Sadik Armagan | 8f397a1 | 2022-06-17 15:38:22 +0100 | [diff] [blame] | 1 | // |
Kevin May | 28c3d0f | 2023-07-04 09:07:30 +0100 | [diff] [blame] | 2 | // Copyright © 2022-2023 Arm Ltd and Contributors. All rights reserved. |
Sadik Armagan | 8f397a1 | 2022-06-17 15:38:22 +0100 | [diff] [blame] | 3 | // SPDX-License-Identifier: MIT |
| 4 | // |
| 5 | |
| 6 | #include "Converter.hpp" |
| 7 | #include <half/half.hpp> |
Jim Flynn | 987398f | 2023-09-16 18:53:52 +0100 | [diff] [blame] | 8 | #include <armnn/Exceptions.hpp> |
Sadik Armagan | 8f397a1 | 2022-06-17 15:38:22 +0100 | [diff] [blame] | 9 | #include <armnnUtils/TensorUtils.hpp> |
| 10 | |
| 11 | namespace armnn_driver |
| 12 | { |
| 13 | |
| 14 | using namespace android::nn; |
| 15 | using Half = half_float::half; |
| 16 | |
| 17 | namespace |
| 18 | { |
| 19 | |
| 20 | } // anonymouse namespace |
| 21 | |
| 22 | bool Converter::ConvertOperation(const Operation& operation, const Model& model, ConversionData& data) |
| 23 | { |
| 24 | switch (operation.type) |
| 25 | { |
| 26 | case OperationType::ABS: |
| 27 | return ConvertElementwiseUnary(operation, model, data, UnaryOperation::Abs); |
| 28 | case OperationType::ADD: |
Kevin May | 28c3d0f | 2023-07-04 09:07:30 +0100 | [diff] [blame] | 29 | return ConvertElementwiseBinary(operation, model, data, armnn::BinaryOperation::Add); |
Sadik Armagan | 8f397a1 | 2022-06-17 15:38:22 +0100 | [diff] [blame] | 30 | case OperationType::ARGMAX: |
| 31 | return ConvertArgMinMax(operation, model, data, ArgMinMaxFunction::Max); |
| 32 | case OperationType::ARGMIN: |
| 33 | return ConvertArgMinMax(operation, model, data, ArgMinMaxFunction::Min); |
| 34 | case OperationType::AVERAGE_POOL_2D: |
| 35 | return ConvertAveragePool2d(operation, model, data); |
Kevin May | 9636a9b | 2022-09-21 15:41:41 +0100 | [diff] [blame] | 36 | case OperationType::BATCH_MATMUL: |
| 37 | return ConvertBatchMatMul(operation, model, data); |
Sadik Armagan | 8f397a1 | 2022-06-17 15:38:22 +0100 | [diff] [blame] | 38 | case OperationType::BATCH_TO_SPACE_ND: |
| 39 | return ConvertBatchToSpaceNd(operation, model, data); |
| 40 | case OperationType::CAST: |
| 41 | return ConvertCast(operation, model, data); |
| 42 | case OperationType::CONCATENATION: |
| 43 | return ConvertConcatenation(operation, model, data); |
| 44 | case OperationType::CONV_2D: |
| 45 | return ConvertConv2d(operation, model, data); |
| 46 | case OperationType::DEPTH_TO_SPACE: |
| 47 | return ConvertDepthToSpace(operation, model, data); |
| 48 | case OperationType::DEPTHWISE_CONV_2D: |
| 49 | return ConvertDepthwiseConv2d(operation, model, data); |
| 50 | case OperationType::DEQUANTIZE: |
| 51 | return ConvertDequantize(operation, model, data); |
| 52 | case OperationType::DIV: |
Kevin May | 28c3d0f | 2023-07-04 09:07:30 +0100 | [diff] [blame] | 53 | return ConvertElementwiseBinary(operation, model, data, armnn::BinaryOperation::Div); |
Sadik Armagan | 8f397a1 | 2022-06-17 15:38:22 +0100 | [diff] [blame] | 54 | case OperationType::ELU: |
| 55 | return ConvertElu(operation, model, data); |
| 56 | case OperationType::EQUAL: |
| 57 | return ConvertComparison(operation, model, data, ComparisonOperation::Equal); |
| 58 | case OperationType::EXP: |
| 59 | return ConvertElementwiseUnary(operation, model, data, UnaryOperation::Exp); |
| 60 | case OperationType::EXPAND_DIMS: |
| 61 | return ConvertExpandDims(operation, model, data); |
| 62 | case OperationType::FILL: |
| 63 | return ConvertFill(operation, model, data); |
| 64 | case OperationType::FLOOR: |
| 65 | return ConvertFloor(operation, model, data); |
| 66 | case OperationType::FULLY_CONNECTED: |
| 67 | return ConvertFullyConnected(operation, model, data); |
| 68 | case OperationType::GATHER: |
| 69 | return ConvertGather(operation, model, data); |
| 70 | case OperationType::GREATER: |
| 71 | return ConvertComparison(operation, model, data, ComparisonOperation::Greater); |
| 72 | case OperationType::GREATER_EQUAL: |
| 73 | return ConvertComparison(operation, model, data, ComparisonOperation::GreaterOrEqual); |
| 74 | case OperationType::GROUPED_CONV_2D: |
| 75 | return ConvertGroupedConv2d(operation, model, data); |
| 76 | case OperationType::HARD_SWISH: |
| 77 | return ConvertHardSwish(operation, model, data); |
| 78 | case OperationType::INSTANCE_NORMALIZATION: |
| 79 | return ConvertInstanceNormalization(operation, model, data); |
| 80 | case OperationType::L2_NORMALIZATION: |
| 81 | return ConvertL2Normalization(operation, model, data); |
| 82 | case OperationType::L2_POOL_2D: |
| 83 | return ConvertL2Pool2d(operation, model, data); |
| 84 | case OperationType::LESS: |
| 85 | return ConvertComparison(operation, model, data, ComparisonOperation::Less); |
| 86 | case OperationType::LESS_EQUAL: |
| 87 | return ConvertComparison(operation, model, data, ComparisonOperation::LessOrEqual); |
| 88 | case OperationType::LOCAL_RESPONSE_NORMALIZATION: |
| 89 | return ConvertLocalResponseNormalization(operation, model, data); |
| 90 | case OperationType::LOG: |
| 91 | return ConvertElementwiseUnary(operation, model, data, UnaryOperation::Log); |
| 92 | case OperationType::LOGICAL_AND: |
| 93 | return ConvertLogicalBinary(operation, model, data, LogicalBinaryOperation::LogicalAnd); |
| 94 | case OperationType::LOGICAL_NOT: |
| 95 | return ConvertElementwiseUnary(operation, model, data, UnaryOperation::LogicalNot); |
| 96 | case OperationType::LOGICAL_OR: |
| 97 | return ConvertLogicalBinary(operation, model, data, LogicalBinaryOperation::LogicalOr); |
| 98 | case OperationType::LOGISTIC: |
| 99 | return ConvertLogistic(operation, model, data); |
| 100 | case OperationType::LOG_SOFTMAX: |
| 101 | return ConvertLogSoftmax(operation, model, data); |
| 102 | case OperationType::LSTM: |
| 103 | return ConvertLstm(operation, model, data); |
| 104 | case OperationType::MAX_POOL_2D: |
| 105 | return ConvertMaxPool2d(operation, model, data); |
| 106 | case OperationType::MAXIMUM: |
Kevin May | 28c3d0f | 2023-07-04 09:07:30 +0100 | [diff] [blame] | 107 | return ConvertElementwiseBinary(operation, model, data, armnn::BinaryOperation::Maximum); |
Sadik Armagan | 8f397a1 | 2022-06-17 15:38:22 +0100 | [diff] [blame] | 108 | case OperationType::MEAN: |
| 109 | return ConvertMean(operation, model, data); |
| 110 | case OperationType::MINIMUM: |
Kevin May | 28c3d0f | 2023-07-04 09:07:30 +0100 | [diff] [blame] | 111 | return ConvertElementwiseBinary(operation, model, data, armnn::BinaryOperation::Minimum); |
Sadik Armagan | 8f397a1 | 2022-06-17 15:38:22 +0100 | [diff] [blame] | 112 | case OperationType::MUL: |
Kevin May | 28c3d0f | 2023-07-04 09:07:30 +0100 | [diff] [blame] | 113 | return ConvertElementwiseBinary(operation, model, data, armnn::BinaryOperation::Mul); |
Sadik Armagan | 8f397a1 | 2022-06-17 15:38:22 +0100 | [diff] [blame] | 114 | case OperationType::NEG: |
| 115 | return ConvertElementwiseUnary(operation, model, data, UnaryOperation::Neg); |
| 116 | case OperationType::NOT_EQUAL: |
| 117 | return ConvertComparison(operation, model, data, ComparisonOperation::NotEqual); |
| 118 | case OperationType::PAD: |
| 119 | return ConvertPad(operation, model, data); |
| 120 | case OperationType::PAD_V2: |
| 121 | return ConvertPadV2(operation, model, data); |
| 122 | case OperationType::PRELU: |
| 123 | return ConvertPrelu(operation, model, data); |
Kevin May | 28c3d0f | 2023-07-04 09:07:30 +0100 | [diff] [blame] | 124 | case OperationType::POW: |
| 125 | return ConvertElementwiseBinary(operation, model, data, BinaryOperation::Power); |
Sadik Armagan | 8f397a1 | 2022-06-17 15:38:22 +0100 | [diff] [blame] | 126 | case OperationType::QUANTIZE: |
| 127 | return ConvertQuantize(operation, model, data); |
| 128 | case OperationType::QUANTIZED_LSTM: |
| 129 | return ConvertQuantizedLstm(operation, model, data); |
| 130 | case OperationType::QUANTIZED_16BIT_LSTM: |
| 131 | return ConvertQuantized16BitLstm(operation, model, data); |
| 132 | case OperationType::RANK: |
| 133 | return ConvertRank(operation, model, data); |
| 134 | case OperationType::REDUCE_MAX: |
| 135 | return ConvertReduce(operation, model, data, armnn::ReduceOperation::Max); |
| 136 | case OperationType::REDUCE_MIN: |
| 137 | return ConvertReduce(operation, model, data, armnn::ReduceOperation::Min); |
Kevin May | 28c3d0f | 2023-07-04 09:07:30 +0100 | [diff] [blame] | 138 | case OperationType::REDUCE_PROD: |
| 139 | return ConvertReduce(operation, model, data, armnn::ReduceOperation::Prod); |
Sadik Armagan | 8f397a1 | 2022-06-17 15:38:22 +0100 | [diff] [blame] | 140 | case OperationType::REDUCE_SUM: |
| 141 | return ConvertReduce(operation, model, data, armnn::ReduceOperation::Sum); |
| 142 | case OperationType::RELU: |
| 143 | return ConvertReLu(operation, model, data); |
| 144 | case OperationType::RELU1: |
| 145 | return ConvertReLu1(operation, model, data); |
| 146 | case OperationType::RELU6: |
| 147 | return ConvertReLu6(operation, model, data); |
| 148 | case OperationType::RESHAPE: |
| 149 | return ConvertReshape(operation, model, data); |
| 150 | case OperationType::RESIZE_BILINEAR: |
| 151 | return ConvertResize(operation, model, data, ResizeMethod::Bilinear); |
| 152 | case OperationType::RESIZE_NEAREST_NEIGHBOR: |
| 153 | return ConvertResize(operation, model, data, ResizeMethod::NearestNeighbor); |
Teresa Charlin | ec0959b | 2023-07-27 13:23:15 +0100 | [diff] [blame] | 154 | case OperationType::REVERSE: |
| 155 | return ConvertReverseV2(operation, model, data); |
Sadik Armagan | 8f397a1 | 2022-06-17 15:38:22 +0100 | [diff] [blame] | 156 | case OperationType::RSQRT: |
| 157 | return ConvertElementwiseUnary(operation, model, data, UnaryOperation::Rsqrt); |
| 158 | case OperationType::SIN: |
| 159 | return ConvertElementwiseUnary(operation, model, data, UnaryOperation::Sin); |
| 160 | case OperationType::SOFTMAX: |
| 161 | return ConvertSoftmax(operation, model, data); |
| 162 | case OperationType::SPACE_TO_BATCH_ND : |
| 163 | return ConvertSpaceToBatchNd(operation, model, data); |
| 164 | case OperationType::SPACE_TO_DEPTH: |
| 165 | return ConvertSpaceToDepth(operation, model, data); |
| 166 | case OperationType::SQRT: |
| 167 | return ConvertSqrt(operation, model, data); |
| 168 | case OperationType::SQUEEZE: |
| 169 | return ConvertSqueeze(operation, model, data); |
| 170 | case OperationType::STRIDED_SLICE: |
| 171 | return ConvertStridedSlice(operation, model, data); |
| 172 | case OperationType::SUB: |
Kevin May | 28c3d0f | 2023-07-04 09:07:30 +0100 | [diff] [blame] | 173 | return ConvertElementwiseBinary(operation, model, data, BinaryOperation::Sub); |
Teresa Charlin | a0cc6ff | 2023-07-27 14:11:54 +0100 | [diff] [blame] | 174 | case OperationType::TILE: |
| 175 | return ConvertTile(operation, model, data); |
Sadik Armagan | 8f397a1 | 2022-06-17 15:38:22 +0100 | [diff] [blame] | 176 | case OperationType::TRANSPOSE: |
| 177 | return ConvertTranspose(operation, model, data); |
| 178 | case OperationType::TRANSPOSE_CONV_2D: |
| 179 | return ConvertTransposeConv2d(operation, model, data); |
| 180 | case OperationType::TANH: |
| 181 | return ConvertTanH(operation, model, data); |
| 182 | default: |
| 183 | VLOG(DRIVER) << "Operation type: " << operation.type << "is not supported in ArmnnDriver"; |
| 184 | return false; |
| 185 | } |
| 186 | } |
| 187 | |
Sadik Armagan | 8f397a1 | 2022-06-17 15:38:22 +0100 | [diff] [blame] | 188 | bool Converter::ConvertArgMinMax(const Operation& operation, |
| 189 | const Model& model, |
| 190 | ConversionData& data, |
| 191 | armnn::ArgMinMaxFunction argMinMaxFunction) |
| 192 | { |
| 193 | VLOG(DRIVER) << "Converter::ConvertArgMinMax()"; |
| 194 | VLOG(DRIVER) << "argMinMaxFunction = " << GetArgMinMaxFunctionAsCString(argMinMaxFunction); |
| 195 | |
| 196 | LayerInputHandle input0 = ConvertToLayerInputHandle(operation, 0, model, data); |
| 197 | |
| 198 | if (!input0.IsValid()) |
