove, tensors are removed. |
| 187 | # But in order to to do this, they cannot be outputs of the sg, |
| 188 | # this need to be fixed prior to the removal. |
| 189 | # Solution is to add a copy op, to maintain the original tensor. |
| 190 | # This is also valid when reshape ifm/ofm is produced respectively |
| 191 | # consumed by CPU |
| 192 | |
| 193 | # Check if operator ifm/ofm are sg ifm/ofm |
| 194 | ifm_is_sg_ifm = op.ifm.ops[0].type in (Op.Placeholder, Op.SubgraphInput, Op.Const) |
| 195 | ifm_is_sg_ofm = any(ifm_cons is None for ifm_cons in op.ifm.consumer_list) |
| 196 | ofm_is_sg_ofm = any(ofm_cons is None for ofm_cons in op.ofm.consumer_list) |
| 197 | # Check if ifm/ofm is produced repectivly consumed by CPU |
| 198 | ifm_is_cpu_produced = any(ifm_prod is not None and not ifm_prod.run_on_npu for ifm_prod in op.ifm.ops) |
| 199 | ofm_is_cpu_consumed = any(ofm_cons is not None and not ofm_cons.run_on_npu for ofm_cons in op.ofm.consumer_list) |
| 200 | |
| 201 | if (ifm_is_sg_ofm or ifm_is_sg_ifm or ifm_is_cpu_produced) and (ofm_is_sg_ofm or ofm_is_cpu_consumed): |
| 202 | # Both ifm and ofm need to persist, but only ifm need a copy, in order to remove the Reshape |
| 203 | insert_add_copy_op_after_tens(op.ifm) |
| 204 | |
| 205 | return op |
| 206 | |
| 207 | |
| 208 | def create_add_for_concat(concat_op, name, ifm, ifm_shape: Shape4D, write_offset: Shape4D): |
| 209 | """Creates an add op for the given concat op/input feature map""" |
| 210 | ofm = concat_op.ofm |
| 211 | ifm2 = create_const_tensor( |
| 212 | name + "_zero_scalar", [1], ofm.dtype, [0], ofm.dtype.as_numpy_type(), quantization=ofm.quantization |
| 213 | ) |
| 214 | add_op = create_add_nop(name) |
| 215 | |
| 216 | add_op.inputs = [ifm, ifm2] |
| 217 | add_op.outputs = [ofm] |
| 218 | add_op.write_offset = write_offset |
| 219 | add_op.write_shape = ifm_shape |
| 220 | ofm.ops.append(add_op) |
| 221 | DebugDatabase.add_optimised(concat_op, add_op) |
| 222 | add_op.ifm_shapes.append(ifm_shape) |
| 223 | add_op.ifm_shapes.append(Shape4D(ifm2.shape)) |
| 224 | add_op.ofm_shapes.append(concat_op.ofm_shapes[0]) |
| 225 | add_op.memory_function = Op.ConcatSliceWrite |
| 226 | return add_op |
| 227 | |
| 228 | |
| 229 | # TODO Could be further optimized checking the type of the consumer, |
| 230 | # rather than just mimic the TFLite behaviour depending on type. |
| 231 | # TOSA bool_t not considered yet |
| 232 | def remove_splitsliceread(op, arch): |
| 233 | |
| 234 | if op.type == Op.SplitSliceRead: |
| 235 | # Check if it is possible to put the SplitSliceRead on the tensor consumer, or if an avgpool need to be inserted |
| 236 | if ( |
| 237 | len(op.ofm.consumer_list) == 1 |
| 238 | and op.ofm.consumer_list[0] is not None |
| 239 | and op.ofm.consumer_list[0].run_on_npu |
| 240 | and op.ofm.consumer_list[0].type != Op.Reshape |
| 241 | and op.ofm_shapes[0] == Shape4D.from_list(op.ofm.shape) |
| 242 | and op.ofm.dtype in (DataType.uint8, DataType.int8, DataType.int16) |
| 243 | ): |
| 244 | # SplitSliceRead can be performed by tensor consumer |
| 245 | cons_op = op.ofm.consumer_list[0] |
| 246 | move_splitsliceread_to_consumer(op, cons_op) |
| 247 | else: |
| 248 | name = op.name + "_add" |
| 249 | ofm = op.ofm |
| 250 | ifm2 = create_const_tensor( |
| 251 | name + "_zero_scalar", [1], ofm.dtype, [0], ofm.dtype.as_numpy_type(), quantization=ofm.quantization |
| 252 | ) |
| 253 | add_op = create_add_nop(name) |
| 254 | add_op.inputs = [op.ifm, ifm2] |
| 255 | add_op.outputs = [ofm] |
| 256 | op.ofm.ops.remove(op) |
| 257 | op.ofm.ops.append(add_op) |
| 258 | add_op.ifm_shapes.append(op.ifm_shapes[0]) |
| 259 | add_op.ifm_shapes.append(Shape4D(ifm2.shape)) |
| 260 | add_op.ofm_shapes.append(op.ofm_shapes[0]) |
| 261 | add_op.read_offsets[0] = op.read_offsets[0] |
| 262 | add_op.read_shapes[0] = op.read_shapes[0] |
| 263 | |
| 264 | op.ifm.consumer_list.remove(op) |
| 265 | DebugDatabase.add_optimised(op, add_op) |
| 266 | |
| 267 | |
Patrik Gustavsson | c2b129d | 2021-09-23 13:52:34 +0200 | [diff] [blame^] | 268 | def rewrite_concat(op): |
Patrik Gustavsson | f1580f0 | 2021-09-01 12:43:02 +0200 | [diff] [blame] | 269 | if not op.run_on_npu or not op.type == Op.Concat: |
| 270 | return |
| 271 | |
Patrik Gustavsson | f1580f0 | 2021-09-01 12:43:02 +0200 | [diff] [blame] | 272 | offset = 0 |
Patrik Gustavsson | f1580f0 | 2021-09-01 12:43:02 +0200 | [diff] [blame] | 273 | inputs = op.inputs |
Patrik Gustavsson | c2b129d | 2021-09-23 13:52:34 +0200 | [diff] [blame^] | 274 | axis_4D = op.attrs["axis4D"] |
Patrik Gustavsson | f1580f0 | 2021-09-01 12:43:02 +0200 | [diff] [blame] | 275 | |
