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 | # Internal representation of a Neural Network Operation. |
Louis Verhaard | e8a5a78 | 2020-11-02 18:04:27 +0100 | [diff] [blame] | 18 | import copy |
Louis Verhaard | aee5d75 | 2020-09-30 09:01:52 +0200 | [diff] [blame] | 19 | from collections import namedtuple |
| 20 | from enum import Enum |
Dwight Lidman | 9b43f84 | 2020-12-08 17:56:44 +0100 | [diff] [blame] | 21 | from typing import Any |
| 22 | from typing import Dict |
| 23 | from typing import List |
Louis Verhaard | e8a5a78 | 2020-11-02 18:04:27 +0100 | [diff] [blame] | 24 | from typing import Optional |
Dwight Lidman | 9b43f84 | 2020-12-08 17:56:44 +0100 | [diff] [blame] | 25 | from typing import TYPE_CHECKING |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 26 | |
Michael McGeagh | 528a56d | 2020-12-16 11:33:21 +0000 | [diff] [blame] | 27 | from .errors import VelaError |
Tim Hall | 4ed38bc | 2020-10-20 18:54:20 +0100 | [diff] [blame] | 28 | from .numeric_util import full_shape |
patrik.gustavsson | eeb8515 | 2020-12-21 17:10:40 +0000 | [diff] [blame] | 29 | from .shape4d import Shape4D |
Tim Hall | 4ed38bc | 2020-10-20 18:54:20 +0100 | [diff] [blame] | 30 | |
Patrik Gustavsson | 2349d42 | 2020-12-01 16:02:29 +0100 | [diff] [blame] | 31 | |
Dwight Lidman | 9b43f84 | 2020-12-08 17:56:44 +0100 | [diff] [blame] | 32 | if TYPE_CHECKING: |
| 33 | from .tensor import Tensor |
| 34 | |
Tim Hall | 4ed38bc | 2020-10-20 18:54:20 +0100 | [diff] [blame] | 35 | PointXY = namedtuple("PointXY", "x y") |
| 36 | PointXYZ = namedtuple("PointXYZ", "x y z") |
| 37 | |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 38 | |
Louis Verhaard | aee5d75 | 2020-09-30 09:01:52 +0200 | [diff] [blame] | 39 | class NpuBlockType(Enum): |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 40 | Default = 0 |
| 41 | ConvolutionMxN = 1 |
| 42 | VectorProduct = 2 |
| 43 | Pooling = 3 |
| 44 | ConvolutionDepthWise = 4 |
| 45 | ElementWise = 5 |
Fredrik Svedberg | a0c3624 | 2020-06-03 15:43:31 +0200 | [diff] [blame] | 46 | ReduceSum = 6 |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 47 | |
| 48 | |
Tim Hall | 4ed38bc | 2020-10-20 18:54:20 +0100 | [diff] [blame] | 49 | class Kernel: |
Louis Verhaard | e8a5a78 | 2020-11-02 18:04:27 +0100 | [diff] [blame] | 50 | """ |
| 51 | Kernel information for NPU operations |
| 52 | """ |
| 53 | |
| 54 | def __init__(self, w: int, h: int, stride_x: int = 1, stride_y: int = 1, dilation_x: int = 1, dilation_y: int = 1): |
| 55 | assert stride_x > 0 and stride_y > 0 |
| 56 | assert dilation_x > 0 and dilation_y > 0 |
Tim Hall | 4ed38bc | 2020-10-20 18:54:20 +0100 | [diff] [blame] | 57 | self.width = w |
| 58 | self.height = h |
Louis Verhaard | e8a5a78 | 2020-11-02 18:04:27 +0100 | [diff] [blame] | 59 | self.stride = PointXY(stride_x, stride_y) |
| 60 | self.dilation = PointXY(dilation_x, dilation_y) |
Tim Hall | 4ed38bc | 2020-10-20 18:54:20 +0100 | [diff] [blame] | 61 | |
Louis Verhaard | e8a5a78 | 2020-11-02 18:04:27 +0100 | [diff] [blame] | 62 | def elements_wh(self) -> int: |
Tim Hall | 4ed38bc | 2020-10-20 18:54:20 +0100 | [diff] [blame] | 63 | return self.width * self.height |
| 64 | |
Louis Verhaard | e8a5a78 | 2020-11-02 18:04:27 +0100 | [diff] [blame] | 65 | def area_width(self) -> int: |
Tim Hall | 4ed38bc | 2020-10-20 18:54:20 +0100 | [diff] [blame] | 66 | return (self.width - 1) * self.dilation.x + 1 |
| 67 | |
Louis Verhaard | e8a5a78 | 2020-11-02 18:04:27 +0100 | [diff] [blame] | 68 | def area_height(self) -> int: |
Tim Hall | 4ed38bc | 2020-10-20 18:54:20 +0100 | [diff] [blame] | 69 | return (self.height - 1) * self.dilation.y + 1 |
| 70 | |
Louis Verhaard | e8a5a78 | 2020-11-02 18:04:27 +0100 | [diff] [blame] | 71 | def __str__(self): |
| 72 | return f"w={self.width}, h={self.height}, stride={tuple(self.stride)}, dilation={tuple(self.dilation)}" |
| 73 | |
Tim Hall | 4ed38bc | 2020-10-20 18:54:20 +0100 | [diff] [blame] | 74 | |
Louis Verhaard | aee5d75 | 2020-09-30 09:01:52 +0200 | [diff] [blame] | 75 | # Classifies operators of type Custom |
| 76 | class CustomType(Enum): |
| 77 | ThirdPartyOp = 0 # Third party custom op |
| 78 | NpuOp = 1 # NPU op |
| 79 | ExistingNpuOp = 2 # NPU op that was part of the input network |
| 80 | |
| 81 | |
| 82 | TensorIndices = namedtuple("TensorIndices", ["ifms", "weights", "biases"]) |
| 83 | |
| 84 | NO_INDICES = TensorIndices([], [], []) |
| 85 | IFM_INDICES = TensorIndices([0], [], []) |
| 86 | IFM_WEIGHTS_INDICES = TensorIndices([0], [1], []) |
| 87 | IFM_WEIGHTS_BIAS_INDICES = TensorIndices([0], [1], [2]) |
| 88 | IFM_IFM2_INDICES = TensorIndices([0, 1], [], []) |
| 89 | CONV2D_BACKPROP_INDICES = TensorIndices([2], [1], [3]) |
| 90 | TRANSPOSE_CONV_INDICES = TensorIndices([0], [1], [3]) |
| 91 | CONCAT_INDICES = TensorIndices([1, 2], [], []) |
| 92 | SPLIT_IFM_INDICES = TensorIndices([1], [], []) |
| 93 | BLOCK_LSTM_INDICES = TensorIndices([3], [4], []) |
| 94 | |
| 95 | |
| 96 | # Static information related to operation codes |
| 97 | class OperatorInfo: |
| 98 | __slots__ = ("id", "block_type", "indices", "is_unary") |
| 99 | _id = 0 |
| 100 | |
| 101 | def __init__(self, block_type=NpuBlockType.Default, indices=NO_INDICES, is_unary=False): |
| 102 | OperatorInfo._id += 1 |
| 103 | self.id = OperatorInfo._id |
| 104 | self.block_type = block_type |
| 105 | self.indices = indices # Indices of the different tensor purposes |