| 199 | { |
| 200 | return Fail("%s: Operation has invalid inputs", __func__); |
| 201 | } |
| 202 | |
| 203 | int32_t axis; |
| 204 | if (!GetInputScalar(operation, 1, OperandType::INT32, axis, model, data)) |
| 205 | { |
| 206 | return Fail("%s: Operation has invalid inputs. Failed to read axis.", __func__); |
| 207 | } |
| 208 | |
| 209 | const armnn::TensorInfo& inputInfo = input0.GetTensorInfo(); |
| 210 | int rank = static_cast<int>(inputInfo.GetNumDimensions()); |
| 211 | |
| 212 | if (((axis < -rank) && (axis < 0)) || ((axis >= rank) && (axis > 0))) |
| 213 | { |
| 214 | // Square bracket denotes inclusive n while parenthesis denotes exclusive n |
| 215 | // E.g. Rank 4 tensor can have axis in range [-4, 3) |
| 216 | // -1 == 3, -2 == 2, -3 == 1, -4 == 0 |
| 217 | return Fail("%s: Axis must be in range [-n, n)", __func__); |
| 218 | } |
| 219 | |
| 220 | const Operand* output = GetOutputOperand(operation, 0, model); |
| 221 | if (!output) |
| 222 | { |
| 223 | return Fail("%s: Could not read output 0", __func__); |
| 224 | } |
| 225 | |
| 226 | const armnn::TensorInfo& inputInfo0 = input0.GetTensorInfo(); |
| 227 | |
| 228 | const armnn::TensorInfo& outputInfo = GetTensorInfoForOperand(*output); |
| 229 | |
| 230 | armnn::ArgMinMaxDescriptor descriptor; |
| 231 | descriptor.m_Function = argMinMaxFunction; |
| 232 | descriptor.m_Axis = axis; |
| 233 | |
| 234 | bool isSupported = false; |
Cathal Corbett | 5383767 | 2022-09-01 11:34:37 +0100 | [diff] [blame] | 235 | armnn::BackendId setBackend; |
Sadik Armagan | 8f397a1 | 2022-06-17 15:38:22 +0100 | [diff] [blame] | 236 | auto validateFunc = [&](const armnn::TensorInfo& outputInfo, bool& isSupported) |
| 237 | { |
| 238 | FORWARD_LAYER_SUPPORT_FUNC(__func__, |
| 239 | IsArgMinMaxSupported, |
| 240 | data.m_Backends, |
| 241 | isSupported, |
Cathal Corbett | 5383767 | 2022-09-01 11:34:37 +0100 | [diff] [blame] | 242 | setBackend, |
Sadik Armagan | 8f397a1 | 2022-06-17 15:38:22 +0100 | [diff] [blame] | 243 | inputInfo0, |
| 244 | outputInfo, |
| 245 | descriptor); |
| 246 | }; |
| 247 | |
| 248 | if(IsDynamicTensor(outputInfo)) |
| 249 | { |
| 250 | isSupported = AreDynamicTensorsSupported(); |
| 251 | } |
| 252 | else |
| 253 | { |
| 254 | validateFunc(outputInfo, isSupported); |
| 255 | } |
| 256 | |
| 257 | if (!isSupported) |
| 258 | { |
| 259 | return false; |
| 260 | } |
| 261 | |
| 262 | armnn::IConnectableLayer* layer = data.m_Network->AddArgMinMaxLayer(descriptor); |
Cathal Corbett | 5383767 | 2022-09-01 11:34:37 +0100 | [diff] [blame] | 263 | layer->SetBackendId(setBackend); |
Sadik Armagan | 8f397a1 | 2022-06-17 15:38:22 +0100 | [diff] [blame] | 264 | assert(layer != nullptr); |
| 265 | |
| 266 | input0.Connect(layer->GetInputSlot(0)); |
| 267 | |
| 268 | return SetupAndTrackLayerOutputSlot(operation, 0, *layer, model, data, nullptr, validateFunc); |
| 269 | } |
| 270 | |
| 271 | bool Converter::ConvertAveragePool2d(const Operation& operation, const Model& model, ConversionData& data) |
| 272 | { |
| 273 | VLOG(DRIVER) << "Converter::ConvertAveragePool2d()"; |
| 274 | return ConvertPooling2d(operation, __func__, PoolingAlgorithm::Average, model, data); |
| 275 | } |
| 276 | |
Kevin May | 9636a9b | 2022-09-21 15:41:41 +0100 | [diff] [blame] | 277 | bool Converter::ConvertBatchMatMul(const Operation& operation, const Model& model, ConversionData& data) |
| 278 | { |
| 279 | VLOG(DRIVER) << "Converter::ConvertBatchMatMul()"; |
| 280 | LayerInputHandle input0 = ConvertToLayerInputHandle(operation, 0, model, data); |
| 281 | LayerInputHandle input1 = ConvertToLayerInputHandle(operation, 1, model, data); |
| 282 | |
| 283 | if (!input0.IsValid() || !input1.IsValid()) |
| 284 | { |
| 285 | return Fail("%s: Operation has invalid inputs", __func__); |
| 286 | } |
| 287 | |
| 288 | const armnn::TensorInfo& inputInfo0 = input0.GetTensorInfo(); |
| 289 | const armnn::TensorInfo& inputInfo1 = input1.GetTensorInfo(); |
| 290 | |
| 291 | unsigned int rankInput0 = inputInfo0.GetNumDimensions(); |
| 292 | if (rankInput0 > 4 || rankInput0 < 2) |
| 293 | { |
| 294 | Fail("%s: Only inputs with rank at least 2 and up to 4 are supported", __func__); |
| 295 | } |
| 296 | |
| 297 | unsigned int rankInput1 = inputInfo1.GetNumDimensions(); |
| 298 | if (rankInput1 > 4 || rankInput1 < 2) |
| 299 | { |
| 300 | Fail("%s: Only inputs with rank at least 2 and up to 4 are supported", __func__); |
| 301 | } |
| 302 | |
| 303 | // Determine data type of input tensor 0 |
| 304 | OperandType input0Type; |
| 305 | if (!GetOperandType(operation, 0, model, input0Type)) |
| 306 | { |
| 307 | return Fail("%s: Operation has invalid inputs", __func__); |
| 308 | } |
| 309 | |
| 310 | // Determine data type of input tensor 0 |
| 311 | OperandType input1Type; |
| 312 | if (!GetOperandType(operation, 0, model, input1Type)) |
| 313 | { |
| 314 | return Fail("%s: Operation has invalid inputs", __func__); |
| 315 | } |
| 316 | |
| 317 | if (input0Type != input1Type) |
| 318 | { |
| 319 | return Fail("%s: Operation has invalid inputs (Inputs must have same OperandCode)", __func__); |
| 320 | } |
| 321 | |
| 322 | const Operand* output = GetOutputOperand(operation, 0, model); |
| 323 | if (!output) |
| 324 | { |
| 325 | return Fail("%s: Could not read output 0", __func__); |
| 326 | } |
| 327 | |
| 328 | const armnn::TensorInfo& outputInfo = GetTensorInfoForOperand(*output); |
| 329 | |
| 330 | armnn::BatchMatMulDescriptor batchMatMulDesc; |
| 331 | |
| 332 | // Inputs 2 and 3 are adjoint in Android NeuralNetworks, but they perform transpose. |
| 333 | // This is why we are linking them with transpose parameters in the descriptor |
| 334 | batchMatMulDesc.m_TransposeX = GetOptionalBool(operation, 2, model, data); |
| 335 | batchMatMulDesc.m_TransposeY = GetOptionalBool(operation, 3, model, data); |
| 336 | |
| 337 | bool isSupported = false; |
Cathal Corbett | 5383767 | 2022-09-01 11:34:37 +0100 | [diff] [blame] | 338 | armnn::BackendId setBackend; |
Kevin May | 9636a9b | 2022-09-21 15:41:41 +0100 | [diff] [blame] | 339 | auto validateFunc = [&](const armnn::TensorInfo& outputInfo, bool& isSupported) |
| 340 | { |
| 341 | FORWARD_LAYER_SUPPORT_FUNC(__func__, |
| 342 | IsBatchMatMulSupported, |
| 343 | data.m_Backends, |
| 344 | isSupported, |
Cathal Corbett | 5383767 | 2022-09-01 11:34:37 +0100 | [diff] [blame] | 345 | setBackend, |
Kevin May | 9636a9b | 2022-09-21 15:41:41 +0100 | [diff] [blame] | 346 | inputInfo0, |
| 347 | inputInfo1, |
| 348 | outputInfo, |
| 349 | batchMatMulDesc); |
| 350 | }; |
| 351 | |
| 352 | if(!IsDynamicTensor(outputInfo)) |
| 353 | { |
| 354 | validateFunc(outputInfo, isSupported); |
| 355 | } |
| 356 | else |
| 357 | { |
| 358 | isSupported = AreDynamicTensorsSupported(); |
| 359 | } |
| 360 | |
| 361 | |
| 362 | if (!isSupported) |
| 363 | { |
| 364 | return false; |
| 365 | } |
| 366 | |
| 367 | armnn::IConnectableLayer* const layer = data.m_Network->AddBatchMatMulLayer(batchMatMulDesc); |
Cathal Corbett | 5383767 | 2022-09-01 11:34:37 +0100 | [diff] [blame] | 368 | layer->SetBackendId(setBackend); |
Kevin May | 9636a9b | 2022-09-21 15:41:41 +0100 | [diff] [blame] | 369 | assert(layer != nullptr); |
| 370 | input0.Connect(layer->GetInputSlot(0)); |
| 371 | input1.Connect(layer->GetInputSlot(1)); |
| 372 | |
| 373 | return SetupAndTrackLayerOutputSlot(operation, 0, *layer, model, data, nullptr, validateFunc); |
| 374 | } |
| 375 | |
Sadik Armagan | 8f397a1 | 2022-06-17 15:38:22 +0100 | [diff] [blame] | 376 | bool Converter::ConvertBatchToSpaceNd(const Operation& operation, const Model& model, ConversionData& data) |
| 377 | { |
| 378 | VLOG(DRIVER) << "Converter::ConvertBatchToSpaceNd()"; |
| 379 | LayerInputHandle input = ConvertToLayerInputHandle(operation, 0, model, data); |
| 380 | if (!input.IsValid()) |
| 381 | { |
| 382 | return Fail("%s: Operation has invalid inputs", __func__); |
| 383 | } |
| 384 | |
| 385 | const Operand* output = GetOutputOperand(operation, 0, model); |
| 386 | if (!output) |
| 387 | { |
| 388 | return Fail("%s: Could not read output 0", __func__); |
| 389 | } |
| 390 | |
| 391 | const armnn::TensorInfo& outputInfo = GetTensorInfoForOperand(*output); |
| 392 | |
| 393 | const Operand* blockOperand = GetInputOperand(operation, 1, model); |
| 394 | if (!blockOperand) |
| 395 | { |
| 396 | return Fail("%s: Could not read input 1", __func__); |
| 397 | } |
| 398 | |
| 399 | // Convert the block operand to int32 |
| 400 | std::vector<int32_t> block; |
| 401 | if (!GetTensorInt32Values(*blockOperand, block, model, data)) |
| 402 | { |
| 403 | return Fail("%s: Input 1 has invalid values", __func__); |
| 404 | } |
| 405 | |
| 406 | const armnn::TensorInfo& inputInfo = input.GetTensorInfo(); |
| 407 | |
| 408 | unsigned int rank = inputInfo.GetNumDimensions(); |
| 409 | if (rank != 4) |
| 410 | { |
| 411 | Fail("%s: Only inputs with rank equal to 4 are supported", __func__); |
| 412 | } |
| 413 | |
| 414 | if (std::any_of(block.cbegin(), block.cend(), [](int32_t i){ return i < 1; })) |
| 415 | { |
| 416 | return Fail("%s: Block sizes for each spatial dimension of the input tensor must be" |
| 417 | " greater than or equal to 1", __func__); |
| 418 | } |
| 419 | |
| 420 | armnn::BatchToSpaceNdDescriptor batchToSpaceNdDesc; |
| 421 | batchToSpaceNdDesc.m_BlockShape.assign(block.cbegin(), block.cend()); |
| 422 | batchToSpaceNdDesc.m_DataLayout = armnn::DataLayout::NHWC; |
| 423 | |
| 424 | if (Is12OrLaterOperand(*output)) |
| 425 | { |
| 426 | batchToSpaceNdDesc.m_DataLayout = OptionalDataLayout(operation, 2, model, data); |
| 427 | } |
| 428 | // Setting crops to 0,0 0,0 as it is not supported in Android NN API |
| 429 | batchToSpaceNdDesc.m_Crops = {{0, 0}, {0, 0}}; |
| 430 | |
| 431 | bool isSupported = false; |
Cathal Corbett | 5383767 | 2022-09-01 11:34:37 +0100 | [diff] [blame] | 432 | armnn::BackendId setBackend; |
Sadik Armagan | 8f397a1 | 2022-06-17 15:38:22 +0100 | [diff] [blame] | 433 | auto validateFunc = [&](const armnn::TensorInfo& outputInfo, bool& isSupported) |
| 434 | { |
| 435 | FORWARD_LAYER_SUPPORT_FUNC(__func__, |
| 436 | IsBatchToSpaceNdSupported, |
| 437 | data.m_Backends, |
| 438 | isSupported, |
Cathal Corbett | 5383767 | 2022-09-01 11:34:37 +0100 | [diff] [blame] | 439 | setBackend, |
Sadik Armagan | 8f397a1 | 2022-06-17 15:38:22 +0100 | [diff] [blame] | 440 | inputInfo, |
| 441 | outputInfo, |
| 442 | batchToSpaceNdDesc); |
| 443 | }; |
| 444 | |
| 445 | if(!IsDynamicTensor(outputInfo)) |
| 446 | { |
| 447 | validateFunc(outputInfo, isSupported); |
| 448 | } |
| 449 | else |
| 450 | { |
| 451 | isSupported = AreDynamicTensorsSupported(); |
| 452 | } |
| 453 | |
| 454 | |
| 455 | if (!isSupported) |
| 456 | { |
| 457 | return false; |
| 458 | } |
| 459 | |
| 460 | armnn::IConnectableLayer* const layer = data.m_Network->AddBatchToSpaceNdLayer(batchToSpaceNdDesc); |
Cathal Corbett | 5383767 | 2022-09-01 11:34:37 +0100 | [diff] [blame] | 461 | layer->SetBackendId(setBackend); |
Sadik Armagan | 8f397a1 | 2022-06-17 15:38:22 +0100 | [diff] [blame] | 462 | assert(layer != nullptr); |
| 463 | input.Connect(layer->GetInputSlot(0)); |
| 464 | |
| 465 | return SetupAndTrackLayerOutputSlot(operation, 0, *layer, model, data, nullptr, validateFunc); |
| 466 | } |
| 467 | |
| 468 | bool Converter::ConvertCast(const Operation& operation, const Model& model, ConversionData& data) |
| 469 | { |
| 470 | VLOG(DRIVER) << "Converter::ConvertCast()"; |
| 471 | |
| 472 | LayerInputHandle input = ConvertToLayerInputHandle(operation, 0, model, data); |
| 473 | |
| 474 | if (!input.IsValid()) |
| 475 | { |
| 476 | return Fail("%s: Operation has invalid inputs", __func__); |
| 477 | } |
| 478 | |
| 479 | const Operand* output = GetOutputOperand(operation, 0, model); |
| 480 | if (!output) |
| 481 | { |
| 482 | return Fail("%s: Could not read output 0", __func__); |