| 276 | for idx, inp in enumerate(inputs): |
Patrik Gustavsson | f1580f0 | 2021-09-01 12:43:02 +0200 | [diff] [blame] | 277 | write_offset = [0, 0, 0, 0] |
| 278 | write_offset[axis_4D] = offset |
| 279 | concat_end = offset + op.ifm_shapes[idx][axis_4D] |
| 280 | create_add_for_concat(op, op.name + str(idx) + "_add", inp, op.ifm_shapes[idx], Shape4D.from_list(write_offset)) |
| 281 | offset = concat_end |
Patrik Gustavsson | c2b129d | 2021-09-23 13:52:34 +0200 | [diff] [blame^] | 282 | assert op.ofm_shapes[0][axis_4D] == offset |
Patrik Gustavsson | f1580f0 | 2021-09-01 12:43:02 +0200 | [diff] [blame] | 283 | |
| 284 | |
Patrik Gustavsson | df99510 | 2021-08-23 15:33:59 +0200 | [diff] [blame] | 285 | def remove_reshapes(op, arch): |
| 286 | if op.run_on_npu and op.type == Op.Reshape: |
Jonas Ohlsson | 0957e3e | 2021-09-01 15:57:21 +0200 | [diff] [blame] | 287 | bypass_memory_only_ops(op) |
Patrik Gustavsson | df99510 | 2021-08-23 15:33:59 +0200 | [diff] [blame] | 288 | |
| 289 | |
Patrik Gustavsson | 8f1f9aa | 2021-06-28 07:41:58 +0200 | [diff] [blame] | 290 | def rewrite_activation(op, arch, nng): |
Patrik Gustavsson | 5e26eda | 2021-06-30 09:07:16 +0200 | [diff] [blame] | 291 | if op.type not in (Op.ReluN, Op.Clamp): |
Patrik Gustavsson | 8f1f9aa | 2021-06-28 07:41:58 +0200 | [diff] [blame] | 292 | return op |
| 293 | |
| 294 | ifm = op.ifm |
Patrik Gustavsson | 8f1f9aa | 2021-06-28 07:41:58 +0200 | [diff] [blame] | 295 | zp = ifm.quantization.zero_point if ifm.quantization.zero_point else 0 |
| 296 | if op.ofm.quantization.zero_point is None: |
| 297 | op.ofm.quantization.zero_point = zp |
| 298 | |
Patrik Gustavsson | 5e26eda | 2021-06-30 09:07:16 +0200 | [diff] [blame] | 299 | if op.type == Op.Clamp: |
Patrik Gustavsson | 8f1f9aa | 2021-06-28 07:41:58 +0200 | [diff] [blame] | 300 | op.attrs["min"] = op.attrs["min_int"] - zp |
| 301 | op.attrs["max"] = op.attrs["max_int"] - zp |
| 302 | elif op.type == Op.ReluN: |
| 303 | op.attrs["max"] = op.attrs["max_int"] - zp |
Patrik Gustavsson | 8f1f9aa | 2021-06-28 07:41:58 +0200 | [diff] [blame] | 304 | |
| 305 | return op |
| 306 | |
| 307 | |
| 308 | def rewrite_rescale(op, arch, nng): |
| 309 | if op.type == Op.Rescale: |
| 310 | ifm = op.ifm |
| 311 | ofm = op.ofm |
| 312 | |
| 313 | # some error checking |
| 314 | assert len(ifm.ops) == 1 |
| 315 | prev_op = ifm.ops[0] |
| 316 | |
| 317 | # TODO currently not supported |
Patrik Gustavsson | 8f1f9aa | 2021-06-28 07:41:58 +0200 | [diff] [blame] | 318 | assert len(ifm.consumer_list) == 1 |
| 319 | |
| 320 | input_zp = op.attrs["input_zp"] |
| 321 | output_zp = op.attrs["output_zp"] |
| 322 | multiplier = op.attrs["multiplier"] |
| 323 | shift = op.attrs["shift"] |
| 324 | scale32 = op.attrs["scale32"] |
| 325 | double_round = op.attrs["double_round"] |
| 326 | per_channel = op.attrs["per_channel"] |
| 327 | |
| 328 | assert ifm.dtype in (DataType.uint8, DataType.int8, DataType.int32) |
| 329 | assert ifm.dtype in (DataType.uint8, DataType.int8) or input_zp == 0 |
| 330 | assert ofm.dtype in (DataType.uint8, DataType.int8) or output_zp == 0 |
| 331 | assert (scale32 and ifm.dtype != DataType.int48) or (not scale32 and not double_round) |
| 332 | |
| 333 | # Check that input tensor has the same zp or no zp |
| 334 | ifm_zp = ifm.quantization.zero_point |
| 335 | if ifm_zp is not None and ifm_zp != input_zp: |
| 336 | print("Error (fuse_rescale): zp of tensors producer/consumer differs unexpectedidly ") |
| 337 | assert False |
| 338 | ifm.quantization.zero_point = input_zp |
Patrik Gustavsson | c74682c | 2021-08-17 14:26:38 +0200 | [diff] [blame] | 339 | ofm.quantization.zero_point = output_zp |
| 340 | for s, m in zip(shift, multiplier): |
| 341 | # TODO these are the TOSA limitations |
| 342 | assert m >= 0 |
| 343 | assert 2 <= s <= 62 |
| 344 | # TODO these are the HW limitations |
| 345 | assert 0 <= s < (1 << 6) |
| 346 | explicit_scaling = ExplicitScaling(per_channel, shift, multiplier) |
Patrik Gustavsson | 8f1f9aa | 2021-06-28 07:41:58 +0200 | [diff] [blame] | 347 | |
Patrik Gustavsson | c74682c | 2021-08-17 14:26:38 +0200 | [diff] [blame] | 348 | if double_round and scale32: |
| 349 | rounding_mode = NpuRoundingMode.TFL |
| 350 | else: |
| 351 | rounding_mode = NpuRoundingMode.NATURAL |
Patrik Gustavsson | 8f1f9aa | 2021-06-28 07:41:58 +0200 | [diff] [blame] | 352 | |
| 353 | if prev_op.type.is_depthwise_conv2d_op() or prev_op.type.is_conv2d_op() or prev_op.type == Op.FullyConnected: |
| 354 | assert len(multiplier) == len(shift) == len(prev_op.bias.values) |
| 355 | |
| 356 | if ifm.dtype == DataType.int32 and per_channel: |