| 106 | self.is_unary = is_unary # Classifies elementwise operators |
| 107 | |
| 108 | |
| 109 | # Internally used operation codes |
| 110 | class Op(Enum): |
| 111 | Abs = OperatorInfo(block_type=NpuBlockType.ElementWise, indices=IFM_INDICES, is_unary=True) |
| 112 | Add = OperatorInfo(block_type=NpuBlockType.ElementWise, indices=IFM_IFM2_INDICES) |
| 113 | AddN = OperatorInfo() |
| 114 | Any = OperatorInfo() |
| 115 | ArgMax = OperatorInfo() |
| 116 | ArgMin = OperatorInfo() |
| 117 | AvgPool = OperatorInfo(block_type=NpuBlockType.Pooling, indices=IFM_INDICES) |
| 118 | BatchMatMul = OperatorInfo() |
| 119 | BatchToSpaceND = OperatorInfo() |
| 120 | BidirectionalSequenceLstm = OperatorInfo(block_type=NpuBlockType.VectorProduct, indices=IFM_WEIGHTS_INDICES) |
| 121 | BidirectionalSequenceRnn = OperatorInfo(block_type=NpuBlockType.VectorProduct, indices=IFM_WEIGHTS_INDICES) |
| 122 | BlockLSTM = OperatorInfo(block_type=NpuBlockType.VectorProduct, indices=BLOCK_LSTM_INDICES) |
| 123 | |
| 124 | CLZ = OperatorInfo( |
| 125 | block_type=NpuBlockType.ElementWise, indices=IFM_INDICES, is_unary=True |
| 126 | ) # NPU specific operation |
| 127 | Call = OperatorInfo() |
| 128 | Cast = OperatorInfo() |
| 129 | Ceil = OperatorInfo() |
Louis Verhaard | e8a5a78 | 2020-11-02 18:04:27 +0100 | [diff] [blame] | 130 | Clip = OperatorInfo() # NPU specific fused activation function for clipping between activation.min/max |
Louis Verhaard | aee5d75 | 2020-09-30 09:01:52 +0200 | [diff] [blame] | 131 | Concat = OperatorInfo(indices=CONCAT_INDICES) |
| 132 | ConcatEmbeddings = OperatorInfo() |
| 133 | ConcatSliceWrite = OperatorInfo(indices=IFM_INDICES) |
Patrik Gustavsson | 2349d42 | 2020-12-01 16:02:29 +0100 | [diff] [blame] | 134 | ConcatTFLite = OperatorInfo(indices=CONCAT_INDICES) |
Louis Verhaard | aee5d75 | 2020-09-30 09:01:52 +0200 | [diff] [blame] | 135 | Const = OperatorInfo() # Constant tensor, only used in CPU subgraphs |
| 136 | Conv2D = OperatorInfo(block_type=NpuBlockType.ConvolutionMxN, indices=IFM_WEIGHTS_INDICES) |
| 137 | Conv2DBackpropInput = OperatorInfo(block_type=NpuBlockType.ConvolutionMxN, indices=CONV2D_BACKPROP_INDICES) |
| 138 | Conv2DBackpropInputSwitchedBias = OperatorInfo( |
| 139 | block_type=NpuBlockType.ConvolutionMxN, indices=TRANSPOSE_CONV_INDICES |
| 140 | ) |
| 141 | Conv2DBias = OperatorInfo(block_type=NpuBlockType.ConvolutionMxN, indices=IFM_WEIGHTS_BIAS_INDICES) |
| 142 | Cos = OperatorInfo() |
| 143 | Custom = OperatorInfo() # Custom 3rd party operator, only used in CPU subgraphs |
| 144 | CustomNpuOp = OperatorInfo() # NPU custom operator, only used in CPU subgraphs |
| 145 | DMA = OperatorInfo() |
| 146 | Delegate = OperatorInfo() |
| 147 | Densify = OperatorInfo() |
| 148 | DepthToSpace = OperatorInfo() |
| 149 | DepthwiseConv2DBias = OperatorInfo(block_type=NpuBlockType.ConvolutionDepthWise, indices=IFM_WEIGHTS_BIAS_INDICES) |
Louis Verhaard | 04f8c00 | 2020-10-09 11:40:21 +0200 | [diff] [blame] | 150 | Dequantize = OperatorInfo(indices=IFM_INDICES) |
Louis Verhaard | aee5d75 | 2020-09-30 09:01:52 +0200 | [diff] [blame] | 151 | Div = OperatorInfo() |
| 152 | Elu = OperatorInfo() |
| 153 | EmbeddingLookup = OperatorInfo() |
| 154 | EmbeddingLookupSparse = OperatorInfo() |
| 155 | Equal = OperatorInfo() |
| 156 | Exp = OperatorInfo() |
| 157 | ExpandDims = OperatorInfo(indices=IFM_INDICES) |
| 158 | FakeQuantWithMinMaxArgs = OperatorInfo() |
| 159 | Fill = OperatorInfo() |
| 160 | Floor = OperatorInfo() |
| 161 | FloorDiv = OperatorInfo() |
| 162 | FloorMod = OperatorInfo() |
| 163 | FullyConnected = OperatorInfo(block_type=NpuBlockType.VectorProduct, indices=IFM_WEIGHTS_BIAS_INDICES) |
| 164 | GatherNd = OperatorInfo() |
| 165 | GatherV2 = OperatorInfo() |
| 166 | Greater = OperatorInfo() |
| 167 | GreaterEqual = OperatorInfo() |
| 168 | HardSwish = OperatorInfo() |
| 169 | HashtableLookup = OperatorInfo() |
| 170 | Identity = OperatorInfo() |
| 171 | If = OperatorInfo() |
| 172 | L2Norm = OperatorInfo() |
| 173 | L2Pool2D = OperatorInfo() |
| 174 | LRN = OperatorInfo() |
| 175 | LSHProjection = OperatorInfo() |
| 176 | LeakyRelu = OperatorInfo(block_type=NpuBlockType.ElementWise, indices=IFM_INDICES, is_unary=True) |
| 177 | Less = OperatorInfo() |
| 178 | LessEqual = OperatorInfo() |
| 179 | Log = OperatorInfo() |
| 180 | LogSoftmax = OperatorInfo() |
| 181 | LogicalAnd = OperatorInfo() |
| 182 | LogicalNot = OperatorInfo() |
| 183 | LogicalOr = OperatorInfo() |
| 184 | Lstm = OperatorInfo(block_type=NpuBlockType.VectorProduct, indices=IFM_WEIGHTS_INDICES) |
| 185 | LUT = OperatorInfo() # NPU specific, operator has LUT, only used in fused activation functions |
| 186 | MatMul = OperatorInfo(block_type=NpuBlockType.VectorProduct, indices=IFM_WEIGHTS_INDICES) |
| 187 | MatrixDiag = OperatorInfo() |
| 188 | MatrixSetDiag = OperatorInfo() |
| 189 | Max = OperatorInfo() |
| 190 | MaxPool = OperatorInfo(block_type=NpuBlockType.Pooling, indices=IFM_INDICES) |
| 191 | Maximum = OperatorInfo(block_type=NpuBlockType.ElementWise, indices=IFM_IFM2_INDICES) |
| 192 | Mean = OperatorInfo() |
| 193 | Min = OperatorInfo() |
| 194 | Minimum = OperatorInfo(block_type=NpuBlockType.ElementWise, indices=IFM_IFM2_INDICES) |
| 195 | MirrorPad = OperatorInfo() |
| 196 | Mul = OperatorInfo(block_type=NpuBlockType.ElementWise, indices=IFM_IFM2_INDICES) |