| 483 | } |
| 484 | |
| 485 | const TensorInfo& inputInfo = input.GetTensorInfo(); |
| 486 | const TensorInfo& outputInfo = GetTensorInfoForOperand(*output); |
| 487 | |
| 488 | bool isSupported = false; |
Cathal Corbett | 5383767 | 2022-09-01 11:34:37 +0100 | [diff] [blame] | 489 | armnn::BackendId setBackend; |
Sadik Armagan | 8f397a1 | 2022-06-17 15:38:22 +0100 | [diff] [blame] | 490 | auto validateFunc = [&](const armnn::TensorInfo& outputInfo, bool& isSupported) |
| 491 | { |
| 492 | FORWARD_LAYER_SUPPORT_FUNC(__func__, |
| 493 | IsCastSupported, |
| 494 | data.m_Backends, |
| 495 | isSupported, |
Cathal Corbett | 5383767 | 2022-09-01 11:34:37 +0100 | [diff] [blame] | 496 | setBackend, |
Sadik Armagan | 8f397a1 | 2022-06-17 15:38:22 +0100 | [diff] [blame] | 497 | inputInfo, |
| 498 | outputInfo); |
| 499 | }; |
| 500 | |
| 501 | if(!IsDynamicTensor(outputInfo)) |
| 502 | { |
| 503 | validateFunc(outputInfo, isSupported); |
| 504 | } |
| 505 | else |
| 506 | { |
| 507 | isSupported = AreDynamicTensorsSupported(); |
| 508 | } |
| 509 | |
| 510 | if (!isSupported) |
| 511 | { |
| 512 | return false; |
| 513 | } |
| 514 | |
| 515 | IConnectableLayer* layer = data.m_Network->AddCastLayer(); |
Cathal Corbett | 5383767 | 2022-09-01 11:34:37 +0100 | [diff] [blame] | 516 | layer->SetBackendId(setBackend); |
Sadik Armagan | 8f397a1 | 2022-06-17 15:38:22 +0100 | [diff] [blame] | 517 | assert(layer != nullptr); |
| 518 | input.Connect(layer->GetInputSlot(0)); |
| 519 | |
| 520 | return SetupAndTrackLayerOutputSlot(operation, 0, *layer, model, data, nullptr, validateFunc); |
| 521 | } |
| 522 | |
| 523 | bool Converter::ConvertComparison(const Operation& operation, |
| 524 | const Model& model, |
| 525 | ConversionData& data, |
| 526 | ComparisonOperation comparisonOperation) |
| 527 | { |
| 528 | VLOG(DRIVER) << "Converter::ConvertComparison()"; |
| 529 | VLOG(DRIVER) << "comparisonOperation = " << GetComparisonOperationAsCString(comparisonOperation); |
| 530 | |
| 531 | LayerInputHandle input0 = ConvertToLayerInputHandle(operation, 0, model, data); |
| 532 | LayerInputHandle input1 = ConvertToLayerInputHandle(operation, 1, model, data); |
| 533 | |
| 534 | if (!(input0.IsValid() && input1.IsValid())) |
| 535 | { |
| 536 | return Fail("%s: Operation has invalid inputs", __func__); |
| 537 | } |
| 538 | |
| 539 | const Operand* output = GetOutputOperand(operation, 0, model); |
| 540 | if (!output) |
| 541 | { |
| 542 | return Fail("%s: Could not read output 0", __func__); |
| 543 | } |
| 544 | |
| 545 | const TensorInfo& inputInfo0 = input0.GetTensorInfo(); |
| 546 | const TensorInfo& inputInfo1 = input1.GetTensorInfo(); |
| 547 | const TensorInfo& outputInfo = GetTensorInfoForOperand(*output); |
| 548 | |
| 549 | ComparisonDescriptor descriptor(comparisonOperation); |
| 550 | |
| 551 | bool isSupported = false; |
Cathal Corbett | 5383767 | 2022-09-01 11:34:37 +0100 | [diff] [blame] | 552 | armnn::BackendId setBackend; |
Sadik Armagan | 8f397a1 | 2022-06-17 15:38:22 +0100 | [diff] [blame] | 553 | auto validateFunc = [&](const armnn::TensorInfo& outputInfo, bool& isSupported) |
| 554 | { |
| 555 | FORWARD_LAYER_SUPPORT_FUNC(__func__, |
| 556 | IsComparisonSupported, |
| 557 | data.m_Backends, |
| 558 | isSupported, |
Cathal Corbett | 5383767 | 2022-09-01 11:34:37 +0100 | [diff] [blame] | 559 | setBackend, |
Sadik Armagan | 8f397a1 | 2022-06-17 15:38:22 +0100 | [diff] [blame] | 560 | inputInfo0, |
| 561 | inputInfo1, |
| 562 | outputInfo, |
| 563 | descriptor); |
| 564 | }; |
| 565 | |
| 566 | if(!IsDynamicTensor(outputInfo)) |
| 567 | { |
| 568 | validateFunc(outputInfo, isSupported); |
| 569 | } |
| 570 | else |
| 571 | { |
| 572 | isSupported = AreDynamicTensorsSupported(); |
| 573 | } |
| 574 | |
| 575 | if (!isSupported) |
| 576 | { |
| 577 | return false; |
| 578 | } |
| 579 | |
| 580 | IConnectableLayer* layer = data.m_Network->AddComparisonLayer(descriptor); |
Cathal Corbett | 5383767 | 2022-09-01 11:34:37 +0100 | [diff] [blame] | 581 | layer->SetBackendId(setBackend); |
Sadik Armagan | 8f397a1 | 2022-06-17 15:38:22 +0100 | [diff] [blame] | 582 | assert(layer != nullptr); |
| 583 | |
| 584 | bool isReshapeSupported = BroadcastTensor(input0, input1, layer, data); |
| 585 | if (!isReshapeSupported) |
| 586 | { |
| 587 | return false; |
| 588 | } |
| 589 | |
| 590 | if(IsDynamicTensor(outputInfo)) |
| 591 | { |
| 592 | input0.Connect(layer->GetInputSlot(0)); |
| 593 | input1.Connect(layer->GetInputSlot(1)); |
| 594 | } |
| 595 | |
| 596 | return SetupAndTrackLayerOutputSlot(operation, 0, *layer, model, data, nullptr, validateFunc); |
| 597 | } |
| 598 | |
| 599 | |
| 600 | bool Converter::ConvertConcatenation(const Operation& operation, const Model& model, ConversionData& data) |
| 601 | { |
| 602 | VLOG(DRIVER) << "Converter::ConvertConcatenation()"; |
| 603 | |
| 604 | // The first N (0..N-1) inputs are tensors. The Nth input is the concatenation axis. |
| 605 | if (operation.inputs.size() <= 1) |
| 606 | { |
| 607 | return Fail("%s: Operation has insufficient arguments", __func__); |
| 608 | } |
| 609 | |
| 610 | // Get inputs and outputs |
| 611 | const std::size_t numInputTensors = operation.inputs.size() - 1; |
| 612 | |
| 613 | int32_t concatDim; |
| 614 | if (!GetInputScalar(operation, numInputTensors, OperandType::INT32, concatDim, model, data)) |
| 615 | { |
| 616 | return Fail("%s: Operation has invalid inputs", __func__); |
| 617 | } |
| 618 | |
| 619 | const Operand* outputOperand = GetOutputOperand(operation, 0, model); |
| 620 | if (!outputOperand) |
| 621 | { |
| 622 | return Fail("%s: Operation has no outputs", __func__); |
| 623 | } |
| 624 | |
| 625 | armnn::TensorInfo outputInfo = GetTensorInfoForOperand(*outputOperand); |
| 626 | armnn::TensorShape outputShape = outputInfo.GetShape(); |
| 627 | const bool isDynamicTensor = IsDynamicTensor(outputInfo); |
| 628 | // |
| 629 | // handle negative concat dims along the lines of tensorflow as described here: |
| 630 | // https://www.tensorflow.org/api_docs/python/tf/concat |
| 631 | // "negative axis refers to axis + rank(values)-th dimension" |
| 632 | // |
| 633 | if (concatDim < 0) |
| 634 | { |
| 635 | concatDim += outputShape.GetNumDimensions(); |
| 636 | } |
| 637 | |
| 638 | if (concatDim >= static_cast<int32_t>(outputShape.GetNumDimensions()) || concatDim < 0) |
| 639 | { |
| 640 | return Fail("%s: Operation has invalid concat axis: %d", __func__, concatDim); |
| 641 | } |
| 642 | |
| 643 | std::vector<LayerInputHandle> inputHandles; |
| 644 | std::vector<armnn::TensorShape> inputShapes; |
| 645 | |
| 646 | inputHandles.reserve(numInputTensors); |
| 647 | inputShapes.reserve(numInputTensors); |
| 648 | |
| 649 | bool inputsHaveBeenReshaped = false; |
| 650 | unsigned int tensorDimensionsAdded = 0; |
| 651 | for (uint32_t i = 0; i < numInputTensors; ++i) |
| 652 | { |
| 653 | const Operand* operand = GetInputOperand(operation, i, model); |
| 654 | if (!operand) |
| 655 | { |
| 656 | return Fail("%s: Operation has invalid inputs", __func__); |
| 657 | } |
| 658 | |
| 659 | LayerInputHandle operandInputHandle = ConvertToLayerInputHandle(operation, i, model, data); |
| 660 | if (!operandInputHandle.IsValid()) |
| 661 | { |
| 662 | return Fail("%s: Operation has invalid inputs", __func__); |
| 663 | } |
| 664 | |
| 665 | armnn::TensorShape operandShape = GetTensorShapeForOperand(*operand); |
| 666 | if (operandShape.GetNumDimensions() == 0) |
| 667 | { |
| 668 | return Fail("%s: Operands with rank 0 are not supported", __func__); |
| 669 | } |
| 670 | |
| 671 | if (RequiresReshape(operandShape)) |
| 672 | { |
| 673 | inputsHaveBeenReshaped = true; |
| 674 | |
| 675 | armnn::TensorInfo reshapeInfo = operandInputHandle.GetTensorInfo(); |
| 676 | |
| 677 | // Expand the tensor to three dimensions |
| 678 | if (operandShape.GetNumDimensions() == 2) |
| 679 | { |
| 680 | reshapeInfo.SetShape(armnn::TensorShape({1, operandShape[0], operandShape[1]})); |
| 681 | tensorDimensionsAdded = 1; |
| 682 | } |
| 683 | else |
| 684 | { |
| 685 | reshapeInfo.SetShape(armnn::TensorShape({1, 1, operandShape[0]})); |
| 686 | tensorDimensionsAdded = 2; |
| 687 | } |
| 688 | |
| 689 | armnn::ReshapeDescriptor reshapeDescriptor; |
| 690 | reshapeDescriptor.m_TargetShape = reshapeInfo.GetShape(); |
| 691 | |
| 692 | bool isSupported = false; |
Cathal Corbett | 5383767 | 2022-09-01 11:34:37 +0100 | [diff] [blame] | 693 | armnn::BackendId setBackendReshape; |
Sadik Armagan | 8f397a1 | 2022-06-17 15:38:22 +0100 | [diff] [blame] | 694 | FORWARD_LAYER_SUPPORT_FUNC(__func__, |
| 695 | IsReshapeSupported, |
| 696 | data.m_Backends, |
| 697 | isSupported, |
Cathal Corbett | 5383767 | 2022-09-01 11:34:37 +0100 | [diff] [blame] | 698 | setBackendReshape, |
Sadik Armagan | 8f397a1 | 2022-06-17 15:38:22 +0100 | [diff] [blame] | 699 | operandInputHandle.GetTensorInfo(), |
| 700 | reshapeInfo, |
| 701 | reshapeDescriptor); |
| 702 | |
| 703 | if (!isSupported) |
| 704 | { |
| 705 | return false; |
| 706 | } |
| 707 | armnn::IConnectableLayer& newReshape = AddReshapeLayer(*data.m_Network, operandInputHandle, reshapeInfo); |
Cathal Corbett | 5383767 | 2022-09-01 11:34:37 +0100 | [diff] [blame] | 708 | newReshape.SetBackendId(setBackendReshape); |
Sadik Armagan | 8f397a1 | 2022-06-17 15:38:22 +0100 | [diff] [blame] | 709 | |
| 710 | // Point to the reshape operation rather then the input operation |
| 711 | operandShape = reshapeInfo.GetShape(); |
| 712 | operandInputHandle = LayerInputHandle(true, &newReshape.GetOutputSlot(0), reshapeInfo); |
| 713 | } |
| 714 | |
| 715 | inputShapes.emplace_back(operandShape); |
| 716 | inputHandles.emplace_back(operandInputHandle); |
| 717 | |
| 718 | if (!inputHandles.back().IsValid()) |
| 719 | { |
| 720 | return Fail("%s: Operation has invalid inputs", __func__); |
| 721 | } |
| 722 | } |
| 723 | |
Kevin May | 28c3d0f | 2023-07-04 09:07:30 +0100 | [diff] [blame] | 724 | if (inputShapes.size() != inputHandles.size()) |
| 725 | { |
| 726 | return Fail("%s: invalid model input shapes size doesn't match input handles size: %i != %i", __func__, |
| 727 | inputShapes.size(), inputHandles.size()); |
| 728 | } |
Sadik Armagan | 8f397a1 | 2022-06-17 15:38:22 +0100 | [diff] [blame] | 729 | |
| 730 | if (inputsHaveBeenReshaped) |
| 731 | { |
| 732 | // Adjust the concatenation dimension by the amount of dimensions added (if any) |
| 733 | concatDim += tensorDimensionsAdded; |
| 734 | |
| 735 | // Add extra dimensions to the output shape to reflect the addition of the reshape layers |
| 736 | if (tensorDimensionsAdded == 1) |
| 737 | { |
| 738 | if (IsDynamicTensor(outputInfo)) |
| 739 | { |
| 740 | outputShape = armnn::TensorShape({1, 0, 0}, {true, false, false}); |
| 741 | } |
| 742 | else |
| 743 | { |
| 744 | outputShape = armnn::TensorShape({1, outputShape[0], outputShape[1]}); |
| 745 | } |
| 746 | } |
| 747 | else if (tensorDimensionsAdded == 2) |
| 748 | { |
| 749 | if (IsDynamicTensor(outputInfo)) |
| 750 | { |
| 751 | outputShape = armnn::TensorShape({1, 1, 0}, {true, true, false}); |
| 752 | } |
| 753 | else |
| 754 | { |
| 755 | outputShape = armnn::TensorShape({1, 1, outputShape[0]}); |
| 756 | } |
| 757 | } |
| 758 | } |
| 759 | |
| 760 | // Check if permutations is required and get the pair of permutations required for the concatenation. |
| 761 | // Permutation is required when the concat dimension is 2 for a 4D tensor or 1 for a 3D tensor. |
| 762 | std::pair<armnn::PermutationVector, armnn::PermutationVector> permutationPair = |
| 763 | std::make_pair(IdentityPermutation4D, IdentityPermutation4D); |
| 764 | bool needPermute = CreateConcatPermutationParameters(inputShapes[0].GetNumDimensions(), |
| 765 | concatDim, |
| 766 | permutationPair); |
| 767 | |
| 768 | // Only relevant to static tensors as dynamic output tensors will be transposed as a result of inferring from input |