Patrik Gustavsson | c74682c | 2021-08-17 14:26:38 +0200 | [diff] [blame] | 357 | prev_op.explicit_scaling = explicit_scaling |
| 358 | prev_op.rounding_mode = rounding_mode |
Patrik Gustavsson | 8f1f9aa | 2021-06-28 07:41:58 +0200 | [diff] [blame] | 359 | |
| 360 | # Bypass op |
| 361 | prev_op.set_output_tensor(ofm) |
| 362 | DebugDatabase.add_optimised(op, prev_op) |
| 363 | return op |
| 364 | else: |
| 365 | print("Warning, unsupported fusing of TOSA Rescale previous operator is of type:", prev_op.type) |
| 366 | assert False |
Patrik Gustavsson | c74682c | 2021-08-17 14:26:38 +0200 | [diff] [blame] | 367 | # TODO which are the cases we need to and can do standalone Rescale? |
| 368 | # TODO should we try to identify a conversion uint8<->int8 accomplished by 2 RESCALE ops? |
| 369 | # origin might be TFLite op QUANTIZE, should we look to see if they can be translated to QUANTIZE? |
| 370 | # limited to these at the moment: |
| 371 | elif ( |
| 372 | (ifm.dtype == DataType.int8 and ofm.dtype == DataType.int8) |
| 373 | or (ifm.dtype == DataType.uint8 and ofm.dtype == DataType.int8) |
| 374 | or (ifm.dtype == DataType.int8 and ofm.dtype == DataType.uint8) |
| 375 | ): |
| 376 | # Create NOP performing the RESCALE |
| 377 | avgpool_op = replace_rescale_with_avg_pool(op) |
| 378 | avgpool_op.rounding_mode = rounding_mode |
Patrik Gustavsson | 8f1f9aa | 2021-06-28 07:41:58 +0200 | [diff] [blame] | 379 | |
Patrik Gustavsson | c74682c | 2021-08-17 14:26:38 +0200 | [diff] [blame] | 380 | if per_channel: |
| 381 | # TODO |
| 382 | avgpool_op.explicit_scaling = explicit_scaling |
| 383 | print("Warning, unsupported TOSA Rescale") |
| 384 | assert False |
| 385 | else: |
| 386 | avgpool_op.explicit_scaling = explicit_scaling |
Patrik Gustavsson | 8f1f9aa | 2021-06-28 07:41:58 +0200 | [diff] [blame] | 387 | else: |
| 388 | print("Warning, unsupported fusing of TOSA Rescale previous operator is of type:", prev_op.type) |
| 389 | assert False |
| 390 | return op |
| 391 | |
| 392 | |
Patrik Gustavsson | e2bfa7e | 2021-09-08 15:04:11 +0200 | [diff] [blame] | 393 | # TODO modified copy of TFLite, solution for TOSA PAD will change so reuse has not been considered |
| 394 | def convert_pad(op, arch, nng): |
| 395 | """ |
| 396 | Rewrites PAD operator to an add that copies the IFM to the OFM |
| 397 | + up to 4 add operators that fill the OFM with zeros at the borders. |
| 398 | """ |
| 399 | |
| 400 | if op.type != Op.Pad: |
| 401 | return op |
| 402 | |
| 403 | # TODO assuming rank <= 4 and N = 1 for rank ==4 |
| 404 | # This is checked in tosa_supported_operators |
| 405 | ifm = op.ifm |
| 406 | assert ifm is not None |
| 407 | ifm_shape = Shape4D(ifm.shape) |
| 408 | ofm = op.ofm |
| 409 | assert ofm is not None |
| 410 | ofm.ops = [] |
| 411 | ofm_shape = op.ofm_shapes[0] |
| 412 | |
| 413 | rank = len(ifm.shape) |
| 414 | padding = op.inputs[1].values |
| 415 | pad_depth = padding[-1] |
| 416 | if not (pad_depth == 0).all(): |
| 417 | print("Warning: For PAD, padding in depth not supported yet") |
| 418 | assert False |
| 419 | |
| 420 | top, bottom = 0, 0 |
| 421 | left, right = 0, 0 |
| 422 | if rank > 1: |
| 423 | left, right = padding[-2][0], padding[-2][1] |
| 424 | if rank > 2: |
| 425 | top, bottom = padding[-3][0], padding[-3][1] |
| 426 | if rank == 4 and not (padding[-4] == 0).all(): |
| 427 | print("Warning: For PAD, padding not supported in first dimension when rank == 4 yet") |
| 428 | assert False |
| 429 | |
| 430 | # Add op that copies IFM to the right place inside the OFM |
| 431 | shp0 = Shape4D(0, 0, 0, 0) |
| 432 | shp_top = shp0.with_height(top) |
| 433 | add_op = create_add_for_concat(op, op.name + "_main", ifm, ifm_shape, shp_top.with_width(left)) |
| 434 | add_op.activation = op.activation |
| 435 | |
| 436 | quant = ofm.quantization |
| 437 | pad_value = ifm.quantization.zero_point |
| 438 | # Add operations that fill the borders of the OFM |
| 439 | if top > 0: |
| 440 | shape = Shape4D(1, top, ofm_shape.width, ofm_shape.depth) |
| 441 | zero_tens = create_const_tensor( |
| 442 | op.name + "_top", |
| 443 | shape.as_list(), |
| 444 | ofm.dtype, |
| 445 | shape.elements() * [pad_value], |
| 446 | np.uint8, |
| 447 | quantization=quant, # TODO |
| 448 | ) |
| 449 | # If top/bottom or left/right are equal, the const tensors can be allocated to the same address |
| 450 | zero_tens.equivalence_id = create_equivalence_id(tuple(zero_tens.values)) |
| 451 | create_add_for_concat(op, op.name + "_top", zero_tens, shape, shp0) |
| 452 | if bottom > 0: |
| 453 | shape = Shape4D(1, bottom, ofm_shape.width, ofm_shape.depth) |
| 454 | zero_tens = create_const_tensor( |