| 197 | Neg = OperatorInfo() |
| 198 | NonMaxSuppressionV4 = OperatorInfo() |
| 199 | NonMaxSuppressionV5 = OperatorInfo() |
| 200 | NotEqual = OperatorInfo() |
| 201 | OneHot = OperatorInfo() |
Patrik Gustavsson | 2349d42 | 2020-12-01 16:02:29 +0100 | [diff] [blame] | 202 | Pack = OperatorInfo(indices=IFM_INDICES) |
Louis Verhaard | aee5d75 | 2020-09-30 09:01:52 +0200 | [diff] [blame] | 203 | PackReshaped = OperatorInfo(indices=IFM_INDICES) |
Louis Verhaard | ae2d553 | 2020-12-11 17:19:54 +0100 | [diff] [blame] | 204 | Pad = OperatorInfo(indices=IFM_INDICES) |
Louis Verhaard | aee5d75 | 2020-09-30 09:01:52 +0200 | [diff] [blame] | 205 | PadV2 = OperatorInfo() |
| 206 | Placeholder = OperatorInfo() # Only used in CPU subgraphs |
| 207 | Pow = OperatorInfo() |
| 208 | Prelu = OperatorInfo() |
| 209 | Prod = OperatorInfo() |
Louis Verhaard | 04f8c00 | 2020-10-09 11:40:21 +0200 | [diff] [blame] | 210 | Quantize = OperatorInfo(indices=IFM_INDICES) |
Louis Verhaard | aee5d75 | 2020-09-30 09:01:52 +0200 | [diff] [blame] | 211 | QuantizedAvgPool = OperatorInfo(block_type=NpuBlockType.Pooling, indices=IFM_INDICES) |
| 212 | QuantizedConv2D = OperatorInfo(block_type=NpuBlockType.ConvolutionMxN, indices=IFM_WEIGHTS_INDICES) |
| 213 | QuantizedMatMul = OperatorInfo(block_type=NpuBlockType.VectorProduct, indices=IFM_WEIGHTS_INDICES) |
| 214 | QuantizedMaxPool = OperatorInfo(block_type=NpuBlockType.Pooling, indices=IFM_INDICES) |
| 215 | QuantizedReshape = OperatorInfo(indices=IFM_INDICES) |
| 216 | Range = OperatorInfo() |
| 217 | Rank = OperatorInfo() |
| 218 | ReduceSum = OperatorInfo(block_type=NpuBlockType.ReduceSum, indices=IFM_INDICES) |
| 219 | Relu = OperatorInfo(indices=IFM_INDICES) |
| 220 | Relu6 = OperatorInfo(indices=IFM_INDICES) |
| 221 | ReluN1To1 = OperatorInfo(indices=IFM_INDICES) |
| 222 | Reshape = OperatorInfo(indices=IFM_INDICES) |
| 223 | ResizeBilinear = OperatorInfo(block_type=NpuBlockType.Pooling, indices=IFM_INDICES) |
| 224 | ResizeNearestNeighbor = OperatorInfo() |
| 225 | ReverseSequence = OperatorInfo() |
| 226 | ReverseV2 = OperatorInfo() |
| 227 | Rnn = OperatorInfo(block_type=NpuBlockType.VectorProduct, indices=IFM_WEIGHTS_INDICES) |
| 228 | Round = OperatorInfo() |
| 229 | Rsqrt = OperatorInfo() |
| 230 | SHL = OperatorInfo(block_type=NpuBlockType.ElementWise, indices=IFM_IFM2_INDICES) # NPU specific operation |
| 231 | SHR = OperatorInfo(block_type=NpuBlockType.ElementWise, indices=IFM_IFM2_INDICES) # NPU specific operation |
| 232 | ScatterNd = OperatorInfo() |
| 233 | SegmentSum = OperatorInfo() |
| 234 | Select = OperatorInfo() |
| 235 | SelectV2 = OperatorInfo() |
| 236 | Shape = OperatorInfo() |
| 237 | Sigmoid = OperatorInfo(indices=IFM_INDICES) |
| 238 | SignBit = OperatorInfo() |
| 239 | Sin = OperatorInfo() |
| 240 | SkipGram = OperatorInfo() |
| 241 | Slice = OperatorInfo(indices=IFM_INDICES) |
Michael McGeagh | 65fd998 | 2020-10-20 11:49:28 +0100 | [diff] [blame] | 242 | Softmax = OperatorInfo(indices=IFM_INDICES) |
Louis Verhaard | aee5d75 | 2020-09-30 09:01:52 +0200 | [diff] [blame] | 243 | SpaceToBatchND = OperatorInfo() |
| 244 | SpaceToDepth = OperatorInfo() |
| 245 | SparseToDense = OperatorInfo() |
| 246 | Split = OperatorInfo(indices=SPLIT_IFM_INDICES) |
| 247 | SplitSliceRead = OperatorInfo(indices=IFM_INDICES) |
Jacob Bohlin | e3de4e5 | 2020-11-27 14:52:06 +0100 | [diff] [blame] | 248 | SplitV = OperatorInfo(indices=IFM_INDICES) |
Louis Verhaard | aee5d75 | 2020-09-30 09:01:52 +0200 | [diff] [blame] | 249 | Sqrt = OperatorInfo() |
| 250 | Square = OperatorInfo() |
| 251 | SquaredDifference = OperatorInfo() |
| 252 | Squeeze = OperatorInfo(indices=IFM_INDICES) |
| 253 | StridedSlice = OperatorInfo(indices=IFM_INDICES) |
Louis Verhaard | aee5d75 | 2020-09-30 09:01:52 +0200 | [diff] [blame] | 254 | Sub = OperatorInfo(block_type=NpuBlockType.ElementWise, indices=IFM_IFM2_INDICES) |
| 255 | SubgraphInput = OperatorInfo() # Only used in CPU subgraphs |
| 256 | Sum = OperatorInfo() |
| 257 | Svdf = OperatorInfo() |
| 258 | Tanh = OperatorInfo(indices=IFM_INDICES) |
| 259 | Tile = OperatorInfo() |
| 260 | TopKV2 = OperatorInfo() |
| 261 | Transpose = OperatorInfo() |
| 262 | UnidirectionalSequenceLstm = OperatorInfo(block_type=NpuBlockType.VectorProduct, indices=IFM_WEIGHTS_INDICES) |
| 263 | UnidirectionalSequenceRnn = OperatorInfo(block_type=NpuBlockType.VectorProduct, indices=IFM_WEIGHTS_INDICES) |
| 264 | Unique = OperatorInfo() |
Patrik Gustavsson | 2349d42 | 2020-12-01 16:02:29 +0100 | [diff] [blame] | 265 | Unpack = OperatorInfo(indices=IFM_INDICES) |
Louis Verhaard | aee5d75 | 2020-09-30 09:01:52 +0200 | [diff] [blame] | 266 | UnpackReshaped = OperatorInfo(indices=IFM_INDICES) |
| 267 | Where = OperatorInfo() |
| 268 | While = OperatorInfo() |
| 269 | ZerosLike = OperatorInfo() |
| 270 | |
| 271 | @property |
| 272 | def info(self): |
| 273 | return self.value |
| 274 | |
| 275 | @property |
| 276 | def npu_block_type(self): |
| 277 | return self.info.block_type |
| 278 | |
| 279 | def is_conv2d_op(self): |
| 280 | return self.info.block_type == NpuBlockType.ConvolutionMxN |
| 281 | |
| 282 | def is_depthwise_conv2d_op(self): |
| 283 | return self.info.block_type == NpuBlockType.ConvolutionDepthWise |
| 284 | |
| 285 | def is_pool_op(self): |
| 286 | return self.info.block_type == NpuBlockType.Pooling |
| 287 | |
| 288 | def is_maxpool_op(self): |