| 769 | if (!isDynamicTensor) |
| 770 | { |
| 771 | if (needPermute) |
| 772 | { |
| 773 | outputShape = armnnUtils::TransposeTensorShape(outputShape, permutationPair.first); |
| 774 | } |
| 775 | |
| 776 | outputInfo.SetShape(outputShape); |
| 777 | } |
| 778 | // this is no-op for identity swizzles, otherwise it replaces both |
| 779 | // the handles and shapes with the swizzled layer output handles and shapes |
| 780 | if (!TransposeInputTensors(data, inputHandles, inputShapes, permutationPair.first)) |
| 781 | { |
| 782 | return false; |
| 783 | } |
| 784 | |
| 785 | // Create an armnn concat layer descriptor - this will also perform validation on the input shapes |
| 786 | armnn::OriginsDescriptor concatDescriptor; |
| 787 | |
| 788 | try |
| 789 | { |
| 790 | // The concat descriptor is always created across the only supported concat dimension |
| 791 | // which is 0, 1 or 3 for a 4-D tensor, or 0 or 2 for a 3-D tensor. |
| 792 | concatDescriptor = armnn::CreateDescriptorForConcatenation(inputShapes.begin(), |
| 793 | inputShapes.end(), |
| 794 | concatDim); |
| 795 | } catch (std::exception& error) |
| 796 | { |
| 797 | return Fail("%s: Error preparing concat descriptor. %s", __func__, error.what()); |
| 798 | } |
| 799 | |
| 800 | // Validate the output shape is correct given the input shapes based on the |
| 801 | // only valid concat dimension which is 0, 1 or 3 for a 4-D tensor, or 0 or 2 for a 3-D tensor. |
| 802 | if (!isDynamicTensor) |
| 803 | { |
| 804 | if (!ValidateConcatOutputShape(inputShapes, outputShape, concatDim)) |
| 805 | { |
| 806 | return Fail("%s: Error validating the output shape for concat", __func__); |
| 807 | } |
| 808 | } |
| 809 | |
| 810 | std::vector<const armnn::TensorInfo*> inputTensorInfos; |
| 811 | std::transform(inputHandles.begin(), inputHandles.end(), std::back_inserter(inputTensorInfos), |
| 812 | [](const LayerInputHandle& h)->const armnn::TensorInfo*{ return &h.GetTensorInfo(); }); |
| 813 | |
| 814 | bool isSupported = false; |
Cathal Corbett | 5383767 | 2022-09-01 11:34:37 +0100 | [diff] [blame] | 815 | armnn::BackendId setBackendConcat; |
Sadik Armagan | 8f397a1 | 2022-06-17 15:38:22 +0100 | [diff] [blame] | 816 | auto validateFunc = [&](const armnn::TensorInfo& outputInfo, bool& isSupported){ |
Cathal Corbett | 5383767 | 2022-09-01 11:34:37 +0100 | [diff] [blame] | 817 | FORWARD_LAYER_SUPPORT_FUNC(__func__, |
| 818 | IsConcatSupported, |
| 819 | data.m_Backends, |
| 820 | isSupported, |
| 821 | setBackendConcat, |
| 822 | inputTensorInfos, |
| 823 | outputInfo, |
| 824 | concatDescriptor); |
Sadik Armagan | 8f397a1 | 2022-06-17 15:38:22 +0100 | [diff] [blame] | 825 | }; |
| 826 | |
| 827 | if (!isDynamicTensor) |
| 828 | { |
| 829 | validateFunc(outputInfo, isSupported); |
| 830 | } |
| 831 | else |
| 832 | { |
| 833 | isSupported = AreDynamicTensorsSupported(); |
| 834 | } |
| 835 | |
| 836 | if (!isSupported) |
| 837 | { |
| 838 | return false; |
| 839 | } |
| 840 | |
| 841 | armnn::IConnectableLayer* layer = data.m_Network->AddConcatLayer(concatDescriptor); |
Cathal Corbett | 5383767 | 2022-09-01 11:34:37 +0100 | [diff] [blame] | 842 | layer->SetBackendId(setBackendConcat); |
Sadik Armagan | 8f397a1 | 2022-06-17 15:38:22 +0100 | [diff] [blame] | 843 | assert(layer != nullptr); |
| 844 | layer->GetOutputSlot(0).SetTensorInfo(outputInfo); |
| 845 | // Connect inputs to the layer |
| 846 | const int numInputSlots = layer->GetNumInputSlots(); |
| 847 | assert(static_cast<std::size_t>(numInputSlots) == inputHandles.size()); |
| 848 | for (int i = 0; i < numInputSlots; ++i) |
| 849 | { |
| 850 | // connect the input directly to the merge (concat) layer |
| 851 | inputHandles[static_cast<unsigned int>(i)].Connect(layer->GetInputSlot(i)); |
| 852 | } |
| 853 | |
| 854 | // Transpose the output shape |
| 855 | auto transposeOutputShape = [&](){ |
| 856 | armnn::TransposeDescriptor transposeDesc; |
| 857 | transposeDesc.m_DimMappings = permutationPair.second; |
| 858 | armnn::TensorInfo inputTransposeInfo = layer->GetOutputSlot(0).GetTensorInfo(); |
| 859 | armnn::TensorInfo outputTransposeInfo = armnnUtils::TransposeTensorShape(inputTransposeInfo, |
| 860 | permutationPair.second); |
| 861 | isSupported = false; |
Cathal Corbett | 5383767 | 2022-09-01 11:34:37 +0100 | [diff] [blame] | 862 | armnn::BackendId setBackendTranspose; |
Sadik Armagan | 8f397a1 | 2022-06-17 15:38:22 +0100 | [diff] [blame] | 863 | FORWARD_LAYER_SUPPORT_FUNC(__func__, |
| 864 | IsTransposeSupported, |
| 865 | data.m_Backends, |
| 866 | isSupported, |
Cathal Corbett | 5383767 | 2022-09-01 11:34:37 +0100 | [diff] [blame] | 867 | setBackendTranspose, |
Sadik Armagan | 8f397a1 | 2022-06-17 15:38:22 +0100 | [diff] [blame] | 868 | inputTransposeInfo, |
| 869 | outputTransposeInfo, |
| 870 | transposeDesc); |
| 871 | if (!isSupported) |
| 872 | { |
| 873 | return false; |
| 874 | } |
| 875 | // Add permutation layer and connect the output to it, the permutation becomes the output layer |
| 876 | armnn::IConnectableLayer& deswizzleLayer = AddTransposeLayer(*data.m_Network, layer->GetOutputSlot(0), |
| 877 | permutationPair.second); |
Cathal Corbett | 5383767 | 2022-09-01 11:34:37 +0100 | [diff] [blame] | 878 | deswizzleLayer.SetBackendId(setBackendTranspose); |
Sadik Armagan | 8f397a1 | 2022-06-17 15:38:22 +0100 | [diff] [blame] | 879 | layer = &deswizzleLayer; |
| 880 | |
| 881 | return true; |
| 882 | }; |
| 883 | |
| 884 | if (needPermute && !isDynamicTensor) |
| 885 | { |
| 886 | transposeOutputShape(); |
| 887 | } |
| 888 | |
| 889 | if (inputsHaveBeenReshaped) |
| 890 | { |
| 891 | if (isDynamicTensor) |
| 892 | { |
| 893 | // Infer the output shapes of concat if outputs are type 1 dynamic |
Jim Flynn | 987398f | 2023-09-16 18:53:52 +0100 | [diff] [blame] | 894 | if (!layer->GetOutputSlot(0).IsTensorInfoSet()) |
| 895 | { |
| 896 | throw armnn::Exception( |
| 897 | "tensor info is not set on output slot, cannot process dynamic tensor after input reshape"); |
| 898 | } |
Sadik Armagan | 8f397a1 | 2022-06-17 15:38:22 +0100 | [diff] [blame] | 899 | if (!ValidateConcatOutputShape(inputShapes, |
| 900 | layer->GetOutputSlot(0).GetTensorInfo().GetShape(), |
| 901 | concatDim)) |
| 902 | { |
| 903 | return Fail("%s: Error validating the output shape for concat", __func__); |
| 904 | } |
| 905 | transposeOutputShape(); |
| 906 | } |
| 907 | |
| 908 | armnn::TensorInfo afterConcatInfo = layer->GetOutputSlot(0).GetTensorInfo(); |
| 909 | // Undo the reshape knowing the amount of dimensions added |
| 910 | if (tensorDimensionsAdded == 1) |
| 911 | { |
| 912 | afterConcatInfo.SetShape( |
| 913 | armnn::TensorShape({afterConcatInfo.GetShape()[1], afterConcatInfo.GetShape()[2]})); |
| 914 | } |
| 915 | else if (tensorDimensionsAdded == 2) |
| 916 | { |
| 917 | afterConcatInfo.SetShape(armnn::TensorShape({afterConcatInfo.GetShape()[2]})); |
| 918 | } |
| 919 | |
| 920 | armnn::ReshapeDescriptor reshapeDescriptor; |
| 921 | reshapeDescriptor.m_TargetShape = afterConcatInfo.GetShape(); |
| 922 | armnn::TensorInfo concatInfo = layer->GetOutputSlot(0).GetTensorInfo(); |
| 923 | |
| 924 | isSupported = false; |
Cathal Corbett | 5383767 | 2022-09-01 11:34:37 +0100 | [diff] [blame] | 925 | armnn::BackendId setBackendReshape2; |
Sadik Armagan | 8f397a1 | 2022-06-17 15:38:22 +0100 | [diff] [blame] | 926 | auto validateReshapeFunc = [&](const armnn::TensorInfo& afterConcatInfo, bool& isSupported){ |
| 927 | FORWARD_LAYER_SUPPORT_FUNC(__func__, |
| 928 | IsReshapeSupported, |
| 929 | data.m_Backends, |
| 930 | isSupported, |
Cathal Corbett | 5383767 | 2022-09-01 11:34:37 +0100 | [diff] [blame] | 931 | setBackendReshape2, |
Sadik Armagan | 8f397a1 | 2022-06-17 15:38:22 +0100 | [diff] [blame] | 932 | concatInfo, |
| 933 | afterConcatInfo, |
| 934 | reshapeDescriptor); |
| 935 | }; |
| 936 | |
| 937 | if (!IsDynamicTensor(afterConcatInfo)) |
| 938 | { |
| 939 | validateReshapeFunc(afterConcatInfo, isSupported); |
| 940 | } |
| 941 | else |
| 942 | { |
| 943 | isSupported = AreDynamicTensorsSupported(); |
| 944 | } |
| 945 | |
| 946 | if (!isSupported) |
| 947 | { |
| 948 | return false; |
| 949 | } |
| 950 | layer = &AddReshapeLayer(*data.m_Network, layer->GetOutputSlot(0), afterConcatInfo); |
Cathal Corbett | 5383767 | 2022-09-01 11:34:37 +0100 | [diff] [blame] | 951 | layer->SetBackendId(setBackendReshape2); |
Sadik Armagan | 8f397a1 | 2022-06-17 15:38:22 +0100 | [diff] [blame] | 952 | return SetupAndTrackLayerOutputSlot(operation, |
| 953 | 0, |
| 954 | *layer, |
| 955 | model, |
| 956 | data, |
| 957 | nullptr, |
| 958 | validateReshapeFunc); |
| 959 | } |
| 960 | |
| 961 | return SetupAndTrackLayerOutputSlot(operation, 0, *layer, model, data, nullptr, validateFunc); |
| 962 | } |
| 963 | |
| 964 | bool Converter::ConvertConv2d(const Operation& operation, const Model& model, ConversionData& data) |
| 965 | { |
| 966 | VLOG(DRIVER) << "Converter::ConvertConv2d()"; |
| 967 | |
| 968 | LayerInputHandle input = ConvertToLayerInputHandle(operation, 0, model, data); |
| 969 | if (!input.IsValid()) |
| 970 | { |
| 971 | return Fail("%s: Operation has invalid inputs", __func__); |
| 972 | } |
| 973 | |
| 974 | const Operand* output = GetOutputOperand(operation, 0, model); |
| 975 | if (!output) |
| 976 | { |
| 977 | return Fail("%s: Could not read output 0", __func__); |
| 978 | } |
| 979 | |
| 980 | const TensorInfo& inputInfo = input.GetTensorInfo(); |
| 981 | const TensorInfo& outputInfo = GetTensorInfoForOperand(*output); |
| 982 | |
| 983 | Convolution2dDescriptor desc; |
| 984 | desc.m_DataLayout = DataLayout::NHWC; |
| 985 | |
| 986 | // Determine whether padding is implicit or explicit |
| 987 | bool implicitPadding = operation.inputs.size() == 7 |
| 988 | || (operation.inputs.size() >= 8 |
| 989 | && GetInputOperand(operation, 7, model)->type == OperandType::BOOL); |
| 990 | |
| 991 | if (implicitPadding) |
| 992 | { |
| 993 | desc.m_DataLayout = OptionalDataLayout(operation, 7, model, data); |
| 994 | } |
| 995 | else if (operation.inputs.size() >= 10) |
| 996 | { |
| 997 | desc.m_DataLayout = OptionalDataLayout(operation, 10, model, data); |
| 998 | } |
| 999 | |
| 1000 | const PermutationVector OHWIToOIHW = {0, 2, 3, 1}; |
| 1001 | |
| 1002 | // ArmNN does not currently support non-fixed weights or bias |
| 1003 | // The NNAPI filter is always OHWI [depth_out, filter_height, filter_width, depth_in] but ArmNN expects the |
| 1004 | // filter's height and width indices to match the input's height and width indices so we permute it to OIHW if |
| 1005 | // the DataLayout is NCHW |
| 1006 | |
Kevin May | 7fbf810 | 2023-08-23 10:07:26 +0100 | [diff] [blame] | 1007 | if (!IsWeightsValid(operation, 1, model, false) && desc.m_DataLayout == DataLayout::NCHW) |
Sadik Armagan | 8f397a1 | 2022-06-17 15:38:22 +0100 | [diff] [blame] | 1008 | { |
| 1009 | return Fail("%s: Operation has unsupported weights OperandLifeTime", __func__); |
| 1010 | } |
| 1011 | |
| 1012 | LayerInputHandle weightsInput = (desc.m_DataLayout == DataLayout::NCHW) |
Sadik Armagan | 1e276f3 | 2022-07-19 12:37:20 +0100 | [diff] [blame] | 1013 | ? ConvertToLayerInputHandle(operation, 1, model, data, OHWIToOIHW, &input) |
| 1014 | : ConvertToLayerInputHandle(operation, 1, model, data, g_DontPermute, &input); |
Sadik Armagan | 8f397a1 | 2022-06-17 15:38:22 +0100 | [diff] [blame] | 1015 | |
| 1016 | if (!weightsInput.IsValid()) |
| 1017 | { |
| 1018 | return Fail("%s: Operation has invalid inputs", __func__); |
| 1019 | } |
| 1020 | |
Sadik Armagan | 1e276f3 | 2022-07-19 12:37:20 +0100 | [diff] [blame] | 1021 | LayerInputHandle biasInput = ConvertToLayerInputHandle(operation, 2, model, data, g_DontPermute, &input); // 1D |
Sadik Armagan | 8f397a1 | 2022-06-17 15:38:22 +0100 | [diff] [blame] | 1022 | if (!biasInput.IsValid()) |
| 1023 | { |