| 455 | op.name + "_bottom", |
| 456 | shape.as_list(), |
| 457 | ofm.dtype, |
| 458 | shape.elements() * [pad_value], |
| 459 | np.uint8, |
| 460 | quantization=quant, |
| 461 | ) |
| 462 | zero_tens.equivalence_id = create_equivalence_id(tuple(zero_tens.values)) |
| 463 | create_add_for_concat(op, op.name + "_bottom", zero_tens, shape, shp0.with_height(ofm_shape.height - bottom)) |
| 464 | if left > 0: |
| 465 | shape = Shape4D(1, ifm_shape.height, left, ofm_shape.depth) |
| 466 | zero_tens = create_const_tensor( |
| 467 | op.name + "_left", shape.as_list(), ofm.dtype, shape.elements() * [pad_value], np.uint8, quantization=quant |
| 468 | ) |
| 469 | zero_tens.equivalence_id = create_equivalence_id(tuple(zero_tens.values)) |
| 470 | create_add_for_concat(op, op.name + "_left", zero_tens, shape, shp_top) |
| 471 | if right > 0: |
| 472 | shape = Shape4D(1, ifm_shape.height, right, ofm_shape.depth) |
| 473 | zero_tens = create_const_tensor( |
| 474 | op.name + "_right", shape.as_list(), ofm.dtype, shape.elements() * [pad_value], np.uint8, quantization=quant |
| 475 | ) |
| 476 | zero_tens.equivalence_id = create_equivalence_id(tuple(zero_tens.values)) |
| 477 | create_add_for_concat(op, op.name + "_right", zero_tens, shape, shp_top.with_width(ofm_shape.width - right)) |
| 478 | |
| 479 | op.type = Op.ConcatTFLite |
| 480 | return add_op |
| 481 | |
| 482 | |
Patrik Gustavsson | f436ada | 2021-09-14 14:56:48 +0200 | [diff] [blame] | 483 | def convert_table_to_lut(op, arch, nng): |
| 484 | # Converts table op to a no-op + LUT |
| 485 | if op.type is not Op.Table: |
| 486 | return op |
| 487 | |
| 488 | table = op.inputs[1] |
| 489 | op.inputs.remove(table) |
| 490 | op.set_ifm_ofm_shapes() |
| 491 | |
| 492 | return convert_to_lut(op, table.values, "table") |
| 493 | |
| 494 | |
Patrik Gustavsson | c2b129d | 2021-09-23 13:52:34 +0200 | [diff] [blame^] | 495 | def decompose_elem_tensors_hwc(op): |
| 496 | """ |
| 497 | Decomposes elementwise op if any of the ifm(s)/ofm are to large in any dimension to be handled by the NPU |
| 498 | """ |
Patrik Gustavsson | 3f22ec2 | 2021-09-21 14:18:44 +0200 | [diff] [blame] | 499 | max_t_size = 65535 |
Patrik Gustavsson | c2b129d | 2021-09-23 13:52:34 +0200 | [diff] [blame^] | 500 | ofm_shape = op.write_shape if op.write_shape is not None else op.ofm_shapes[0] |
| 501 | ifm_shape = op.read_shapes[0] if op.read_shapes[0] is not None else op.ifm_shapes[0] |
Patrik Gustavsson | 3f22ec2 | 2021-09-21 14:18:44 +0200 | [diff] [blame] | 502 | ifm2_shape = op.ifm_shapes[1] if op.ifm_shapes[1] else None |
Patrik Gustavsson | c2b129d | 2021-09-23 13:52:34 +0200 | [diff] [blame^] | 503 | ifm2_shape = op.read_shapes[1] if op.read_shapes[1] is not None else ifm2_shape |
Patrik Gustavsson | 3f22ec2 | 2021-09-21 14:18:44 +0200 | [diff] [blame] | 504 | limit_shape = Shape4D(1, max_t_size, max_t_size, max_t_size) |
| 505 | |
| 506 | if any(dim_size > max_t_size for dim_size in ofm_shape.as_list()): |
| 507 | ofm_split = ofm_shape.floordiv_const(max_t_size).add(1, 1, 1, 1) |
| 508 | |
| 509 | for height in range(ofm_split.height): |
| 510 | for width in range(ofm_split.width): |
| 511 | for depth in range(ofm_split.depth): |
| 512 | ofm_offset = Shape4D(0, height * max_t_size, width * max_t_size, depth * max_t_size) |
| 513 | ofm_part_shape = ofm_shape.clip(ofm_offset, limit_shape) |
| 514 | ofm_cut = (ofm_offset, ofm_part_shape) |
| 515 | |
| 516 | ifm_d = depth * max_t_size if ifm_shape.depth == ofm_shape.depth else 0 |
| 517 | ifm_w = width * max_t_size if ifm_shape.width == ofm_shape.width else 0 |
| 518 | ifm_h = height * max_t_size if ifm_shape.height == ofm_shape.height else 0 |
| 519 | ifm_offset = Shape4D(0, ifm_h, ifm_w, ifm_d) |
| 520 | ifm_part_shape = ifm_shape.clip(ifm_offset, limit_shape) |
| 521 | ifm_cut = (ifm_offset, ifm_part_shape) |
| 522 | |
| 523 | if ifm2_shape is not None: |
| 524 | ifm2_d = depth * max_t_size if ifm2_shape.depth == ofm_shape.depth else 0 |
| 525 | ifm2_w = width * max_t_size if ifm2_shape.width == ofm_shape.width else 0 |
| 526 | ifm2_h = height * max_t_size if ifm2_shape.height == ofm_shape.height else 0 |
| 527 | ifm2_offset = Shape4D(0, ifm2_h, ifm2_w, ifm2_d) |
| 528 | ifm2_part_shape = ifm2_shape.clip(ifm2_offset, limit_shape) |
| 529 | ifm2_cut = (ifm2_offset, ifm2_part_shape) |
| 530 | else: |
Patrik Gustavsson | 3f22ec2 | 2021-09-21 14:18:44 +0200 | [diff] [blame] | 531 | ifm2_cut = (None, None) |
| 532 | |
| 533 | create_elem_part_op(op, ifm_cut, ifm2_cut, ofm_cut) |
| 534 | op.ofm.ops.remove(op) |
| 535 | op.ifm.consumer_list.remove(op) |
| 536 | if op.ifm2 is not None: |
| 537 | op.ifm2.consumer_list.remove(op) |