| 289 | return self in (Op.MaxPool, Op.QuantizedMaxPool) |
| 290 | |
| 291 | def is_avgpool_op(self): |
| 292 | return self in (Op.QuantizedAvgPool, Op.AvgPool) |
| 293 | |
| 294 | def is_elementwise_op(self): |
| 295 | return self.info.block_type == NpuBlockType.ElementWise |
| 296 | |
| 297 | def is_unary_elementwise_op(self): |
| 298 | return self.info.block_type == NpuBlockType.ElementWise and self.info.is_unary |
| 299 | |
| 300 | def is_binary_elementwise_op(self): |
| 301 | return self.info.block_type == NpuBlockType.ElementWise and not self.info.is_unary |
| 302 | |
| 303 | def is_relu_op(self): |
Louis Verhaard | e8a5a78 | 2020-11-02 18:04:27 +0100 | [diff] [blame] | 304 | return self in (Op.Relu, Op.Relu6, Op.ReluN1To1, Op.Clip) |
Louis Verhaard | aee5d75 | 2020-09-30 09:01:52 +0200 | [diff] [blame] | 305 | |
| 306 | def is_activation_op(self): |
| 307 | return self.is_relu_op() or self in (Op.Tanh, Op.Sigmoid, Op.Softmax, Op.LUT) |
| 308 | |
| 309 | def is_split_op(self): |
Patrik Gustavsson | 2349d42 | 2020-12-01 16:02:29 +0100 | [diff] [blame] | 310 | return self in (Op.Split, Op.SplitV, Op.StridedSlice, Op.Slice, Op.UnpackReshaped, Op.Unpack) |
Louis Verhaard | aee5d75 | 2020-09-30 09:01:52 +0200 | [diff] [blame] | 311 | |
| 312 | def is_concat_op(self): |
Patrik Gustavsson | 2349d42 | 2020-12-01 16:02:29 +0100 | [diff] [blame] | 313 | return self in (Op.Concat, Op.ConcatTFLite, Op.PackReshaped, Op.Pack) |
Louis Verhaard | aee5d75 | 2020-09-30 09:01:52 +0200 | [diff] [blame] | 314 | |
| 315 | def needs_bias(self): |
| 316 | return bool(self.info.indices.biases) |
| 317 | |
Patrik Gustavsson | 2349d42 | 2020-12-01 16:02:29 +0100 | [diff] [blame] | 318 | def needs_shapes(self): |
| 319 | return bool(self.info.indices.ifms) |
| 320 | |
Louis Verhaard | aee5d75 | 2020-09-30 09:01:52 +0200 | [diff] [blame] | 321 | @classmethod |
| 322 | def op_set(cls, predicate): |
| 323 | # Returns the set of all operator codes that fulfill the given predicate |
| 324 | return {op_type for op_type in Op if predicate(op_type)} |
| 325 | |
| 326 | def __str__(self): |
| 327 | return self.name |
| 328 | |
| 329 | __repr__ = __str__ |
| 330 | |
| 331 | def __lt__(self, other): |
| 332 | return self.value.id < other.value.id |
| 333 | |
| 334 | |
Michael McGeagh | 1689548 | 2020-12-14 15:51:20 +0000 | [diff] [blame] | 335 | class Padding(Enum): |
| 336 | SAME = 0 |
| 337 | VALID = 1 |
Louis Verhaard | ae2d553 | 2020-12-11 17:19:54 +0100 | [diff] [blame] | 338 | EXPLICIT = 2 # Padding is specified in a PAD operation (only used for NPU operations) |
Michael McGeagh | 1689548 | 2020-12-14 15:51:20 +0000 | [diff] [blame] | 339 | |
| 340 | |
Louis Verhaard | e8a5a78 | 2020-11-02 18:04:27 +0100 | [diff] [blame] | 341 | class ActivationFunction: |
| 342 | """Fused activation function""" |
| 343 | |
| 344 | def __init__(self, op_type: Op): |
| 345 | self.op_type = op_type # The activation operation to be performed |
| 346 | # min/max are optional; if present they are non-quantized values |
| 347 | self.min: Optional[float] = None |
| 348 | self.max: Optional[float] = None |
| 349 | # Table lookup index, only applicable for Op.LUT activation, 0-7 |
| 350 | self.lut_index: int = 0 |
| 351 | |
| 352 | def clone(self): |
| 353 | res = copy.copy(self) |
| 354 | return res |
| 355 | |
| 356 | |
| 357 | def create_activation_function(op_type: Op) -> ActivationFunction: |
| 358 | """Creates activation function with min/max depending on op_type""" |
| 359 | act = ActivationFunction(op_type) |
| 360 | if op_type == Op.Relu: |
| 361 | act.min = 0.0 |
| 362 | elif op_type == Op.Relu6: |
| 363 | act.min = 0.0 |
| 364 | act.max = 6.0 |
| 365 | elif op_type == Op.ReluN1To1: |
| 366 | act.min = -1.0 |
| 367 | act.max = 1.0 |
| 368 | elif op_type == Op.Tanh: |
| 369 | act.min = -1.0 |
| 370 | act.max = 1.0 |
| 371 | elif op_type == Op.Sigmoid: |
| 372 | act.min = 0.0 |
| 373 | act.max = 1.0 |
| 374 | return act |
| 375 | |
| 376 | |
patrik.gustavsson | eeb8515 | 2020-12-21 17:10:40 +0000 | [diff] [blame] | 377 | def get_slice_offsets(input_shape: List[int], offset_tens: int, offset_mask: int, is_begin: bool = True): |
Louis Verhaard | fa2f92a | 2020-09-21 11:56:18 +0200 | [diff] [blame] | 378 | # For strided slice operator: get start or end offsets |
| 379 | offsets = len(input_shape) * [0] if is_begin else input_shape[:] |
| 380 | for idx in range(len(input_shape)): |
| 381 | # If the i:th bit in the mask is set then the value on offset_tens[i] should be ignored |
| 382 | if (offset_mask & (1 << idx)) == 0: |
| 383 | offsets[idx] = offset_tens.values[idx] |
| 384 | if offsets[idx] < 0: |
| 385 | # Convert offset to positive value |
| 386 | offsets[idx] += input_shape[idx] |
| 387 | return offsets |
| 388 | |
| 389 | |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 390 | class Operation: |
| 391 | """Class representing a Neural Network operation. Has a name, a type, |
Dwight Lidman | c6ac194 | 2020-10-02 14:55:45 +0200 | [diff] [blame] | 392 | input and output tensors, as well as an attribute dictionary.""" |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 393 | |
Fredrik Svedberg | a0c3624 | 2020-06-03 15:43:31 +0200 | [diff] [blame] | 394 | __slots__ = ( |
| 395 | "type", |
| 396 | "name", |
| 397 | "op_index", |
| 398 | "attrs", |
| 399 | "inputs", |
| 400 | "outputs", |
| 401 | "flops", |