| 1024 | return Fail("%s: Operation has invalid inputs", __func__); |
| 1025 | } |
| 1026 | |
| 1027 | biasInput.SanitizeQuantizationScale(weightsInput, input); |
| 1028 | armnn::TensorInfo weightsInfo = weightsInput.GetTensorInfo(); |
| 1029 | armnn::TensorInfo biasInfo = biasInput.GetTensorInfo(); |
| 1030 | |
| 1031 | ActivationFn activation; |
| 1032 | if (implicitPadding) |
| 1033 | { |
| 1034 | ::android::nn::PaddingScheme paddingScheme; |
| 1035 | if (!GetInputPaddingScheme(operation, 3, paddingScheme, model, data) |
| 1036 | || !GetInputScalar(operation, 4, OperandType::INT32, desc.m_StrideX, model, data) |
| 1037 | || !GetInputScalar(operation, 5, OperandType::INT32, desc.m_StrideY, model, data) |
| 1038 | || !GetInputActivationFunction(operation, 6, activation, model, data) |
| 1039 | || !GetOptionalConvolutionDilationParams(operation, 8, desc, model, data)) |
| 1040 | { |
| 1041 | return Fail("%s: Operation has invalid inputs (implicit padding)", __func__); |
| 1042 | } |
| 1043 | |
| 1044 | armnnUtils::DataLayoutIndexed dataLayoutIndexed(desc.m_DataLayout); |
| 1045 | unsigned int widthIndex = dataLayoutIndexed.GetWidthIndex(); |
| 1046 | unsigned int heightIndex = dataLayoutIndexed.GetHeightIndex(); |
| 1047 | const uint32_t kernelX = weightsInfo.GetShape()[widthIndex]; |
| 1048 | const uint32_t kernelY = weightsInfo.GetShape()[heightIndex]; |
| 1049 | const uint32_t inputX = inputInfo.GetShape()[widthIndex]; |
| 1050 | const uint32_t inputY = inputInfo.GetShape()[heightIndex]; |
| 1051 | |
| 1052 | CalcPadding(inputX, kernelX, desc.m_StrideX, desc.m_DilationX, desc.m_PadLeft, desc.m_PadRight, paddingScheme); |
| 1053 | CalcPadding(inputY, kernelY, desc.m_StrideY, desc.m_DilationY, desc.m_PadTop, desc.m_PadBottom, paddingScheme); |
| 1054 | |
| 1055 | } |
| 1056 | else if (operation.inputs.size() >= 10) |
| 1057 | { |
| 1058 | // explicit padding |
| 1059 | if (!GetInputScalar(operation, 3, OperandType::INT32, desc.m_PadLeft, model, data) |
| 1060 | || !GetInputScalar(operation, 4, OperandType::INT32, desc.m_PadRight, model, data) |
| 1061 | || !GetInputScalar(operation, 5, OperandType::INT32, desc.m_PadTop, model, data) |
| 1062 | || !GetInputScalar(operation, 6, OperandType::INT32, desc.m_PadBottom, model, data) |
| 1063 | || !GetInputScalar(operation, 7, OperandType::INT32, desc.m_StrideX, model, data) |
| 1064 | || !GetInputScalar(operation, 8, OperandType::INT32, desc.m_StrideY, model, data) |
| 1065 | || !GetInputActivationFunction(operation, 9, activation, model, data) |
| 1066 | || !GetOptionalConvolutionDilationParams(operation, 11, desc, model, data)) |
| 1067 | { |
| 1068 | return Fail("%s: Operation has invalid inputs (explicit padding)", __func__); |
| 1069 | } |
| 1070 | } |
| 1071 | else |
| 1072 | { |
| 1073 | return Fail("%s: Unsupported number of operation inputs", __func__); |
| 1074 | } |
| 1075 | |
| 1076 | desc.m_BiasEnabled = true; |
| 1077 | Optional<TensorInfo> biases(biasInfo); |
| 1078 | |
Sadik Armagan | b016157 | 2022-08-03 11:27:05 +0100 | [diff] [blame] | 1079 | bool requiresValidation = true; |
| 1080 | const Operand* weightsOperand = GetInputOperand(operation, 1, model); |
| 1081 | const Operand* biasOperand = GetInputOperand(operation, 2, model); |
| 1082 | if (IsConnectedToDequantize(weightsInput.GetOutputSlot()) |
| 1083 | || IsConnectedToDequantize(biasInput.GetOutputSlot())) |
Sadik Armagan | 8f397a1 | 2022-06-17 15:38:22 +0100 | [diff] [blame] | 1084 | { |
Sadik Armagan | b016157 | 2022-08-03 11:27:05 +0100 | [diff] [blame] | 1085 | // Do not require validation for now. There will be an optimization step |
| 1086 | // [ConvertConstDequantisationLayersToConstLayers] will convert layers to Constant layers |
| 1087 | // then at the end of the optimization there will be layer supported validation. |
| 1088 | requiresValidation = false; |
| 1089 | VLOG(DRIVER) << "Converter::ConvertConv2d(): Weights and Biases are as INPUTS."; |
| 1090 | } |
| 1091 | |
Cathal Corbett | 5383767 | 2022-09-01 11:34:37 +0100 | [diff] [blame] | 1092 | armnn::BackendId setBackend; |
Sadik Armagan | b016157 | 2022-08-03 11:27:05 +0100 | [diff] [blame] | 1093 | auto validateFunc = [&](const armnn::TensorInfo& outputInfo, bool& isSupported) { |
Sadik Armagan | 8f397a1 | 2022-06-17 15:38:22 +0100 | [diff] [blame] | 1094 | FORWARD_LAYER_SUPPORT_FUNC(__func__, |
| 1095 | IsConvolution2dSupported, |
| 1096 | data.m_Backends, |
| 1097 | isSupported, |
Cathal Corbett | 5383767 | 2022-09-01 11:34:37 +0100 | [diff] [blame] | 1098 | setBackend, |
Sadik Armagan | 8f397a1 | 2022-06-17 15:38:22 +0100 | [diff] [blame] | 1099 | inputInfo, |
| 1100 | outputInfo, |
| 1101 | desc, |
| 1102 | weightsInfo, |
| 1103 | biases); |
| 1104 | }; |
| 1105 | |
Sadik Armagan | b016157 | 2022-08-03 11:27:05 +0100 | [diff] [blame] | 1106 | if (requiresValidation) |
Sadik Armagan | 8f397a1 | 2022-06-17 15:38:22 +0100 | [diff] [blame] | 1107 | { |
Sadik Armagan | b016157 | 2022-08-03 11:27:05 +0100 | [diff] [blame] | 1108 | VLOG(DRIVER) << "Converter::ConvertConv2d(): Requires Validation!"; |
| 1109 | bool isSupported = false; |
| 1110 | if (!IsDynamicTensor(outputInfo)) |
| 1111 | { |
| 1112 | validateFunc(outputInfo, isSupported); |
| 1113 | } |
| 1114 | else |
| 1115 | { |
| 1116 | isSupported = AreDynamicTensorsSupported(); |
| 1117 | } |
Sadik Armagan | 8f397a1 | 2022-06-17 15:38:22 +0100 | [diff] [blame] | 1118 | |
Sadik Armagan | b016157 | 2022-08-03 11:27:05 +0100 | [diff] [blame] | 1119 | if (!isSupported) |
| 1120 | { |
| 1121 | return false; |
| 1122 | } |
Sadik Armagan | 8f397a1 | 2022-06-17 15:38:22 +0100 | [diff] [blame] | 1123 | } |
| 1124 | |
| 1125 | armnn::IConnectableLayer* startLayer = data.m_Network->AddConvolution2dLayer(desc); |
Cathal Corbett | 5383767 | 2022-09-01 11:34:37 +0100 | [diff] [blame] | 1126 | startLayer->SetBackendId(setBackend); |
Sadik Armagan | 8f397a1 | 2022-06-17 15:38:22 +0100 | [diff] [blame] | 1127 | |
| 1128 | if (!startLayer) |
| 1129 | { |
| 1130 | return Fail("%s: AddConvolution2dLayer failed", __func__); |
| 1131 | } |
| 1132 | |
| 1133 | input.Connect(startLayer->GetInputSlot(0)); |
| 1134 | weightsInput.Connect(startLayer->GetInputSlot(1)); |
| 1135 | biasInput.Connect(startLayer->GetInputSlot(2)); |
| 1136 | |
| 1137 | return SetupAndTrackLayerOutputSlot(operation, 0, *startLayer, model, data, nullptr, validateFunc, activation); |
| 1138 | } |
| 1139 | |
| 1140 | bool Converter::ConvertDepthToSpace(const Operation& operation, const Model& model, ConversionData& data) |
| 1141 | { |
| 1142 | VLOG(DRIVER) << "Converter::ConvertDepthToSpace()"; |
| 1143 | |
| 1144 | LayerInputHandle input = ConvertToLayerInputHandle(operation, 0, model, data); |
| 1145 | if (!input.IsValid() ) |
| 1146 | { |
| 1147 | return Fail("%s: Operation has invalid inputs", __func__); |
| 1148 | } |
| 1149 | |
| 1150 | const armnn::TensorInfo& inputInfo = input.GetTensorInfo(); |
| 1151 | unsigned int rank = inputInfo.GetNumDimensions(); |
| 1152 | if (rank != 4) |
| 1153 | { |
| 1154 | return Fail("%s: Only inputs with rank 4 are supported", __func__); |
| 1155 | } |
| 1156 | |
| 1157 | const Operand* output = GetOutputOperand(operation, 0, model); |
| 1158 | if (!output) |
| 1159 | { |
| 1160 | return Fail("%s: Could not read output 0", __func__); |
| 1161 | } |
| 1162 | |
| 1163 | const armnn::TensorInfo& outputInfo = GetTensorInfoForOperand(*output); |
| 1164 | |
| 1165 | armnn::DepthToSpaceDescriptor descriptor; |
| 1166 | |
| 1167 | GetInputScalar(operation, 1, OperandType::INT32, descriptor.m_BlockSize, model, data); |
| 1168 | if (descriptor.m_BlockSize <= 1) |
| 1169 | { |
Kevin May | 21e3999 | 2023-07-31 12:20:03 +0100 | [diff] [blame] | 1170 | return Fail("%s: Block size must be at least 1 in all dimensions", __func__); |
Sadik Armagan | 8f397a1 | 2022-06-17 15:38:22 +0100 | [diff] [blame] | 1171 | } |
| 1172 | |
| 1173 | descriptor.m_DataLayout = armnn::DataLayout::NHWC; |
| 1174 | if (Is12OrLaterOperand(*output)) |
| 1175 | { |
| 1176 | descriptor.m_DataLayout = OptionalDataLayout(operation, 2, model, data); |
| 1177 | } |
| 1178 | |
| 1179 | bool isSupported = false; |
Cathal Corbett | 5383767 | 2022-09-01 11:34:37 +0100 | [diff] [blame] | 1180 | armnn::BackendId setBackend; |
Sadik Armagan | 8f397a1 | 2022-06-17 15:38:22 +0100 | [diff] [blame] | 1181 | auto validateFunc = [&](const armnn::TensorInfo& outputInfo, bool& isSupported) |
| 1182 | { |
| 1183 | FORWARD_LAYER_SUPPORT_FUNC(__func__, |
| 1184 | IsDepthToSpaceSupported, |
| 1185 | data.m_Backends, |
| 1186 | isSupported, |
Cathal Corbett | 5383767 | 2022-09-01 11:34:37 +0100 | [diff] [blame] | 1187 | setBackend, |
Sadik Armagan | 8f397a1 | 2022-06-17 15:38:22 +0100 | [diff] [blame] | 1188 | inputInfo, |
| 1189 | outputInfo, |
| 1190 | descriptor); |
| 1191 | }; |
| 1192 | |
| 1193 | if(!IsDynamicTensor(outputInfo)) |
| 1194 | { |
| 1195 | validateFunc(outputInfo, isSupported); |
| 1196 | } |
| 1197 | else |
| 1198 | { |
| 1199 | isSupported = AreDynamicTensorsSupported(); |
| 1200 | } |
| 1201 | |
| 1202 | if (!isSupported) |
| 1203 | { |
| 1204 | return false; |
| 1205 | } |
| 1206 | |
| 1207 | armnn::IConnectableLayer* const layer = data.m_Network->AddDepthToSpaceLayer(descriptor); |
Cathal Corbett | 5383767 | 2022-09-01 11:34:37 +0100 | [diff] [blame] | 1208 | layer->SetBackendId(setBackend); |
Sadik Armagan | 8f397a1 | 2022-06-17 15:38:22 +0100 | [diff] [blame] | 1209 | assert(layer != nullptr); |
| 1210 | input.Connect(layer->GetInputSlot(0)); |
| 1211 | |
| 1212 | return SetupAndTrackLayerOutputSlot(operation, 0, *layer, model, data, nullptr, validateFunc); |
| 1213 | } |
| 1214 | |
| 1215 | bool Converter::ConvertDepthwiseConv2d(const Operation& operation, const Model& model, ConversionData& data) |
| 1216 | { |
| 1217 | VLOG(DRIVER) << "Converter::ConvertDepthwiseConv2d()"; |
| 1218 | |
| 1219 | LayerInputHandle input = ConvertToLayerInputHandle(operation, 0, model, data); |
| 1220 | |
| 1221 | if (!input.IsValid()) |
| 1222 | { |
| 1223 | return Fail("%s: Operation has invalid inputs", __func__); |
| 1224 | } |
| 1225 | |
| 1226 | const Operand* output = GetOutputOperand(operation, 0, model); |
| 1227 | |
| 1228 | if (!output) |
| 1229 | { |
| 1230 | return Fail("%s: Could not read output 0", __func__); |
| 1231 | } |
| 1232 | |
| 1233 | const armnn::TensorInfo& inputInfo = input.GetTensorInfo(); |
| 1234 | const armnn::TensorInfo& outputInfo = GetTensorInfoForOperand(*output); |
| 1235 | |
| 1236 | // ArmNN does not currently support non-fixed weights or bias |
Kevin May | 7fbf810 | 2023-08-23 10:07:26 +0100 | [diff] [blame] | 1237 | if (!IsWeightsValid(operation, 1, model, false)) |
| 1238 | { |
| 1239 | return Fail("%s: This Operation has unsupported weights OperandLifeTime", __func__); |
| 1240 | } |
| 1241 | |
Sadik Armagan | 8f397a1 | 2022-06-17 15:38:22 +0100 | [diff] [blame] | 1242 | // Find the shape of the weights tensor. In AndroidNN this will be [ 1, H, W, I * M ] |
| 1243 | const Operand* weightsOperand = GetInputOperand(operation, 1, model); |
| 1244 | |
Sadik Armagan | 8f397a1 | 2022-06-17 15:38:22 +0100 | [diff] [blame] | 1245 | // Basic sanity check on the weights shape. |
| 1246 | // ANEURALNETWORKS_DEPTHWISE_CONV_2D specifies a 4-D tensor, of shape |
| 1247 | // [1, filter_height, filter_width, depth_out] |
| 1248 | if (weightsOperand->dimensions[0] != 1) |
| 1249 | { |
| 1250 | return Fail("%s: Filter operand dimension 0 is invalid, should be 1", __func__); |
| 1251 | } |
| 1252 | |
| 1253 | armnn::DepthwiseConvolution2dDescriptor desc; |
| 1254 | desc.m_DataLayout = armnn::DataLayout::NHWC; |
| 1255 | |
| 1256 | // Determine whether padding is implicit or explicit |
| 1257 | bool implicitPadding = operation.inputs.size() == 8 |
| 1258 | || (operation.inputs.size() >= 9 |
| 1259 | && GetInputOperand(operation, 8, model)->type == OperandType::BOOL); |