| 538 | return |
| 539 | |
| 540 | |
| 541 | def create_elem_part_op(op, ifm_cut, ifm2_cut, ofm_cut): |
| 542 | part_op = op.clone() |
| 543 | ifm_read_offset = op.read_offsets[0] if op.read_offsets[0] is not None else Shape4D(0, 0, 0, 0) |
| 544 | ofm_write_offset = op.write_offset if op.write_offset is not None else Shape4D(0, 0, 0, 0) |
| 545 | ifm_offset, ifm_shape = ifm_cut |
| 546 | ofm_offset, ofm_shape = ofm_cut |
| 547 | |
| 548 | part_op.read_offsets[0] = ifm_read_offset + ifm_offset |
| 549 | part_op.read_shapes[0] = ifm_shape |
| 550 | part_op.write_offset = ofm_write_offset + ofm_offset |
| 551 | part_op.write_shape = ofm_shape |
Patrik Gustavsson | 46408a8 | 2021-09-20 10:47:47 +0200 | [diff] [blame] | 552 | part_op.ifm_shapes = op.ifm_shapes.copy() |
| 553 | part_op.ofm_shapes = op.ofm_shapes.copy() |
| 554 | part_op.ifm.consumer_list.append(part_op) |
| 555 | op.ofm.ops.append(part_op) |
Patrik Gustavsson | 3f22ec2 | 2021-09-21 14:18:44 +0200 | [diff] [blame] | 556 | |
| 557 | ifm2_offset, ifm2_shape = ifm2_cut |
| 558 | if ifm2_offset: |
| 559 | ifm2_read_offset = op.read_offsets[1] if op.read_offsets[1] is not None else Shape4D(0, 0, 0, 0) |
| 560 | part_op.read_offsets[1] = ifm2_read_offset + ifm2_offset |
| 561 | part_op.read_shapes[1] = ifm2_shape |
Patrik Gustavsson | 46408a8 | 2021-09-20 10:47:47 +0200 | [diff] [blame] | 562 | part_op.ifm2.consumer_list.append(part_op) |
| 563 | |
| 564 | |
| 565 | def get_nhwc_stride(shape): |
| 566 | stride_x = shape.depth |
| 567 | stride_y = shape.width * stride_x |
| 568 | stride_n = shape.height * stride_y |
| 569 | return Shape4D(stride_n, stride_y, stride_x, 1) |
| 570 | |
| 571 | |
Patrik Gustavsson | 3f22ec2 | 2021-09-21 14:18:44 +0200 | [diff] [blame] | 572 | def get_elem_shapes_removed_singles(op): |
Patrik Gustavsson | 46408a8 | 2021-09-20 10:47:47 +0200 | [diff] [blame] | 573 | """ |
Patrik Gustavsson | 3f22ec2 | 2021-09-21 14:18:44 +0200 | [diff] [blame] | 574 | Returns the shapes of ifm(s)/ofms after removing all the dimensions that are 1 for all ifm(s)/ofm |
Patrik Gustavsson | 46408a8 | 2021-09-20 10:47:47 +0200 | [diff] [blame] | 575 | """ |
Patrik Gustavsson | 3f22ec2 | 2021-09-21 14:18:44 +0200 | [diff] [blame] | 576 | binary = op.ifm2 is not None |
Patrik Gustavsson | c2b129d | 2021-09-23 13:52:34 +0200 | [diff] [blame^] | 577 | ofm_shape = op.ofm_shapes[0].as_list() if len(op.ofm_shapes) > 0 else op.ofm.shape |
| 578 | ifm_shape = op.ifm_shapes[0].as_list() if len(op.ifm_shapes) > 0 else op.ifm.shape |
| 579 | if binary: |
| 580 | ifm2_shape = op.ifm_shapes[1].as_list() if len(op.ofm_shapes) else op.ifm2.shape |
| 581 | |
| 582 | rank = len(ofm_shape) |
Patrik Gustavsson | 3f22ec2 | 2021-09-21 14:18:44 +0200 | [diff] [blame] | 583 | new_ofm_shape = [] |
| 584 | new_ifm_shape = [] |
| 585 | new_ifm2_shape = [] |
| 586 | for idx in range(rank): |
Patrik Gustavsson | c2b129d | 2021-09-23 13:52:34 +0200 | [diff] [blame^] | 587 | if ofm_shape[idx] != 1: |
| 588 | new_ofm_shape.append(ofm_shape[idx]) |
| 589 | new_ifm_shape.append(ifm_shape[idx]) |
Patrik Gustavsson | 3f22ec2 | 2021-09-21 14:18:44 +0200 | [diff] [blame] | 590 | if binary: |
Patrik Gustavsson | c2b129d | 2021-09-23 13:52:34 +0200 | [diff] [blame^] | 591 | new_ifm2_shape.append(ifm2_shape[idx]) |
| 592 | |
Patrik Gustavsson | 3f22ec2 | 2021-09-21 14:18:44 +0200 | [diff] [blame] | 593 | if new_ofm_shape == []: |
| 594 | new_ofm_shape = [1] |
| 595 | new_ifm_shape = [1] |
| 596 | new_ifm2_shape = [1] if binary else None |
Patrik Gustavsson | 46408a8 | 2021-09-20 10:47:47 +0200 | [diff] [blame] | 597 | |
Patrik Gustavsson | 3f22ec2 | 2021-09-21 14:18:44 +0200 | [diff] [blame] | 598 | return new_ofm_shape, new_ifm_shape, new_ifm2_shape |
Patrik Gustavsson | 46408a8 | 2021-09-20 10:47:47 +0200 | [diff] [blame] | 599 | |
| 600 | |
Patrik Gustavsson | 3f22ec2 | 2021-09-21 14:18:44 +0200 | [diff] [blame] | 601 | def decomp_dims_elementwise(op): |
Patrik Gustavsson | 46408a8 | 2021-09-20 10:47:47 +0200 | [diff] [blame] | 602 | """ |
Patrik Gustavsson | 3f22ec2 | 2021-09-21 14:18:44 +0200 | [diff] [blame] | 603 | Decompose elementwise ops with Rank > 3 (H,W,D). |
Patrik Gustavsson | 46408a8 | 2021-09-20 10:47:47 +0200 | [diff] [blame] | 604 | If Rank > 3, all the dimensions above H are viewed as the N dimension. |
| 605 | the elementwise operation will be decomposed to N (of ofm) elementwise operations. |
| 606 | By reading and writing with offsets from/to the ifm(s)/ofm. |
Patrik Gustavsson | 3f22ec2 | 2021-09-21 14:18:44 +0200 | [diff] [blame] | 607 | Note: Broadcast need to be handled for binary elementwise ops, and TOSA allowes for broadcast by both ifm and ifm2 |