| 402 | "scheduled_pass", |
| 403 | "run_on_npu", |
Louis Verhaard | aee5d75 | 2020-09-30 09:01:52 +0200 | [diff] [blame] | 404 | "activation", |
| 405 | "memory_function", |
| 406 | "forced_output_quantization", |
Fredrik Svedberg | a0c3624 | 2020-06-03 15:43:31 +0200 | [diff] [blame] | 407 | "activation_lut", |
Tim Hall | 4ed38bc | 2020-10-20 18:54:20 +0100 | [diff] [blame] | 408 | "_kernel", |
Patrik Gustavsson | 2349d42 | 2020-12-01 16:02:29 +0100 | [diff] [blame] | 409 | "ifm_shapes", |
| 410 | "ofm_shapes", |
Fredrik Svedberg | a0c3624 | 2020-06-03 15:43:31 +0200 | [diff] [blame] | 411 | ) |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 412 | |
Louis Verhaard | e8a5a78 | 2020-11-02 18:04:27 +0100 | [diff] [blame] | 413 | def __init__(self, op_type: Op, name: str): |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 414 | self.type = op_type |
| 415 | self.name = name |
Dwight Lidman | 9b43f84 | 2020-12-08 17:56:44 +0100 | [diff] [blame] | 416 | self.attrs: Dict[str, Any] = {} |
| 417 | self.inputs: List[Tensor] = [] |
| 418 | self.outputs: List[Tensor] = [] |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 419 | self.flops = 0 |
| 420 | self.run_on_npu = True |
Louis Verhaard | aee5d75 | 2020-09-30 09:01:52 +0200 | [diff] [blame] | 421 | # Fused activation function. If not none: operator code. |
Louis Verhaard | e8a5a78 | 2020-11-02 18:04:27 +0100 | [diff] [blame] | 422 | self.activation: Optional[ActivationFunction] = None |
Louis Verhaard | aee5d75 | 2020-09-30 09:01:52 +0200 | [diff] [blame] | 423 | # Fused memory function, if not None: operator code |
| 424 | self.memory_function = None |
| 425 | # If not none: contains QuantizationParameters to be used as output quantization |
| 426 | # (which overrides the ofm tensor's quantization), used in LUT |
| 427 | self.forced_output_quantization = None |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 428 | self.scheduled_pass = None |
Tim Hall | c8310b1 | 2020-06-17 14:53:11 +0100 | [diff] [blame] | 429 | self.op_index = None # input network operator index |
Fredrik Svedberg | a0c3624 | 2020-06-03 15:43:31 +0200 | [diff] [blame] | 430 | self.activation_lut = None |
Tim Hall | 4ed38bc | 2020-10-20 18:54:20 +0100 | [diff] [blame] | 431 | self._kernel = None |
patrik.gustavsson | eeb8515 | 2020-12-21 17:10:40 +0000 | [diff] [blame] | 432 | self.ifm_shapes: List[Shape4D] = [] |
| 433 | self.ofm_shapes: List[Shape4D] = [] |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 434 | |
| 435 | def clone(self, suffix="_clone"): |
| 436 | res = Operation(self.type, self.name + suffix) |
| 437 | |
| 438 | res.attrs = dict(self.attrs) |
| 439 | res.inputs = list(self.inputs) |
| 440 | res.outputs = list(self.outputs) |
| 441 | res.flops = self.flops |
Louis Verhaard | aee5d75 | 2020-09-30 09:01:52 +0200 | [diff] [blame] | 442 | res.run_on_npu = self.run_on_npu |
Louis Verhaard | e8a5a78 | 2020-11-02 18:04:27 +0100 | [diff] [blame] | 443 | res.activation = None if self.activation is None else self.activation.clone() |
Louis Verhaard | aee5d75 | 2020-09-30 09:01:52 +0200 | [diff] [blame] | 444 | res.memory_function = self.memory_function |
| 445 | res.forced_output_quantization = self.forced_output_quantization |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 446 | res.scheduled_pass = self.scheduled_pass |
Tim Hall | c8310b1 | 2020-06-17 14:53:11 +0100 | [diff] [blame] | 447 | res.op_index = None # not relevant as not part of input network |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 448 | |
| 449 | return res |
| 450 | |
| 451 | def __str__(self): |
Louis Verhaard | aee5d75 | 2020-09-30 09:01:52 +0200 | [diff] [blame] | 452 | return "<nng.Operation '{}' type={}>".format(self.name, self.type) |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 453 | |
| 454 | __repr__ = __str__ |
| 455 | |
Michael McGeagh | 65fd998 | 2020-10-20 11:49:28 +0100 | [diff] [blame] | 456 | def get_kernel_size(self): |
Tim Hall | 4ed38bc | 2020-10-20 18:54:20 +0100 | [diff] [blame] | 457 | weights = self.weights |
| 458 | if weights and self.type.npu_block_type in (NpuBlockType.ConvolutionDepthWise, NpuBlockType.ConvolutionMxN): |
| 459 | weight_shape = full_shape(4, weights.shape, 1) |
Michael McGeagh | 65fd998 | 2020-10-20 11:49:28 +0100 | [diff] [blame] | 460 | h = weight_shape[-4] |
| 461 | w = weight_shape[-3] |
Louis Verhaard | e8a5a78 | 2020-11-02 18:04:27 +0100 | [diff] [blame] | 462 | elif self.type.npu_block_type in (NpuBlockType.Pooling, NpuBlockType.ReduceSum) and "ksize" in self.attrs: |
| 463 | h, w = self.attrs["ksize"][1:3] |
Tim Hall | 4ed38bc | 2020-10-20 18:54:20 +0100 | [diff] [blame] | 464 | else: |
Michael McGeagh | 65fd998 | 2020-10-20 11:49:28 +0100 | [diff] [blame] | 465 | h = self.attrs.get("filter_height", 1) |
| 466 | w = self.attrs.get("filter_width", 1) |
| 467 | return w, h |
| 468 | |
| 469 | def get_kernel_stride(self): |
| 470 | if "strides" in self.attrs: |
| 471 | _, h, w, _ = self.attrs["strides"] |
| 472 | else: |
| 473 | h = self.attrs.get("stride_h", 1) |
| 474 | w = self.attrs.get("stride_w", 1) |
| 475 | return w, h |
| 476 | |
| 477 | def get_kernel_dilation(self): |
| 478 | if "dilation" in self.attrs: |
| 479 | _, h, w, _ = self.attrs["dilation"] |
| 480 | else: |
| 481 | h = self.attrs.get("dilation_h_factor", 1) |
| 482 | w = self.attrs.get("dilation_w_factor", 1) |