| 1260 | |
| 1261 | // Look ahead to find the optional DataLayout, if present |
| 1262 | const uint32_t dataLayoutFlagIndex = implicitPadding ? 8 : 11; |
| 1263 | desc.m_DataLayout = OptionalDataLayout(operation, dataLayoutFlagIndex, model, data); |
| 1264 | |
| 1265 | armnnUtils::DataLayoutIndexed dataLayoutIndexed(desc.m_DataLayout); |
| 1266 | unsigned int widthIndex = dataLayoutIndexed.GetWidthIndex(); |
| 1267 | unsigned int heightIndex = dataLayoutIndexed.GetHeightIndex(); |
| 1268 | |
Sadik Armagan | 1e276f3 | 2022-07-19 12:37:20 +0100 | [diff] [blame] | 1269 | LayerInputHandle weightsInput = ConvertToLayerInputHandle(operation, 1, model, data, g_DontPermute, &input); |
Sadik Armagan | 8f397a1 | 2022-06-17 15:38:22 +0100 | [diff] [blame] | 1270 | if (!weightsInput.IsValid()) |
| 1271 | { |
| 1272 | return Fail("%s: Operation has invalid inputs", __func__); |
| 1273 | } |
| 1274 | |
| 1275 | const Operand* biasOperand = GetInputOperand(operation, 2, model); |
| 1276 | if (!biasOperand) |
| 1277 | { |
| 1278 | return Fail("%s: Could not read bias", __func__); |
| 1279 | } |
| 1280 | |
Sadik Armagan | 1e276f3 | 2022-07-19 12:37:20 +0100 | [diff] [blame] | 1281 | LayerInputHandle biasInput = ConvertToLayerInputHandle(operation, 2, model, data, g_DontPermute, &input); // 1D |
Sadik Armagan | 8f397a1 | 2022-06-17 15:38:22 +0100 | [diff] [blame] | 1282 | if (!biasInput.IsValid()) |
| 1283 | { |
| 1284 | return Fail("%s: Operation has invalid inputs", __func__); |
| 1285 | } |
| 1286 | |
| 1287 | biasInput.SanitizeQuantizationScale(weightsInput, input); |
| 1288 | armnn::TensorInfo weightsInfo = weightsInput.GetTensorInfo(); |
| 1289 | armnn::TensorInfo biasInfo = biasInput.GetTensorInfo(); |
| 1290 | |
| 1291 | ActivationFn activation; |
| 1292 | if (implicitPadding) |
| 1293 | { |
| 1294 | ::android::nn::PaddingScheme paddingScheme; |
| 1295 | if (!GetInputPaddingScheme(operation, 3, paddingScheme, model, data) |
| 1296 | || !GetInputScalar(operation, 4, OperandType::INT32, desc.m_StrideX, model, data) |
| 1297 | || !GetInputScalar(operation, 5, OperandType::INT32, desc.m_StrideY, model, data) |
| 1298 | || !GetInputActivationFunction(operation, 7, activation, model, data) |
| 1299 | || !GetOptionalConvolutionDilationParams(operation, 9, desc, model, data)) |
| 1300 | { |
| 1301 | return Fail("%s: Operation has invalid inputs (implicit padding)", __func__); |
| 1302 | } |
| 1303 | |
| 1304 | const uint32_t kernelX = weightsInfo.GetShape()[2]; |
| 1305 | const uint32_t kernelY = weightsInfo.GetShape()[1]; |
| 1306 | const uint32_t inputX = inputInfo.GetShape()[widthIndex]; |
| 1307 | const uint32_t inputY = inputInfo.GetShape()[heightIndex]; |
| 1308 | |
| 1309 | CalcPadding(inputX, kernelX, desc.m_StrideX, desc.m_DilationX, desc.m_PadLeft, desc.m_PadRight, paddingScheme); |
| 1310 | CalcPadding(inputY, kernelY, desc.m_StrideY, desc.m_DilationY, desc.m_PadTop, desc.m_PadBottom, paddingScheme); |
| 1311 | } |
| 1312 | else if (operation.inputs.size() >= 11) |
| 1313 | { |
| 1314 | // explicit padding |
| 1315 | if (!GetInputScalar(operation, 3, OperandType::INT32, desc.m_PadLeft, model, data) |
| 1316 | || !GetInputScalar(operation, 4, OperandType::INT32, desc.m_PadRight, model, data) |
| 1317 | || !GetInputScalar(operation, 5, OperandType::INT32, desc.m_PadTop, model, data) |
| 1318 | || !GetInputScalar(operation, 6, OperandType::INT32, desc.m_PadBottom, model, data) |
| 1319 | || !GetInputScalar(operation, 7, OperandType::INT32, desc.m_StrideX, model, data) |
| 1320 | || !GetInputScalar(operation, 8, OperandType::INT32, desc.m_StrideY, model, data) |
| 1321 | || !GetInputActivationFunction(operation, 10, activation, model, data) |
| 1322 | || !GetOptionalConvolutionDilationParams(operation, 12, desc, model, data)) |
| 1323 | { |
| 1324 | return Fail("%s: Operation has invalid inputs (explicit padding)", __func__); |
| 1325 | } |
| 1326 | } |
| 1327 | else |
| 1328 | { |
| 1329 | return Fail("%s: Unsupported number of operation inputs", __func__); |
| 1330 | } |
| 1331 | |
| 1332 | desc.m_BiasEnabled = true; |
| 1333 | Optional<TensorInfo> biases(biasInfo); |
| 1334 | |
Sadik Armagan | b016157 | 2022-08-03 11:27:05 +0100 | [diff] [blame] | 1335 | bool requiresValidation = true; |
| 1336 | if (IsConnectedToDequantize(weightsInput.GetOutputSlot()) || IsConnectedToDequantize(biasInput.GetOutputSlot())) |
Sadik Armagan | 8f397a1 | 2022-06-17 15:38:22 +0100 | [diff] [blame] | 1337 | { |
Sadik Armagan | b016157 | 2022-08-03 11:27:05 +0100 | [diff] [blame] | 1338 | // Do not require validation for now. There will be an optimization step |
| 1339 | // [ConvertConstDequantisationLayersToConstLayers] will convert layers to Constant layers |
| 1340 | // then at the end of the optimization there will be layer supported validation. |
| 1341 | requiresValidation = false; |
| 1342 | VLOG(DRIVER) << "Converter::ConvertDepthwiseConv2d(): Weights and Biases are as INPUTS."; |
| 1343 | } |
| 1344 | |
Cathal Corbett | 5383767 | 2022-09-01 11:34:37 +0100 | [diff] [blame] | 1345 | armnn::BackendId setBackend; |
Sadik Armagan | b016157 | 2022-08-03 11:27:05 +0100 | [diff] [blame] | 1346 | auto validateFunc = [&](const armnn::TensorInfo& outputInfo, bool& isSupported) { |
Sadik Armagan | 8f397a1 | 2022-06-17 15:38:22 +0100 | [diff] [blame] | 1347 | FORWARD_LAYER_SUPPORT_FUNC(__func__, |
| 1348 | IsDepthwiseConvolutionSupported, |
| 1349 | data.m_Backends, |
| 1350 | isSupported, |
Cathal Corbett | 5383767 | 2022-09-01 11:34:37 +0100 | [diff] [blame] | 1351 | setBackend, |
Sadik Armagan | 8f397a1 | 2022-06-17 15:38:22 +0100 | [diff] [blame] | 1352 | inputInfo, |
| 1353 | outputInfo, |
| 1354 | desc, |
| 1355 | weightsInfo, |
| 1356 | biases); |
| 1357 | }; |
| 1358 | |
Sadik Armagan | b016157 | 2022-08-03 11:27:05 +0100 | [diff] [blame] | 1359 | if (requiresValidation) |
Sadik Armagan | 8f397a1 | 2022-06-17 15:38:22 +0100 | [diff] [blame] | 1360 | { |
Sadik Armagan | b016157 | 2022-08-03 11:27:05 +0100 | [diff] [blame] | 1361 | VLOG(DRIVER) << "Converter::ConvertDepthwiseConv2d(): Requires Validation!"; |
| 1362 | bool isSupported = false; |
| 1363 | if (!IsDynamicTensor(outputInfo)) |
| 1364 | { |
| 1365 | validateFunc(outputInfo, isSupported); |
| 1366 | } |
| 1367 | else |
| 1368 | { |
| 1369 | isSupported = AreDynamicTensorsSupported(); |
| 1370 | } |
Sadik Armagan | 8f397a1 | 2022-06-17 15:38:22 +0100 | [diff] [blame] | 1371 | |
Sadik Armagan | b016157 | 2022-08-03 11:27:05 +0100 | [diff] [blame] | 1372 | if (!isSupported) |
| 1373 | { |
| 1374 | return false; |
| 1375 | } |
Sadik Armagan | 8f397a1 | 2022-06-17 15:38:22 +0100 | [diff] [blame] | 1376 | } |
| 1377 | |
| 1378 | armnn::IConnectableLayer* startLayer = data.m_Network->AddDepthwiseConvolution2dLayer(desc); |
Cathal Corbett | 5383767 | 2022-09-01 11:34:37 +0100 | [diff] [blame] | 1379 | startLayer->SetBackendId(setBackend); |
Sadik Armagan | 8f397a1 | 2022-06-17 15:38:22 +0100 | [diff] [blame] | 1380 | |
| 1381 | if (!startLayer) |
| 1382 | { |
| 1383 | return Fail("%s: AddDepthwiseConvolution2dLayer failed", __func__); |
| 1384 | } |
| 1385 | |
| 1386 | input.Connect(startLayer->GetInputSlot(0)); |
| 1387 | |
| 1388 | // Connect weights and bias inputs |
| 1389 | weightsInput.Connect(startLayer->GetInputSlot(1)); |
| 1390 | biasInput.Connect(startLayer->GetInputSlot(2)); |
| 1391 | |
| 1392 | return SetupAndTrackLayerOutputSlot(operation, 0, *startLayer, model, data, nullptr, validateFunc, activation); |
| 1393 | } |
| 1394 | |
| 1395 | bool Converter::ConvertDequantize(const Operation& operation, const Model& model, ConversionData& data) |
| 1396 | { |
| 1397 | VLOG(DRIVER) << "Converter::ConvertDequantize()"; |
| 1398 | |
| 1399 | LayerInputHandle input = ConvertToLayerInputHandle(operation, 0, model, data); |
| 1400 | if (!input.IsValid()) |
| 1401 | { |
| 1402 | return Fail("%s: Operation has invalid input", __func__); |
| 1403 | } |
| 1404 | |
| 1405 | const armnn::TensorInfo& inputInfo = input.GetTensorInfo(); |
| 1406 | const armnn::Optional<unsigned int>& quantizationDim = inputInfo.GetQuantizationDim(); |
| 1407 | if (quantizationDim.has_value() && quantizationDim.value() != 0) |
| 1408 | { |
| 1409 | return Fail("%s: Operation has quantization dimension different than 0", __func__); |
| 1410 | } |
| 1411 | |
| 1412 | const Operand* const outputOperand = GetOutputOperand(operation, 0, model); |
| 1413 | if (!outputOperand) |
| 1414 | { |
| 1415 | return Fail("%s: Operation has invalid outputs", __func__); |
| 1416 | } |
| 1417 | |
| 1418 | const armnn::TensorInfo& outputInfo = GetTensorInfoForOperand(*outputOperand); |
| 1419 | |
| 1420 | bool isSupported = false; |
Cathal Corbett | 5383767 | 2022-09-01 11:34:37 +0100 | [diff] [blame] | 1421 | armnn::BackendId setBackend; |
Sadik Armagan | 8f397a1 | 2022-06-17 15:38:22 +0100 | [diff] [blame] | 1422 | auto validateFunc = [&](const armnn::TensorInfo& outputInfo, bool& isSupported) |
| 1423 | { |
| 1424 | FORWARD_LAYER_SUPPORT_FUNC(__func__, |
| 1425 | IsDequantizeSupported, |
| 1426 | data.m_Backends, |
| 1427 | isSupported, |
Cathal Corbett | 5383767 | 2022-09-01 11:34:37 +0100 | [diff] [blame] | 1428 | setBackend, |
Sadik Armagan | 8f397a1 | 2022-06-17 15:38:22 +0100 | [diff] [blame] | 1429 | inputInfo, |
| 1430 | outputInfo); |
| 1431 | }; |
| 1432 | |
| 1433 | if(IsDynamicTensor(outputInfo)) |
| 1434 | { |
| 1435 | isSupported = AreDynamicTensorsSupported(); |
| 1436 | } |
| 1437 | else |
| 1438 | { |
| 1439 | validateFunc(outputInfo, isSupported); |
| 1440 | } |
| 1441 | |
| 1442 | if (!isSupported) |
| 1443 | { |
| 1444 | return false; |
| 1445 | } |
| 1446 | |
| 1447 | armnn::IConnectableLayer* const layer = data.m_Network->AddDequantizeLayer(); |
Cathal Corbett | 5383767 | 2022-09-01 11:34:37 +0100 | [diff] [blame] | 1448 | layer->SetBackendId(setBackend); |
Sadik Armagan | 8f397a1 | 2022-06-17 15:38:22 +0100 | [diff] [blame] | 1449 | assert(layer != nullptr); |
| 1450 | input.Connect(layer->GetInputSlot(0)); |
| 1451 | |
| 1452 | return SetupAndTrackLayerOutputSlot(operation, 0, *layer, model, data, nullptr, validateFunc); |
| 1453 | } |
| 1454 | |
Kevin May | 28c3d0f | 2023-07-04 09:07:30 +0100 | [diff] [blame] | 1455 | bool Converter::ConvertElementwiseBinary(const Operation& operation, |
| 1456 | const Model& model, |
| 1457 | ConversionData& data, |
| 1458 | armnn::BinaryOperation binaryOperation) |
Sadik Armagan | 8f397a1 | 2022-06-17 15:38:22 +0100 | [diff] [blame] | 1459 | { |
Kevin May | 28c3d0f | 2023-07-04 09:07:30 +0100 | [diff] [blame] | 1460 | VLOG(DRIVER) << "Converter::ConvertElementwiseBinary()"; |
| 1461 | VLOG(DRIVER) << "binaryOperation = " << GetBinaryOperationAsCString(binaryOperation); |
Sadik Armagan | 8f397a1 | 2022-06-17 15:38:22 +0100 | [diff] [blame] | 1462 | |
| 1463 | LayerInputHandle input0 = ConvertToLayerInputHandle(operation, 0, model, data); |
| 1464 | LayerInputHandle input1 = ConvertToLayerInputHandle(operation, 1, model, data); |
| 1465 | |
| 1466 | if (!input0.IsValid() || !input1.IsValid()) |
| 1467 | { |
| 1468 | return Fail("%s: Operation has invalid inputs", __func__); |
| 1469 | } |
| 1470 | |
Kevin May | 28c3d0f | 2023-07-04 09:07:30 +0100 | [diff] [blame] | 1471 | // The FuseActivation parameter is always the input index 2, and it should be optional |
Sadik Armagan | 8f397a1 | 2022-06-17 15:38:22 +0100 | [diff] [blame] | 1472 | ActivationFn activationFunction; |
| 1473 | if (!GetOptionalInputActivation(operation, 2, activationFunction, model, data)) |
| 1474 | { |
Kevin May | 28c3d0f | 2023-07-04 09:07:30 +0100 | [diff] [blame] | 1475 | return Fail("%s: Operation has invalid optional input: activation function", __func__); |
Sadik Armagan | 8f397a1 | 2022-06-17 15:38:22 +0100 | [diff] [blame] | 1476 | } |
| 1477 | |
| 1478 | const Operand* output = GetOutputOperand(operation, 0, model); |