Patrik Gustavsson | 46408a8 | 2021-09-20 10:47:47 +0200 | [diff] [blame] | 608 | """ |
| 609 | |
| 610 | ifm = op.ifm |
| 611 | ifm2 = op.ifm2 |
| 612 | ofm = op.ofm |
Patrik Gustavsson | 3f22ec2 | 2021-09-21 14:18:44 +0200 | [diff] [blame] | 613 | binary = op.ifm2 is not None |
Patrik Gustavsson | 46408a8 | 2021-09-20 10:47:47 +0200 | [diff] [blame] | 614 | |
Patrik Gustavsson | 3f22ec2 | 2021-09-21 14:18:44 +0200 | [diff] [blame] | 615 | # Remove dimensions that are all 1 |
| 616 | new_ofm_shape, new_ifm_shape, new_ifm2_shape = get_elem_shapes_removed_singles(op) |
| 617 | rank = len(new_ofm_shape) |
| 618 | |
Patrik Gustavsson | 46408a8 | 2021-09-20 10:47:47 +0200 | [diff] [blame] | 619 | if rank > 3: |
| 620 | n = rank - 3 |
Patrik Gustavsson | 3f22ec2 | 2021-09-21 14:18:44 +0200 | [diff] [blame] | 621 | ofm_decomp_shape = Shape4D(new_ofm_shape[0:n]) |
Patrik Gustavsson | 46408a8 | 2021-09-20 10:47:47 +0200 | [diff] [blame] | 622 | ofm_decomp_stride = get_nhwc_stride(ofm_decomp_shape) |
Patrik Gustavsson | 3f22ec2 | 2021-09-21 14:18:44 +0200 | [diff] [blame] | 623 | ofm_part_shape = Shape4D(new_ofm_shape[n:]) |
| 624 | op.ofm_shapes.append(Shape4D([ofm_decomp_shape.elements()] + new_ofm_shape[n:])) |
Patrik Gustavsson | 46408a8 | 2021-09-20 10:47:47 +0200 | [diff] [blame] | 625 | |
Patrik Gustavsson | 3f22ec2 | 2021-09-21 14:18:44 +0200 | [diff] [blame] | 626 | if binary: |
| 627 | ifm_decomp_shape = Shape4D(new_ifm_shape[0:n]) |
| 628 | ifm2_decomp_shape = Shape4D(new_ifm2_shape[0:n]) |
| 629 | ifm_decomp_stride = get_nhwc_stride(ifm_decomp_shape) |
| 630 | ifm2_decomp_stride = get_nhwc_stride(ifm2_decomp_shape) |
| 631 | ifm_part_shape = Shape4D(new_ifm_shape[n:]) |
| 632 | ifm2_part_shape = Shape4D(new_ifm2_shape[n:]) |
| 633 | op.ifm_shapes.append(Shape4D([ifm_decomp_shape.elements()] + new_ifm_shape[n:])) |
| 634 | op.ifm_shapes.append(Shape4D([ifm2_decomp_shape.elements()] + new_ifm2_shape[n:])) |
| 635 | else: |
| 636 | op.ifm_shapes.append(Shape4D([ofm_decomp_shape.elements()] + new_ofm_shape[n:])) |
Patrik Gustavsson | 46408a8 | 2021-09-20 10:47:47 +0200 | [diff] [blame] | 637 | |
Patrik Gustavsson | 3f22ec2 | 2021-09-21 14:18:44 +0200 | [diff] [blame] | 638 | op_list = [] |
Patrik Gustavsson | 46408a8 | 2021-09-20 10:47:47 +0200 | [diff] [blame] | 639 | for height in range(ofm_decomp_shape.height): |
| 640 | for width in range(ofm_decomp_shape.width): |
| 641 | for depth in range(ofm_decomp_shape.depth): |
| 642 | ofm_offset = Shape4D(0, height, width, depth) |
Patrik Gustavsson | 3f22ec2 | 2021-09-21 14:18:44 +0200 | [diff] [blame] | 643 | ofm_offset = Shape4D(ofm_offset.dot_prod(ofm_decomp_stride), 0, 0, 0) |
| 644 | ofm_cut = (ofm_offset, ofm_part_shape) |
Patrik Gustavsson | 46408a8 | 2021-09-20 10:47:47 +0200 | [diff] [blame] | 645 | |
Patrik Gustavsson | 3f22ec2 | 2021-09-21 14:18:44 +0200 | [diff] [blame] | 646 | if binary: |
| 647 | ifm_d = depth if ifm_decomp_shape.depth == ofm_decomp_shape.depth else 0 |
| 648 | ifm_w = width if ifm_decomp_shape.width == ofm_decomp_shape.width else 0 |
| 649 | ifm_h = height if ifm_decomp_shape.height == ofm_decomp_shape.height else 0 |
| 650 | ifm_offset = Shape4D(0, ifm_h, ifm_w, ifm_d) |
| 651 | ifm_offset = Shape4D(ifm_offset.dot_prod(ifm_decomp_stride), 0, 0, 0) |
| 652 | ifm_cut = (ifm_offset, ifm_part_shape) |
Patrik Gustavsson | 46408a8 | 2021-09-20 10:47:47 +0200 | [diff] [blame] | 653 | |
Patrik Gustavsson | 3f22ec2 | 2021-09-21 14:18:44 +0200 | [diff] [blame] | 654 | ifm2_d = depth if ifm2_decomp_shape.depth == ofm_decomp_shape.depth else 0 |
| 655 | ifm2_w = width if ifm2_decomp_shape.width == ofm_decomp_shape.width else 0 |
| 656 | ifm2_h = height if ifm2_decomp_shape.height == ofm_decomp_shape.height else 0 |
| 657 | ifm2_offset = Shape4D(0, ifm2_h, ifm2_w, ifm2_d) |
| 658 | ifm2_offset = Shape4D(ifm2_offset.dot_prod(ifm2_decomp_stride), 0, 0, 0) |
| 659 | ifm2_cut = (ifm2_offset, ifm2_part_shape) |
| 660 | op_list.append(create_elem_part_op(op, ifm_cut, ifm2_cut, ofm_cut)) |
| 661 | else: |
| 662 | op_list.append(create_elem_part_op(op, ofm_cut, None, ofm_cut)) |
Patrik Gustavsson | 46408a8 | 2021-09-20 10:47:47 +0200 | [diff] [blame] | 663 | |
Patrik Gustavsson | 46408a8 | 2021-09-20 10:47:47 +0200 | [diff] [blame] | 664 | ofm.ops.remove(op) |
Patrik Gustavsson | 3f22ec2 | 2021-09-21 14:18:44 +0200 | [diff] [blame] | 665 | ifm.consumer_list.remove(op) |
| 666 | if binary: |
| 667 | ifm2.consumer_list.remove(op) |
| 668 | else: |
| 669 | op.ofm_shapes.append(Shape4D(new_ofm_shape)) |
| 670 | op.ifm_shapes.append(Shape4D(new_ifm_shape)) |