| 483 | return w, h |
| 484 | |
| 485 | @property |
| 486 | def kernel(self): |
| 487 | k_w, k_h = self.get_kernel_size() |
| 488 | s_w, s_h = self.get_kernel_stride() |
| 489 | d_w, d_h = self.get_kernel_dilation() |
| 490 | self._kernel = Kernel(k_w, k_h, s_w, s_h, d_w, d_h) |
Tim Hall | 4ed38bc | 2020-10-20 18:54:20 +0100 | [diff] [blame] | 491 | return self._kernel |
| 492 | |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 493 | def get_ifm_ifm2_weights_ofm(self): |
Louis Verhaard | aee5d75 | 2020-09-30 09:01:52 +0200 | [diff] [blame] | 494 | return self.ifm, self.ifm2, self.weights, self.ofm |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 495 | |
| 496 | def get_ifm_weights_biases_ofm(self): |
Louis Verhaard | aee5d75 | 2020-09-30 09:01:52 +0200 | [diff] [blame] | 497 | return self.ifm, self.weights, self.bias, self.ofm |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 498 | |
Jacob Bohlin | 49d9212 | 2020-08-19 14:36:46 +0200 | [diff] [blame] | 499 | def get_ifm_ifm2_weights_biases_ofm(self): |
Louis Verhaard | aee5d75 | 2020-09-30 09:01:52 +0200 | [diff] [blame] | 500 | return self.ifm, self.ifm2, self.weights, self.bias, self.ofm |
Jacob Bohlin | 49d9212 | 2020-08-19 14:36:46 +0200 | [diff] [blame] | 501 | |
Louis Verhaard | aee5d75 | 2020-09-30 09:01:52 +0200 | [diff] [blame] | 502 | def get_ifm_ofm(self): |
| 503 | return self.ifm, self.ofm |
Jacob Bohlin | 49d9212 | 2020-08-19 14:36:46 +0200 | [diff] [blame] | 504 | |
Louis Verhaard | aee5d75 | 2020-09-30 09:01:52 +0200 | [diff] [blame] | 505 | @property |
| 506 | def ifm(self): |
| 507 | # Gets the IFM tensor, or None if not applicable |
| 508 | return self.get_input(self.type.info.indices.ifms, 0) |
Jacob Bohlin | 49d9212 | 2020-08-19 14:36:46 +0200 | [diff] [blame] | 509 | |
Louis Verhaard | aee5d75 | 2020-09-30 09:01:52 +0200 | [diff] [blame] | 510 | @property |
| 511 | def ifm2(self): |
| 512 | # Gets the IFM2 tensor, or None if not applicable |
| 513 | return self.get_input(self.type.info.indices.ifms, 1) |
Louis Verhaard | 98a3499 | 2020-09-01 10:39:04 +0200 | [diff] [blame] | 514 | |
Louis Verhaard | aee5d75 | 2020-09-30 09:01:52 +0200 | [diff] [blame] | 515 | @property |
| 516 | def bias(self): |
| 517 | # Gets the bias tensor, or None if not applicable |
| 518 | return self.get_input(self.type.info.indices.biases, 0) |
| 519 | |
| 520 | @property |
| 521 | def weights(self): |
| 522 | # Gets the weight tensor, or None if not applicable |
| 523 | return self.get_input(self.type.info.indices.weights, 0) |
| 524 | |
| 525 | def get_ifm_tensors(self): |
| 526 | # Gets the IFM tensors, or empty list if not applicable |
| 527 | return self._index_list_to_tensors(self.type.info.indices.ifms) |
| 528 | |
| 529 | def get_weight_tensors(self): |
| 530 | # Gets the weight tensors, or empty list if not applicable |
| 531 | return self._index_list_to_tensors(self.type.info.indices.weights) |
| 532 | |
| 533 | def get_bias_tensors(self): |
| 534 | # Gets the bias tensors, or empty list if not applicable |
| 535 | return self._index_list_to_tensors(self.type.info.indices.biases) |
| 536 | |
| 537 | def _index_list_to_tensors(self, index_list): |
| 538 | return [self.inputs[ix] for ix in index_list if ix < len(self.inputs)] |
| 539 | |
| 540 | def get_input(self, index_list, ix): |
| 541 | if ix >= len(index_list): |
| 542 | return None |
| 543 | if index_list[ix] >= len(self.inputs): |
| 544 | return None |
| 545 | return self.inputs[index_list[ix]] |
| 546 | |
| 547 | @property |
| 548 | def ofm(self): |
| 549 | # Gets the OFM tensor, or None if not applicable |
| 550 | return self.outputs[0] if self.outputs else None |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 551 | |
| 552 | def get_concat_inputs_axis(self): |
Louis Verhaard | aee5d75 | 2020-09-30 09:01:52 +0200 | [diff] [blame] | 553 | assert self.type.is_concat_op() |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 554 | |
Louis Verhaard | aee5d75 | 2020-09-30 09:01:52 +0200 | [diff] [blame] | 555 | if self.type == Op.Concat: |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 556 | axis_tensor = self.inputs[0] |
| 557 | inputs = self.inputs[1:] |
Louis Verhaard | aee5d75 | 2020-09-30 09:01:52 +0200 | [diff] [blame] | 558 | elif self.type == Op.ConcatTFLite: |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 559 | inputs = self.inputs |
| 560 | axis = self.attrs["axis"] |
Louis Verhaard | aee5d75 | 2020-09-30 09:01:52 +0200 | [diff] [blame] | 561 | elif self.type == Op.PackReshaped: |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 562 | # Requires fixup_pack_input to be called before this point |
| 563 | inputs = self.inputs |
| 564 | axis = self.attrs["axis"] |
| 565 | assert len(self.inputs) == self.attrs["values_count"] |
| 566 | else: |
Louis Verhaard | aee5d75 | 2020-09-30 09:01:52 +0200 | [diff] [blame] | 567 | assert len(axis_tensor.ops) == 1 and axis_tensor.ops[0].type == Op.Const |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 568 | axis = int(axis_tensor.values) |
| 569 | |
| 570 | return inputs, axis |
| 571 | |
Louis Verhaard | b2fb212 | 2020-06-04 15:51:24 +0200 | [diff] [blame] | 572 | def get_dilation_h_w(self): |
| 573 | _, dilation_h, dilation_w, _ = self.attrs.get("dilation", (1, 1, 1, 1)) |
| 574 | return dilation_h, dilation_w |
| 575 | |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 576 | def get_split_inputs_axis(self): |