| 1479 | if (!output) |
| 1480 | { |
Kevin May | 28c3d0f | 2023-07-04 09:07:30 +0100 | [diff] [blame] | 1481 | return Fail("%s: Could not read output", __func__); |
Sadik Armagan | 8f397a1 | 2022-06-17 15:38:22 +0100 | [diff] [blame] | 1482 | } |
| 1483 | |
| 1484 | const armnn::TensorInfo& outputInfo = GetTensorInfoForOperand(*output); |
| 1485 | |
Kevin May | 28c3d0f | 2023-07-04 09:07:30 +0100 | [diff] [blame] | 1486 | armnn::ElementwiseBinaryDescriptor descriptor(binaryOperation); |
| 1487 | |
Sadik Armagan | 8f397a1 | 2022-06-17 15:38:22 +0100 | [diff] [blame] | 1488 | bool isSupported = false; |
| 1489 | auto validateFunc = [&](const armnn::TensorInfo& outputInfo, bool& isSupported) |
| 1490 | { |
| 1491 | FORWARD_LAYER_SUPPORT_FUNC(__func__, |
Kevin May | 28c3d0f | 2023-07-04 09:07:30 +0100 | [diff] [blame] | 1492 | IsElementwiseBinarySupported, |
Sadik Armagan | 8f397a1 | 2022-06-17 15:38:22 +0100 | [diff] [blame] | 1493 | data.m_Backends, |
| 1494 | isSupported, |
Kevin May | 28c3d0f | 2023-07-04 09:07:30 +0100 | [diff] [blame] | 1495 | armnn::BackendId(), |
Sadik Armagan | 8f397a1 | 2022-06-17 15:38:22 +0100 | [diff] [blame] | 1496 | input0.GetTensorInfo(), |
| 1497 | input1.GetTensorInfo(), |
Kevin May | 28c3d0f | 2023-07-04 09:07:30 +0100 | [diff] [blame] | 1498 | outputInfo, |
| 1499 | binaryOperation); |
Sadik Armagan | 8f397a1 | 2022-06-17 15:38:22 +0100 | [diff] [blame] | 1500 | }; |
| 1501 | |
Kevin May | 28c3d0f | 2023-07-04 09:07:30 +0100 | [diff] [blame] | 1502 | if (!IsDynamicTensor(outputInfo)) |
Sadik Armagan | 8f397a1 | 2022-06-17 15:38:22 +0100 | [diff] [blame] | 1503 | { |
| 1504 | validateFunc(outputInfo, isSupported); |
| 1505 | } |
| 1506 | else |
| 1507 | { |
| 1508 | isSupported = AreDynamicTensorsSupported(); |
| 1509 | } |
| 1510 | |
| 1511 | if (!isSupported) |
| 1512 | { |
| 1513 | return false; |
| 1514 | } |
| 1515 | |
Kevin May | 28c3d0f | 2023-07-04 09:07:30 +0100 | [diff] [blame] | 1516 | armnn::IConnectableLayer* layer = data.m_Network->AddElementwiseBinaryLayer(descriptor); |
| 1517 | if (!layer) |
| 1518 | { |
| 1519 | return Fail("%s: Could not add the ElementwiseBinaryLayer", __func__); |
| 1520 | } |
| 1521 | bool isReshapeSupported = BroadcastTensor(input0, input1, layer, data); |
Sadik Armagan | 8f397a1 | 2022-06-17 15:38:22 +0100 | [diff] [blame] | 1522 | if (!isReshapeSupported) |
| 1523 | { |
| 1524 | return false; |
| 1525 | } |
| 1526 | |
Kevin May | 28c3d0f | 2023-07-04 09:07:30 +0100 | [diff] [blame] | 1527 | return SetupAndTrackLayerOutputSlot(operation, 0, *layer, model, |
Sadik Armagan | 8f397a1 | 2022-06-17 15:38:22 +0100 | [diff] [blame] | 1528 | data, nullptr, validateFunc, activationFunction); |
| 1529 | } |
| 1530 | |
| 1531 | bool Converter::ConvertElementwiseUnary(const Operation& operation, |
| 1532 | const Model& model, |
| 1533 | ConversionData& data, |
| 1534 | UnaryOperation unaryOperation) |
| 1535 | { |
| 1536 | VLOG(DRIVER) << "Converter::ConvertElementwiseUnary()"; |
| 1537 | VLOG(DRIVER) << "unaryOperation = " << GetUnaryOperationAsCString(unaryOperation); |
| 1538 | |
| 1539 | LayerInputHandle input = ConvertToLayerInputHandle(operation, 0, model, data); |
| 1540 | |
| 1541 | if (!input.IsValid()) |
| 1542 | { |
| 1543 | return Fail("%s: Operation has invalid input", __func__); |
| 1544 | } |
| 1545 | |
| 1546 | const Operand* output = GetOutputOperand(operation, 0, model); |
| 1547 | if (!output) |
| 1548 | { |
| 1549 | return Fail("%s: Could not read output 0", __func__); |
| 1550 | } |
| 1551 | |
| 1552 | const TensorInfo& inputInfo = input.GetTensorInfo(); |
| 1553 | const TensorInfo& outputInfo = GetTensorInfoForOperand(*output); |
| 1554 | |
| 1555 | ElementwiseUnaryDescriptor descriptor(unaryOperation); |
| 1556 | |
| 1557 | bool isSupported = false; |
Cathal Corbett | 5383767 | 2022-09-01 11:34:37 +0100 | [diff] [blame] | 1558 | armnn::BackendId setBackend; |
Sadik Armagan | 8f397a1 | 2022-06-17 15:38:22 +0100 | [diff] [blame] | 1559 | auto validateFunc = [&](const armnn::TensorInfo& outputInfo, bool& isSupported) |
| 1560 | { |
| 1561 | FORWARD_LAYER_SUPPORT_FUNC(__func__, |
| 1562 | IsElementwiseUnarySupported, |
| 1563 | data.m_Backends, |
| 1564 | isSupported, |
Cathal Corbett | 5383767 | 2022-09-01 11:34:37 +0100 | [diff] [blame] | 1565 | setBackend, |
Sadik Armagan | 8f397a1 | 2022-06-17 15:38:22 +0100 | [diff] [blame] | 1566 | inputInfo, |
| 1567 | outputInfo, |
| 1568 | descriptor); |
| 1569 | }; |
| 1570 | |
| 1571 | if(!IsDynamicTensor(outputInfo)) |
| 1572 | { |
| 1573 | validateFunc(outputInfo, isSupported); |
| 1574 | } |
| 1575 | else |
| 1576 | { |
| 1577 | isSupported = AreDynamicTensorsSupported(); |
| 1578 | } |
| 1579 | |
| 1580 | if (!isSupported) |
| 1581 | { |
| 1582 | return false; |
| 1583 | } |
| 1584 | |
| 1585 | IConnectableLayer* layer = data.m_Network->AddElementwiseUnaryLayer(descriptor); |
Cathal Corbett | 5383767 | 2022-09-01 11:34:37 +0100 | [diff] [blame] | 1586 | layer->SetBackendId(setBackend); |
Sadik Armagan | 8f397a1 | 2022-06-17 15:38:22 +0100 | [diff] [blame] | 1587 | assert(layer != nullptr); |
| 1588 | input.Connect(layer->GetInputSlot(0)); |
| 1589 | |
| 1590 | return SetupAndTrackLayerOutputSlot(operation, 0, *layer, model, data, nullptr, validateFunc); |
| 1591 | } |
| 1592 | |
| 1593 | bool Converter::ConvertElu(const Operation& operation, const Model& model, ConversionData& data) |
| 1594 | { |
| 1595 | VLOG(DRIVER) << "Converter::ConvertElu()"; |
| 1596 | |
| 1597 | LayerInputHandle input0 = ConvertToLayerInputHandle(operation, 0, model, data); |
| 1598 | if (!input0.IsValid()) |
| 1599 | { |
| 1600 | return Fail("%s: Operation has invalid inputs", __func__); |
| 1601 | } |
| 1602 | |
| 1603 | // Determine data type of input tensor |
| 1604 | OperandType inputType; |
| 1605 | if (!GetOperandType(operation, 0, model, inputType)) |
| 1606 | { |
| 1607 | return Fail("%s: Operation has invalid inputs", __func__); |
| 1608 | } |
| 1609 | |
| 1610 | ActivationDescriptor desc; |
| 1611 | desc.m_Function = ActivationFunction::Elu; |
| 1612 | |
| 1613 | // Read alpha |
| 1614 | if (inputType == OperandType::TENSOR_FLOAT16) |
| 1615 | { |
| 1616 | Half alpha; |
| 1617 | |
| 1618 | if (!GetInputScalar(operation, 1, OperandType::FLOAT16, alpha, model, data)) |
| 1619 | { |
| 1620 | return Fail("%s: Operation has invalid inputs (FLOAT16)", __func__); |
| 1621 | } |
| 1622 | |
| 1623 | desc.m_A = static_cast<float>(alpha); |
| 1624 | } |
| 1625 | else if (inputType == OperandType::TENSOR_FLOAT32) |
| 1626 | { |
| 1627 | if (!GetInputScalar(operation, 1, OperandType::FLOAT32, desc.m_A, model, data)) |
| 1628 | { |
| 1629 | return Fail("%s: Operation has invalid inputs (FLOAT32)", __func__); |
| 1630 | } |
| 1631 | } |
| 1632 | else |
| 1633 | { |
| 1634 | return Fail("%s: Unsupported input tensor type: %d", __func__, inputType); |
| 1635 | } |
| 1636 | |
| 1637 | return ::ConvertToActivation(operation, __func__, desc, model, data); |
| 1638 | } |
| 1639 | |
| 1640 | bool Converter::ConvertExpandDims(const Operation& operation, const Model& model, ConversionData& data) |
| 1641 | { |
| 1642 | VLOG(DRIVER) << "Converter::ConvertExpandDims()"; |
| 1643 | |
| 1644 | LayerInputHandle input = ConvertToLayerInputHandle(operation, 0, model, data); |
| 1645 | |
| 1646 | if (!input.IsValid()) |
| 1647 | { |
| 1648 | return Fail("%s: Operation has invalid input", __func__); |
| 1649 | } |
| 1650 | |
| 1651 | const Operand* output = GetOutputOperand(operation, 0, model); |
| 1652 | if (!output) |
| 1653 | { |
| 1654 | return Fail("%s: Operation has invalid output", __func__); |
| 1655 | } |
| 1656 | |
| 1657 | const TensorInfo& outputInfo = GetTensorInfoForOperand(*output); |
| 1658 | |
| 1659 | int32_t axis; |
| 1660 | if (!GetInputScalar(operation, 1, OperandType::INT32, axis, model, data)) |
| 1661 | { |
| 1662 | return Fail("%s: failed to get axis input value", __func__); |
| 1663 | } |
| 1664 | |
| 1665 | TensorShape targetShape; |
| 1666 | |
| 1667 | try |
| 1668 | { |
| 1669 | targetShape = armnnUtils::ExpandDims(input.GetTensorInfo().GetShape(), axis); |
| 1670 | } |
| 1671 | catch (const std::exception& e) |
| 1672 | { |
| 1673 | return Fail("%s: %s", __func__, e.what()); |
| 1674 | } |
| 1675 | |
| 1676 | ReshapeDescriptor reshapeDescriptor; |
| 1677 | reshapeDescriptor.m_TargetShape = targetShape; |
| 1678 | |
| 1679 | bool isSupported = false; |
Cathal Corbett | 5383767 | 2022-09-01 11:34:37 +0100 | [diff] [blame] | 1680 | armnn::BackendId setBackend; |
Sadik Armagan | 8f397a1 | 2022-06-17 15:38:22 +0100 | [diff] [blame] | 1681 | auto validateFunc = [&](const armnn::TensorInfo& outputInfo, bool& isSupported) |
| 1682 | { |
| 1683 | FORWARD_LAYER_SUPPORT_FUNC(__func__, |
| 1684 | IsReshapeSupported, |
| 1685 | data.m_Backends, |
| 1686 | isSupported, |
Cathal Corbett | 5383767 | 2022-09-01 11:34:37 +0100 | [diff] [blame] | 1687 | setBackend, |
Sadik Armagan | 8f397a1 | 2022-06-17 15:38:22 +0100 | [diff] [blame] | 1688 | input.GetTensorInfo(), |
| 1689 | outputInfo, |
| 1690 | reshapeDescriptor); |
| 1691 | }; |
| 1692 | |
| 1693 | if(!IsDynamicTensor(outputInfo)) |
| 1694 | { |
| 1695 | if (targetShape != outputInfo.GetShape()) |
| 1696 | { |
| 1697 | return Fail("%s: Shape of the output operand does not match the resolved expanded shape", __func__); |
| 1698 | } |
| 1699 | validateFunc(outputInfo, isSupported); |
| 1700 | } |
| 1701 | else |
| 1702 | { |
| 1703 | isSupported = AreDynamicTensorsSupported(); |
| 1704 | } |
| 1705 | |
| 1706 | if (!isSupported) |
| 1707 | { |
| 1708 | return false; |
| 1709 | } |
| 1710 | |
| 1711 | IConnectableLayer* layer = data.m_Network->AddReshapeLayer(reshapeDescriptor); |
Cathal Corbett | 5383767 | 2022-09-01 11:34:37 +0100 | [diff] [blame] | 1712 | layer->SetBackendId(setBackend); |
Sadik Armagan | 8f397a1 | 2022-06-17 15:38:22 +0100 | [diff] [blame] | 1713 | assert(layer != nullptr); |
| 1714 | input.Connect(layer->GetInputSlot(0)); |
| 1715 | |
| 1716 | return SetupAndTrackLayerOutputSlot(operation, 0, *layer, model, data, nullptr, validateFunc); |
| 1717 | } |
| 1718 | |
| 1719 | bool Converter::ConvertFill(const Operation& operation, const Model& model, ConversionData& data) |
| 1720 | { |
| 1721 | VLOG(DRIVER) << "Converter::ConvertFill()"; |
| 1722 | LayerInputHandle input = ConvertToLayerInputHandle(operation, 0, model, data); |
| 1723 | if (!input.IsValid()) |
| 1724 | { |
| 1725 | return Fail("%s: Operation has invalid inputs", __func__); |
| 1726 | } |
| 1727 | |
| 1728 | const Operand* output = GetOutputOperand(operation, 0, model); |
| 1729 | if (!output) |
| 1730 | { |
| 1731 | return Fail("%s: Could not read output", __func__); |
| 1732 | } |
| 1733 | |
| 1734 | const TensorInfo& inputInfo = input.GetTensorInfo(); |
| 1735 | const TensorInfo& outputInfo = GetTensorInfoForOperand(*output); |
| 1736 | if (IsDynamicTensor(outputInfo)) |
| 1737 | { |
| 1738 | return Fail("%s: Dynamic output tensors are not supported", __func__); |
| 1739 | } |
| 1740 | |
| 1741 | // Determine data type of output tensor |
| 1742 | OperandType outputType = output->type; |
| 1743 | FillDescriptor descriptor; |
| 1744 | // Read the scalar fill value |
| 1745 | if (outputType == OperandType::TENSOR_FLOAT16) |
| 1746 | { |
| 1747 | Half value; |
| 1748 | |
| 1749 | if (!GetInputScalar(operation, 1, OperandType::FLOAT16, value, model, data)) |
| 1750 | { |
| 1751 | return Fail("%s: Operation has invalid inputs %d", __func__, outputType); |
| 1752 | } |
| 1753 | |
| 1754 | descriptor.m_Value = static_cast<float>(value); |
| 1755 | } |
| 1756 | else if (outputType == OperandType::TENSOR_FLOAT32) |
| 1757 | { |
| 1758 | if (!GetInputScalar(operation, 1, OperandType::FLOAT32, descriptor.m_Value, model, data)) |
| 1759 | { |
| 1760 | return Fail("%s: Operation has invalid inputs %d", __func__, outputType); |