| 671 | op.ifm_shapes.append(Shape4D(new_ifm2_shape)) |
| 672 | |
| 673 | return [op] |
Patrik Gustavsson | 46408a8 | 2021-09-20 10:47:47 +0200 | [diff] [blame] | 674 | |
| 675 | |
| 676 | def decomp_elementwise(tens, arch, nng): |
| 677 | """ |
Patrik Gustavsson | c2b129d | 2021-09-23 13:52:34 +0200 | [diff] [blame^] | 678 | Decompose elementwise ops with Rank > 3 (H,W,C). |
Patrik Gustavsson | 3f22ec2 | 2021-09-21 14:18:44 +0200 | [diff] [blame] | 679 | Decompose size of tensors exceeding NPU max size |
Patrik Gustavsson | 46408a8 | 2021-09-20 10:47:47 +0200 | [diff] [blame] | 680 | """ |
Patrik Gustavsson | c2b129d | 2021-09-23 13:52:34 +0200 | [diff] [blame^] | 681 | tens_ops = tens.ops.copy() |
| 682 | for op in tens_ops: |
| 683 | if op.type.is_elementwise_op(): |
| 684 | decomp_list = decomp_dims_elementwise(op) |
| 685 | for part_op in decomp_list: |
| 686 | decompose_elem_tensors_hwc(part_op) |
| 687 | return tens |
| 688 | |
| 689 | |
| 690 | def reshape_concat_shape(shape, rank, axis): |
| 691 | new_h = 1 |
| 692 | for i in range(axis): |
| 693 | new_h *= shape[i] |
| 694 | new_c = 1 |
| 695 | for i in range(axis + 1, rank): |
| 696 | new_c *= shape[i] |
| 697 | if axis == (rank - 1): |
| 698 | new_shape = [new_h, shape[axis], 1] |
| 699 | else: |
| 700 | new_shape = [new_h, shape[axis], new_c] |
| 701 | return new_shape |
| 702 | |
| 703 | |
| 704 | def reshape_concat(op): |
| 705 | """ |
| 706 | Reshapes concat ops with Rank > 3 (H,W,C). |
| 707 | """ |
| 708 | ofm = op.ofm |
| 709 | rank = len(ofm.shape) |
| 710 | axis = op.attrs["axis"] |
| 711 | if axis < 0: |
| 712 | axis += rank |
| 713 | |
| 714 | if rank > 3: |
| 715 | # Reshape so that axis in to be concatenated is the W dimension |
| 716 | # Reshape inputs |
| 717 | for inp in op.inputs: |
| 718 | new_shape = reshape_concat_shape(inp.shape, rank, axis) |
| 719 | op.ifm_shapes.append(Shape4D(new_shape)) |
| 720 | # Reshape output |
| 721 | new_shape = reshape_concat_shape(ofm.shape, rank, axis) |
| 722 | op.ofm_shapes.append(Shape4D(new_shape)) |
| 723 | op.attrs["axis4D"] = 2 |
| 724 | else: |
| 725 | for inp in op.inputs: |
| 726 | op.ifm_shapes.append(Shape4D(inp.shape)) |
| 727 | op.ofm_shapes.append(Shape4D(ofm.shape)) |
| 728 | op.attrs["axis4D"] = axis + (4 - rank) |
| 729 | |
| 730 | |
| 731 | def decomp_rewrite_concat(tens, arch, nng): |
| 732 | """ |
| 733 | Decompose concat ops with Rank > 3 (H,W,C). |
| 734 | Rewrite of concat to elementwise operations |
| 735 | """ |
| 736 | if len(tens.ops) == 1 and tens.ops[0].type == Op.Concat: |
Patrik Gustavsson | 3f22ec2 | 2021-09-21 14:18:44 +0200 | [diff] [blame] | 737 | op = tens.ops[0] |
Patrik Gustavsson | 46408a8 | 2021-09-20 10:47:47 +0200 | [diff] [blame] | 738 | |
Patrik Gustavsson | c2b129d | 2021-09-23 13:52:34 +0200 | [diff] [blame^] | 739 | reshape_concat(op) |
| 740 | rewrite_concat(op) |
Patrik Gustavsson | 3f22ec2 | 2021-09-21 14:18:44 +0200 | [diff] [blame] | 741 | |
Patrik Gustavsson | c2b129d | 2021-09-23 13:52:34 +0200 | [diff] [blame^] | 742 | op.ofm.ops.remove(op) |
| 743 | for inp in op.inputs: |
| 744 | inp.consumer_list.remove(op) |
| 745 | |
Patrik Gustavsson | 46408a8 | 2021-09-20 10:47:47 +0200 | [diff] [blame] | 746 | return tens |
| 747 | |
| 748 | |
Patrik Gustavsson | c74682c | 2021-08-17 14:26:38 +0200 | [diff] [blame] | 749 | def fixup_quantization(op, arch, nng): |
| 750 | if op.ifm and op.ifm.quantization.zero_point is None: |
| 751 | op.ifm.quantization.zero_point = 0 |
| 752 | if op.ifm2 and op.ifm2.quantization.zero_point is None: |
Patrik Gustavsson | f436ada | 2021-09-14 14:56:48 +0200 | [diff] [blame] | 753 | op.ifm2.quantization.zero_point = 0 |
| 754 | if not op.forced_output_quantization: |
| 755 | if op.ofm and op.ofm.quantization and op.ofm.quantization.zero_point is None: |
| 756 | op.ofm.quantization.zero_point = 0 |
Patrik Gustavsson | c74682c | 2021-08-17 14:26:38 +0200 | [diff] [blame] | 757 | return op |
| 758 | |
| 759 | |
Patrik Gustavsson | 8f1f9aa | 2021-06-28 07:41:58 +0200 | [diff] [blame] | 760 | def supported_operator_check(op, arch, nng): |
| 761 | op.run_on_npu = arch.tosa_supported_operators.is_operator_supported(op) |
Patrik Gustavsson | df99510 | 2021-08-23 15:33:59 +0200 | [diff] [blame] | 762 | assert op.run_on_npu or op.type in (Op.Placeholder, Op.SubgraphInput, Op.Const) |
Patrik Gustavsson | 8f1f9aa | 2021-06-28 07:41:58 +0200 | [diff] [blame] | 763 | return op |
| 764 | |
| 765 | |
| 766 | def tosa_optimise_graph(nng, arch): |
Patrik Gustavsson | 46408a8 | 2021-09-20 10:47:47 +0200 | [diff] [blame] | 767 | |