Louis Verhaard | aee5d75 | 2020-09-30 09:01:52 +0200 | [diff] [blame] | 577 | assert self.type.is_split_op() |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 578 | |
| 579 | offset_start = None |
| 580 | offset_end = None |
| 581 | axis = None |
Louis Verhaard | aee5d75 | 2020-09-30 09:01:52 +0200 | [diff] [blame] | 582 | if self.type == Op.Split: |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 583 | num_splits = self.attrs.get("num_splits") |
| 584 | axis_tens = self.inputs[0] |
Louis Verhaard | aee5d75 | 2020-09-30 09:01:52 +0200 | [diff] [blame] | 585 | assert len(axis_tens.ops) == 1 and axis_tens.ops[0].type == Op.Const |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 586 | axis = int(axis_tens.values) |
| 587 | input_tens = self.inputs[1] |
| 588 | outputs = self.outputs |
| 589 | assert num_splits == len(outputs) |
| 590 | |
Louis Verhaard | aee5d75 | 2020-09-30 09:01:52 +0200 | [diff] [blame] | 591 | elif self.type == Op.SplitV: |
Charles Xu | 53d4752 | 2020-05-04 11:32:05 +0200 | [diff] [blame] | 592 | num_splits = self.attrs.get("num_splits") |
| 593 | input_tens = self.inputs[0] |
| 594 | size_tens = self.inputs[1] |
Louis Verhaard | aee5d75 | 2020-09-30 09:01:52 +0200 | [diff] [blame] | 595 | assert len(size_tens.ops) == 1 and size_tens.ops[0].type == Op.Const |
Charles Xu | 53d4752 | 2020-05-04 11:32:05 +0200 | [diff] [blame] | 596 | sizes = size_tens.values |
Patrik Gustavsson | 271ddc3 | 2020-09-01 09:15:27 +0200 | [diff] [blame] | 597 | |
Charles Xu | 53d4752 | 2020-05-04 11:32:05 +0200 | [diff] [blame] | 598 | axis_tens = self.inputs[2] |
Louis Verhaard | aee5d75 | 2020-09-30 09:01:52 +0200 | [diff] [blame] | 599 | assert len(axis_tens.ops) == 1 and axis_tens.ops[0].type == Op.Const |
Charles Xu | 53d4752 | 2020-05-04 11:32:05 +0200 | [diff] [blame] | 600 | axis = int(axis_tens.values) |
Patrik Gustavsson | 271ddc3 | 2020-09-01 09:15:27 +0200 | [diff] [blame] | 601 | |
| 602 | for idx, size in enumerate(sizes): |
| 603 | # One but only one size might be set to -1, indicating that size should be inferred |
| 604 | if size == -1: |
| 605 | sizes[idx] = input_tens.shape[axis] - (sum(sizes) + 1) |
| 606 | break |
| 607 | |
Charles Xu | 53d4752 | 2020-05-04 11:32:05 +0200 | [diff] [blame] | 608 | outputs = self.outputs |
| 609 | assert num_splits == len(outputs) |
| 610 | assert sum(sizes) == input_tens.shape[axis] |
| 611 | |
Louis Verhaard | aee5d75 | 2020-09-30 09:01:52 +0200 | [diff] [blame] | 612 | elif self.type == Op.Slice: |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 613 | input_tens, begin_tens, size_tens = self.inputs |
| 614 | outputs = self.outputs |
| 615 | offset_start = [0] * len(input_tens.shape) |
| 616 | offset_end = [0] * len(input_tens.shape) |
| 617 | |
| 618 | for idx in range(len(begin_tens.values)): |
| 619 | # Check if the op should slice in dimension idx |
| 620 | if size_tens.values[idx] != input_tens.shape[idx]: |
| 621 | offset_start[idx] = begin_tens.values[idx] |
| 622 | offset_end[idx] = size_tens.values[idx] + offset_start[idx] |
| 623 | |
Louis Verhaard | aee5d75 | 2020-09-30 09:01:52 +0200 | [diff] [blame] | 624 | elif self.type == Op.StridedSlice: |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 625 | input_tens, begin_tens, end_tens, strides_tens = self.inputs |
| 626 | outputs = self.outputs |
| 627 | out_tens = outputs[0] |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 628 | |
| 629 | # Extract masks |
| 630 | begin_mask = self.attrs["begin_mask"] |
| 631 | ellipsis_mask = self.attrs["ellipsis_mask"] |
| 632 | end_mask = self.attrs["end_mask"] |
| 633 | new_axis_mask = self.attrs["new_axis_mask"] |
| 634 | shrink_axis_mask = self.attrs["shrink_axis_mask"] |
Patrik Gustavsson | cf72890 | 2020-04-30 08:57:23 +0200 | [diff] [blame] | 635 | |
| 636 | # shrink_axis_mask/new_axis_mask/ellipsis_mask is not supported by the Operation class but the operation |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 637 | # may have the attribute modified and handled in the graph optimization phase. |
Patrik Gustavsson | cf72890 | 2020-04-30 08:57:23 +0200 | [diff] [blame] | 638 | assert shrink_axis_mask == new_axis_mask == ellipsis_mask == 0 |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 639 | assert len(input_tens.shape) == len(out_tens.shape) |
Louis Verhaard | fa2f92a | 2020-09-21 11:56:18 +0200 | [diff] [blame] | 640 | offset_start = get_slice_offsets(input_tens.shape, begin_tens, begin_mask, is_begin=True) |
| 641 | offset_end = get_slice_offsets(input_tens.shape, end_tens, end_mask, is_begin=False) |
Louis Verhaard | aee5d75 | 2020-09-30 09:01:52 +0200 | [diff] [blame] | 642 | elif self.type == Op.UnpackReshaped: |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 643 | # Requires fixup_unpack_output to be called before this point |
| 644 | input_tens = self.inputs[0] |
| 645 | outputs = self.outputs |
| 646 | axis = self.attrs["axis"] |
| 647 | num_splits = self.attrs["num"] |
| 648 | # Number of outputs have to equal the value of the dimension to unpack |
| 649 | assert num_splits == len(outputs) == input_tens.shape[axis] |
| 650 | else: |
| 651 | assert False |
| 652 | |
| 653 | return input_tens, outputs, axis, offset_start, offset_end |
Fredrik Svedberg | a0c3624 | 2020-06-03 15:43:31 +0200 | [diff] [blame] | 654 | |
| 655 | def set_activation_lut(self, lut_tensor): |