Patrik Gustavsson | c2b129d | 2021-09-23 13:52:34 +0200 | [diff] [blame^] | 768 | # TODO the supported operator checking need to be split in semantic and HW checks |
| 769 | for idx, sg in enumerate(nng.subgraphs): |
| 770 | nng.subgraphs[idx] = rewrite_graph.rewrite_graph_pre_order( |
| 771 | nng, sg, arch, [], [supported_operator_check], rewrite_unsupported=False, |
| 772 | ) |
| 773 | |
| 774 | # Decomposing and rewrite of concat |
| 775 | for idx, sg in enumerate(nng.subgraphs): |
| 776 | nng.subgraphs[idx] = rewrite_graph.rewrite_graph_pre_order( |
| 777 | nng, sg, arch, [decomp_rewrite_concat], [], rewrite_unsupported=False |
| 778 | ) |
| 779 | |
| 780 | # Decomposing of elementwise |
Patrik Gustavsson | 46408a8 | 2021-09-20 10:47:47 +0200 | [diff] [blame] | 781 | for idx, sg in enumerate(nng.subgraphs): |
| 782 | nng.subgraphs[idx] = rewrite_graph.rewrite_graph_pre_order( |
| 783 | nng, sg, arch, [decomp_elementwise], [], rewrite_unsupported=False |
| 784 | ) |
| 785 | |
Patrik Gustavsson | 8f1f9aa | 2021-06-28 07:41:58 +0200 | [diff] [blame] | 786 | for idx, sg in enumerate(nng.subgraphs): |
| 787 | nng.subgraphs[idx] = rewrite_graph.rewrite_graph_pre_order( |
Patrik Gustavsson | c2b129d | 2021-09-23 13:52:34 +0200 | [diff] [blame^] | 788 | nng, sg, arch, [], [set_ifm_ofm_op_shapes], rewrite_unsupported=False, |
Patrik Gustavsson | 8f1f9aa | 2021-06-28 07:41:58 +0200 | [diff] [blame] | 789 | ) |
| 790 | |
Patrik Gustavsson | df99510 | 2021-08-23 15:33:59 +0200 | [diff] [blame] | 791 | # Removal of Transpose |
| 792 | for idx, sg in enumerate(nng.subgraphs): |
| 793 | nng.subgraphs[idx] = rewrite_graph.rewrite_graph_pre_order( |
| 794 | nng, sg, arch, [], [remove_const_transpose], rewrite_unsupported=False, |
| 795 | ) |
| 796 | |
| 797 | # Handle sg input output |
| 798 | for idx, sg in enumerate(nng.subgraphs): |
| 799 | nng.subgraphs[idx] = rewrite_graph.rewrite_graph_pre_order( |
Patrik Gustavsson | f1580f0 | 2021-09-01 12:43:02 +0200 | [diff] [blame] | 800 | nng, sg, arch, [], [fix_sg_input_output_tosa], rewrite_unsupported=False, |
Patrik Gustavsson | df99510 | 2021-08-23 15:33:59 +0200 | [diff] [blame] | 801 | ) |
| 802 | |
| 803 | # Removal of reshapes |
| 804 | for sg in nng.subgraphs: |
| 805 | rewrite_graph.visit_graph_post_order(sg.output_tensors, arch, [], [remove_reshapes]) |
| 806 | sg.refresh_after_modification() |
| 807 | |
Patrik Gustavsson | f366fb1 | 2021-09-07 13:30:29 +0200 | [diff] [blame] | 808 | # TODO, when and where to best handle calc_scaling_avgpool |
| 809 | for idx, sg in enumerate(nng.subgraphs): |
| 810 | nng.subgraphs[idx] = rewrite_graph.rewrite_graph_pre_order( |
| 811 | nng, sg, arch, [], [calc_scaling_avgpool], rewrite_unsupported=False, |
| 812 | ) |
| 813 | |
Patrik Gustavsson | 8f1f9aa | 2021-06-28 07:41:58 +0200 | [diff] [blame] | 814 | # Rewite Operators step |
Patrik Gustavsson | f436ada | 2021-09-14 14:56:48 +0200 | [diff] [blame] | 815 | op_rewrite_list = [set_tensor_equivalence, rewrite_rescale, convert_depthwise_to_conv, convert_table_to_lut] |
Patrik Gustavsson | 8f1f9aa | 2021-06-28 07:41:58 +0200 | [diff] [blame] | 816 | |
| 817 | for idx, sg in enumerate(nng.subgraphs): |
| 818 | nng.subgraphs[idx] = rewrite_graph.rewrite_graph_pre_order( |
| 819 | nng, sg, arch, [], op_rewrite_list, rewrite_unsupported=False, |
| 820 | ) |
| 821 | |
Patrik Gustavsson | c74682c | 2021-08-17 14:26:38 +0200 | [diff] [blame] | 822 | # Post-processing step 1 |
Patrik Gustavsson | 8f1f9aa | 2021-06-28 07:41:58 +0200 | [diff] [blame] | 823 | for idx, sg in enumerate(nng.subgraphs): |
| 824 | nng.subgraphs[idx] = rewrite_graph.rewrite_graph_pre_order( |
Patrik Gustavsson | e2bfa7e | 2021-09-08 15:04:11 +0200 | [diff] [blame] | 825 | nng, sg, arch, [], [rewrite_activation, convert_pad, add_padding_fields], |
Patrik Gustavsson | 8f1f9aa | 2021-06-28 07:41:58 +0200 | [diff] [blame] | 826 | ) |
| 827 | |
Patrik Gustavsson | f1580f0 | 2021-09-01 12:43:02 +0200 | [diff] [blame] | 828 | # Removal of Slice, need to be done after optimisation has been performed, |
| 829 | # since ifm/ofm_shapes are of importance to this function |
| 830 | for sg in nng.subgraphs: |
| 831 | rewrite_graph.visit_graph_post_order(sg.output_tensors, arch, [], [remove_splitsliceread]) |
| 832 | sg.refresh_after_modification() |
| 833 | |
Patrik Gustavsson | c74682c | 2021-08-17 14:26:38 +0200 | [diff] [blame] | 834 | # Post-processing step 2 |
| 835 | for idx, sg in enumerate(nng.subgraphs): |
| 836 | nng.subgraphs[idx] = rewrite_graph.rewrite_graph_pre_order(nng, sg, arch, [], [fixup_quantization],) |
| 837 | |
Patrik Gustavsson | 8f1f9aa | 2021-06-28 07:41:58 +0200 | [diff] [blame] | 838 | return nng |