Louis Verhaard | e8a5a78 | 2020-11-02 18:04:27 +0100 | [diff] [blame] | 656 | self.activation = ActivationFunction(Op.LUT) |
Fredrik Svedberg | a0c3624 | 2020-06-03 15:43:31 +0200 | [diff] [blame] | 657 | self.activation_lut = lut_tensor |
Michael McGeagh | c5b549b | 2020-08-07 11:54:28 +0100 | [diff] [blame] | 658 | self.add_input_tensor(lut_tensor) |
Michael McGeagh | 5778ffd | 2020-08-06 17:31:02 +0100 | [diff] [blame] | 659 | |
| 660 | def add_input_tensor(self, tens): |
| 661 | self.inputs.append(tens) |
| 662 | if self not in tens.consumer_list: |
| 663 | tens.consumer_list.append(self) |
| 664 | |
Jacob Bohlin | 67e0d8f | 2020-08-20 10:53:02 +0200 | [diff] [blame] | 665 | def set_input_tensor(self, tens, idx): |
| 666 | tens_to_remove = self.inputs[idx] |
| 667 | if tens_to_remove in tens.consumer_list: |
| 668 | tens.consumer_list.remove(tens_to_remove) |
| 669 | |
| 670 | self.inputs[idx] = tens |
| 671 | if self not in tens.consumer_list: |
| 672 | tens.consumer_list.append(self) |
| 673 | |
Michael McGeagh | 5778ffd | 2020-08-06 17:31:02 +0100 | [diff] [blame] | 674 | def set_output_tensor(self, tens): |
| 675 | tens.ops = [self] |
| 676 | self.outputs = [tens] |
Jacob Bohlin | a41cd4d | 2020-08-26 18:21:28 +0200 | [diff] [blame] | 677 | |
Louis Verhaard | 98a3499 | 2020-09-01 10:39:04 +0200 | [diff] [blame] | 678 | def get_output_quantization(self): |
Louis Verhaard | aee5d75 | 2020-09-30 09:01:52 +0200 | [diff] [blame] | 679 | if self.forced_output_quantization is not None: |
| 680 | return self.forced_output_quantization |
| 681 | return self.ofm.quantization |
Michael McGeagh | 528a56d | 2020-12-16 11:33:21 +0000 | [diff] [blame] | 682 | |
| 683 | def error(self, msg): |
| 684 | """ |
| 685 | Raises a VelaError exception for errors encountered when parsing an Operation |
| 686 | |
| 687 | :param self: Operation object that resulted in the error |
| 688 | :param msg: str object that contains a description of the specific error encountered |
| 689 | """ |
| 690 | |
| 691 | def _print_tensors(tensors): |
| 692 | lines = [] |
| 693 | for idx, tens in enumerate(tensors): |
| 694 | tens_name = getattr(tens, "name", "Not a Tensor") |
| 695 | lines.append(f" {idx} = {tens_name}") |
| 696 | return lines |
| 697 | |
| 698 | if self.op_index is None: |
| 699 | lines = [f"Invalid {self.type} (name = {self.name}) operator in the internal representation. {msg}"] |
| 700 | else: |
| 701 | lines = [f"Invalid {self.type} (op_index = {self.op_index}) operator in the input network. {msg}"] |
| 702 | |
| 703 | lines += [" Input tensors:"] |
| 704 | lines += _print_tensors(self.inputs) |
| 705 | |
| 706 | lines += [" Output tensors:"] |
| 707 | lines += _print_tensors(self.outputs) |
| 708 | |
| 709 | raise VelaError("\n".join(lines)) |
Patrik Gustavsson | 2349d42 | 2020-12-01 16:02:29 +0100 | [diff] [blame] | 710 | |
| 711 | def set_ifm_ofm_shapes(self): |
patrik.gustavsson | eeb8515 | 2020-12-21 17:10:40 +0000 | [diff] [blame] | 712 | self.ifm_shapes = [] |
| 713 | self.ofm_shapes = [] |
| 714 | |
Patrik Gustavsson | 2349d42 | 2020-12-01 16:02:29 +0100 | [diff] [blame] | 715 | ifm_tensor, ifm2_tensor, weight_tensor, ofm_tensor = self.get_ifm_ifm2_weights_ofm() |
| 716 | |
| 717 | # set all shapes to op, as 4D |
| 718 | if self.type == Op.FullyConnected: |
| 719 | n_in_elems = weight_tensor.shape[-2] |
| 720 | elms = ifm_tensor.elements() |
| 721 | batch_size = elms // n_in_elems |
| 722 | assert batch_size * n_in_elems == elms |
| 723 | |
patrik.gustavsson | eeb8515 | 2020-12-21 17:10:40 +0000 | [diff] [blame] | 724 | self.ifm_shapes.append(Shape4D([batch_size, 1, 1, n_in_elems])) |
| 725 | self.ofm_shapes.append(Shape4D(ofm_tensor.get_full_shape())) |
Patrik Gustavsson | 2349d42 | 2020-12-01 16:02:29 +0100 | [diff] [blame] | 726 | elif self.type == Op.Softmax: |
patrik.gustavsson | eeb8515 | 2020-12-21 17:10:40 +0000 | [diff] [blame] | 727 | self.ifm_shapes.append(Shape4D(ifm_tensor.get_full_shape())) |
| 728 | self.ofm_shapes.append(Shape4D(ofm_tensor.get_full_shape())) |
Patrik Gustavsson | 2349d42 | 2020-12-01 16:02:29 +0100 | [diff] [blame] | 729 | elif self.type.is_split_op or self.type.is_concat_op(): |
| 730 | for inp in self.inputs: |
| 731 | if inp is not None: |
patrik.gustavsson | eeb8515 | 2020-12-21 17:10:40 +0000 | [diff] [blame] | 732 | self.ifm_shapes.append(Shape4D(full_shape(4, inp.shape, 1))) |
Patrik Gustavsson | 2349d42 | 2020-12-01 16:02:29 +0100 | [diff] [blame] | 733 | else: |
| 734 | self.ifm_shapes.append(None) |
| 735 | for out in self.outputs: |
| 736 | if out is not None: |
patrik.gustavsson | eeb8515 | 2020-12-21 17:10:40 +0000 | [diff] [blame] | 737 | self.ofm_shapes.append(Shape4D(full_shape(4, out.shape, 1))) |
Patrik Gustavsson | 2349d42 | 2020-12-01 16:02:29 +0100 | [diff] [blame] | 738 | else: |
| 739 | self.ofm_shapes.append(None) |
| 740 | else: |
patrik.gustavsson | eeb8515 | 2020-12-21 17:10:40 +0000 | [diff] [blame] | 741 | self.ifm_shapes.append(Shape4D(full_shape(4, ifm_tensor.shape, 1))) |
Patrik Gustavsson | 2349d42 | 2020-12-01 16:02:29 +0100 | [diff] [blame] | 742 | if ifm2_tensor is not None: |
patrik.gustavsson | eeb8515 | 2020-12-21 17:10:40 +0000 | [diff] [blame] | 743 | self.ifm_shapes.append(Shape4D(full_shape(4, ifm2_tensor.shape, 1))) |
| 744 | self.ofm_shapes.append(Shape4D(full_shape(4, ofm_tensor.shape, 1))) |