| 1761 | } |
| 1762 | } |
| 1763 | else if (outputType == OperandType::TENSOR_INT32) |
| 1764 | { |
| 1765 | int32_t value; |
| 1766 | |
| 1767 | if (!GetInputScalar(operation, 1, OperandType::INT32, value, model, data)) |
| 1768 | { |
| 1769 | return Fail("%s: Operation has invalid inputs %d", __func__, outputType); |
| 1770 | } |
| 1771 | |
| 1772 | descriptor.m_Value = static_cast<float>(value); |
| 1773 | } |
| 1774 | else |
| 1775 | { |
| 1776 | return Fail("%s: Unsupported input tensor type: %d", __func__, outputType); |
| 1777 | } |
| 1778 | |
| 1779 | bool isSupported = false; |
Cathal Corbett | 5383767 | 2022-09-01 11:34:37 +0100 | [diff] [blame] | 1780 | armnn::BackendId setBackend; |
Sadik Armagan | 8f397a1 | 2022-06-17 15:38:22 +0100 | [diff] [blame] | 1781 | FORWARD_LAYER_SUPPORT_FUNC(__func__, |
| 1782 | IsFillSupported, |
| 1783 | data.m_Backends, |
| 1784 | isSupported, |
Cathal Corbett | 5383767 | 2022-09-01 11:34:37 +0100 | [diff] [blame] | 1785 | setBackend, |
Sadik Armagan | 8f397a1 | 2022-06-17 15:38:22 +0100 | [diff] [blame] | 1786 | inputInfo, |
| 1787 | outputInfo, |
| 1788 | descriptor); |
| 1789 | if (!isSupported) |
| 1790 | { |
| 1791 | return false; |
| 1792 | } |
| 1793 | |
| 1794 | IConnectableLayer* const layer = data.m_Network->AddFillLayer(descriptor); |
Cathal Corbett | 5383767 | 2022-09-01 11:34:37 +0100 | [diff] [blame] | 1795 | layer->SetBackendId(setBackend); |
Sadik Armagan | 8f397a1 | 2022-06-17 15:38:22 +0100 | [diff] [blame] | 1796 | assert(layer != nullptr); |
| 1797 | input.Connect(layer->GetInputSlot(0)); |
| 1798 | |
| 1799 | return SetupAndTrackLayerOutputSlot(operation, 0, *layer, model, data); |
| 1800 | } |
| 1801 | |
| 1802 | bool Converter::ConvertFloor(const Operation& operation, const Model& model, ConversionData& data) |
| 1803 | { |
| 1804 | VLOG(DRIVER) << "Converter::ConvertFloor()"; |
| 1805 | LayerInputHandle input = ConvertToLayerInputHandle(operation, 0, model, data); |
| 1806 | if (!input.IsValid()) |
| 1807 | { |
| 1808 | return Fail("%s: Operation has invalid inputs", __func__); |
| 1809 | } |
| 1810 | |
| 1811 | const Operand* const outputOperand = GetOutputOperand(operation, 0, model); |
| 1812 | if (!outputOperand) |
| 1813 | { |
| 1814 | return Fail("%s: Operation has invalid outputs", __func__); |
| 1815 | } |
| 1816 | |
| 1817 | const armnn::TensorInfo& outputInfo = GetTensorInfoForOperand(*outputOperand); |
| 1818 | |
| 1819 | bool isSupported = false; |
Cathal Corbett | 5383767 | 2022-09-01 11:34:37 +0100 | [diff] [blame] | 1820 | armnn::BackendId setBackend; |
Sadik Armagan | 8f397a1 | 2022-06-17 15:38:22 +0100 | [diff] [blame] | 1821 | auto validateFunc = [&](const armnn::TensorInfo& outputInfo, bool& isSupported) |
| 1822 | { |
| 1823 | FORWARD_LAYER_SUPPORT_FUNC(__func__, |
| 1824 | IsFloorSupported, |
| 1825 | data.m_Backends, |
| 1826 | isSupported, |
Cathal Corbett | 5383767 | 2022-09-01 11:34:37 +0100 | [diff] [blame] | 1827 | setBackend, |
Sadik Armagan | 8f397a1 | 2022-06-17 15:38:22 +0100 | [diff] [blame] | 1828 | input.GetTensorInfo(), |
| 1829 | outputInfo); |
| 1830 | }; |
| 1831 | |
| 1832 | if(!IsDynamicTensor(outputInfo)) |
| 1833 | { |
| 1834 | validateFunc(outputInfo, isSupported); |
| 1835 | } |
| 1836 | else |
| 1837 | { |
| 1838 | isSupported = AreDynamicTensorsSupported(); |
| 1839 | } |
| 1840 | |
| 1841 | if (!isSupported) |
| 1842 | { |
| 1843 | return false; |
| 1844 | } |
| 1845 | |
| 1846 | armnn::IConnectableLayer* layer = data.m_Network->AddFloorLayer(); |
Cathal Corbett | 5383767 | 2022-09-01 11:34:37 +0100 | [diff] [blame] | 1847 | layer->SetBackendId(setBackend); |
Sadik Armagan | 8f397a1 | 2022-06-17 15:38:22 +0100 | [diff] [blame] | 1848 | assert(layer != nullptr); |
| 1849 | input.Connect(layer->GetInputSlot(0)); |
| 1850 | |
| 1851 | return SetupAndTrackLayerOutputSlot(operation, 0, *layer, model, data, nullptr, validateFunc); |
| 1852 | } |
| 1853 | |
| 1854 | bool Converter::ConvertFullyConnected(const Operation& operation, const Model& model, ConversionData& data) |
| 1855 | { |
| 1856 | VLOG(DRIVER) << "Converter::ConvertFullyConnected()"; |
| 1857 | LayerInputHandle input = ConvertToLayerInputHandle(operation, 0, model, data); |
| 1858 | if (!input.IsValid()) |
| 1859 | { |
| 1860 | return Fail("%s: Operation has invalid inputs", __func__); |
| 1861 | } |
| 1862 | |
| 1863 | const Operand* output = GetOutputOperand(operation, 0, model); |
| 1864 | if (!output) |
| 1865 | { |
| 1866 | return Fail("%s: Could not read output 0", __func__); |
| 1867 | } |
| 1868 | |
| 1869 | const armnn::TensorInfo& inputInfo = input.GetTensorInfo(); |
| 1870 | const armnn::TensorInfo& outputInfo = GetTensorInfoForOperand(*output); |
| 1871 | |
| 1872 | LayerInputHandle weightsInput = LayerInputHandle(); |
| 1873 | const Operand* weightsOperand = GetInputOperand(operation, 1, model); |
| 1874 | if (!weightsOperand) |
| 1875 | { |
| 1876 | return Fail("%s: Could not read weights", __func__); |
| 1877 | } |
| 1878 | |
| 1879 | // If weights are constant a separate constant layer will be created to store data. |
| 1880 | // Otherwise handle non const weights as inputs. |
| 1881 | weightsInput = ConvertToLayerInputHandle(operation, 1, model, data); |
| 1882 | if (!weightsInput.IsValid()) |
| 1883 | { |
| 1884 | return Fail("%s: Operation has invalid inputs", __func__); |
| 1885 | } |
| 1886 | |
| 1887 | LayerInputHandle biasInput = LayerInputHandle(); |
| 1888 | const Operand* biasOperand = GetInputOperand(operation, 2, model); |
| 1889 | if (!biasOperand) |
| 1890 | { |
| 1891 | return Fail("%s: Could not read bias", __func__); |
| 1892 | } |
| 1893 | |
| 1894 | // If bias are constant a separate constant layer will be created to store data. |
| 1895 | // Otherwise handle non const bias as inputs. |
| 1896 | biasInput = ConvertToLayerInputHandle(operation, 2, model, data); // 1D |
| 1897 | if (!biasInput.IsValid()) |
| 1898 | { |
| 1899 | return Fail("%s: Operation has invalid inputs", __func__); |
| 1900 | } |
| 1901 | |
| 1902 | armnn::TensorInfo weightsInfo = weightsInput.GetTensorInfo(); |
| 1903 | armnn::TensorInfo reshapedInfo = inputInfo; |
| 1904 | try |
| 1905 | { |
| 1906 | reshapedInfo.SetShape(FlattenFullyConnectedInput(inputInfo.GetShape(), weightsInfo.GetShape())); |
| 1907 | } |
| 1908 | catch (const std::exception& e) |
| 1909 | { |
| 1910 | return Fail("%s: %s", __func__, e.what()); |
| 1911 | } |
| 1912 | |
| 1913 | // Ensuring that the bias value is within 1% of the weights input (small float differences can exist) |
| 1914 | armnn::TensorInfo biasInfo = biasInput.GetTensorInfo(); |
| 1915 | SanitizeBiasQuantizationScale(biasInfo, weightsInfo, reshapedInfo); |
| 1916 | |
| 1917 | ActivationFn activationFunction; |
| 1918 | if (!GetInputActivationFunction(operation, 3, activationFunction, model, data)) |
| 1919 | { |
| 1920 | return Fail("%s: Operation has invalid inputs", __func__); |
| 1921 | } |
| 1922 | |
| 1923 | armnn::FullyConnectedDescriptor desc; |
| 1924 | desc.m_TransposeWeightMatrix = true; |
| 1925 | desc.m_BiasEnabled = true; |
| 1926 | desc.m_ConstantWeights = IsOperandConstant(*weightsOperand); |
| 1927 | |
| 1928 | bool isSupported = false; |
Cathal Corbett | 5383767 | 2022-09-01 11:34:37 +0100 | [diff] [blame] | 1929 | armnn::BackendId setBackend; |
Sadik Armagan | 8f397a1 | 2022-06-17 15:38:22 +0100 | [diff] [blame] | 1930 | auto validateFunc = [&](const armnn::TensorInfo& outputInfo, bool& isSupported) |
| 1931 | { |
| 1932 | if (!VerifyFullyConnectedShapes(reshapedInfo.GetShape(), |
| 1933 | weightsInfo.GetShape(), |
| 1934 | outputInfo.GetShape(), |
| 1935 | desc.m_TransposeWeightMatrix)) |
| 1936 | { |
| 1937 | isSupported = false; |
| 1938 | Fail("%s: Expected outputShape does not match actual outputShape", __func__); |
| 1939 | return; |
| 1940 | } |
| 1941 | |
| 1942 | FORWARD_LAYER_SUPPORT_FUNC(__func__, |
| 1943 | IsFullyConnectedSupported, |
| 1944 | data.m_Backends, |
| 1945 | isSupported, |
Cathal Corbett | 5383767 | 2022-09-01 11:34:37 +0100 | [diff] [blame] | 1946 | setBackend, |
Sadik Armagan | 8f397a1 | 2022-06-17 15:38:22 +0100 | [diff] [blame] | 1947 | reshapedInfo, |
| 1948 | outputInfo, |
| 1949 | weightsInfo, |
| 1950 | biasInfo, |
| 1951 | desc); |
| 1952 | }; |
| 1953 | |
| 1954 | if(!IsDynamicTensor(outputInfo)) |
| 1955 | { |
| 1956 | validateFunc(outputInfo, isSupported); |
| 1957 | } |
| 1958 | else |
| 1959 | { |
| 1960 | isSupported = AreDynamicTensorsSupported(); |
| 1961 | } |
| 1962 | |
| 1963 | if (!isSupported) |
| 1964 | { |
| 1965 | return false; |
| 1966 | } |
| 1967 | |
| 1968 | // Add FullyConnected layer. Weights and bias will be connected as constant layers or non const inputs. |
| 1969 | armnn::IConnectableLayer* startLayer = data.m_Network->AddFullyConnectedLayer(desc); |
Cathal Corbett | 5383767 | 2022-09-01 11:34:37 +0100 | [diff] [blame] | 1970 | startLayer->SetBackendId(setBackend); |
Sadik Armagan | 8f397a1 | 2022-06-17 15:38:22 +0100 | [diff] [blame] | 1971 | |
| 1972 | if (inputInfo.GetNumDimensions() > 2U) |
| 1973 | { |
| 1974 | armnn::ReshapeDescriptor reshapeDescriptor; |
| 1975 | reshapeDescriptor.m_TargetShape = reshapedInfo.GetShape(); |
| 1976 | |
| 1977 | armnn::IConnectableLayer* reshapeLayer = data.m_Network->AddReshapeLayer(reshapeDescriptor); |
| 1978 | assert(reshapeLayer != nullptr); |
| 1979 | input.Connect(reshapeLayer->GetInputSlot(0)); |
| 1980 | reshapeLayer->GetOutputSlot(0).SetTensorInfo(reshapedInfo); |
| 1981 | reshapeLayer->GetOutputSlot(0).Connect(startLayer->GetInputSlot(0)); |
| 1982 | } |
| 1983 | else |
| 1984 | { |
| 1985 | input.Connect(startLayer->GetInputSlot(0)); |
| 1986 | } |
| 1987 | |
| 1988 | // Connect weights and bias inputs |
| 1989 | weightsInput.Connect(startLayer->GetInputSlot(1)); |
| 1990 | biasInput.Connect(startLayer->GetInputSlot(2)); |
| 1991 | |
| 1992 | return SetupAndTrackLayerOutputSlot(operation, 0, *startLayer, model, |
| 1993 | data, nullptr, validateFunc, activationFunction); |
| 1994 | } |
| 1995 | |
| 1996 | bool Converter::ConvertGather(const Operation& operation, const Model& model, ConversionData& data) |
| 1997 | { |
| 1998 | VLOG(DRIVER) << "Converter::ConvertGather()"; |
| 1999 | |
| 2000 | LayerInputHandle input = ConvertToLayerInputHandle(operation, 0, model, data); |
| 2001 | if (!input.IsValid()) |
| 2002 | { |
| 2003 | return Fail("%s: Operation has invalid input", __func__); |
| 2004 | } |
| 2005 | auto inputDimensions = input.GetTensorInfo().GetNumDimensions(); |
| 2006 | |
| 2007 | LayerInputHandle indices = ConvertToLayerInputHandle(operation, 2, model, data); |
| 2008 | if (!indices.IsValid()) |
| 2009 | { |
| 2010 | return Fail("%s: Operation has invalid indices", __func__); |
| 2011 | } |
| 2012 | auto indicesDimensions = indices.GetTensorInfo().GetNumDimensions(); |
| 2013 | |
| 2014 | const Operand* output = GetOutputOperand(operation, 0, model); |
| 2015 | if (!output) |
| 2016 | { |
| 2017 | return Fail("%s: Operation has invalid output", __func__); |
| 2018 | } |
| 2019 | const TensorInfo& outputInfo = GetTensorInfoForOperand(*output); |
| 2020 | auto outputDimensions = outputInfo.GetNumDimensions(); |
| 2021 | if (outputDimensions != inputDimensions + indicesDimensions - 1) |
| 2022 | { |
| 2023 | return Fail("%s: Operation has invalid output dimensions: %d. Output must be an (%d + %d - 1)-D tensor", |
| 2024 | __func__, outputDimensions, inputDimensions, indicesDimensions); |
| 2025 | } |
| 2026 | |
| 2027 | int32_t axis; |
| 2028 | if (!GetInputScalar(operation, 1, OperandType::INT32, axis, model, data)) |
| 2029 | { |
| 2030 | return Fail("%s: Operation has invalid or unsupported axis operand", __func__); |
| 2031 | } |
Kevin May | 28c3d0f | 2023-07-04 09:07:30 +0100 | [diff] [blame] | 2032 | int32_t inputDimensions_int = static_cast<int32_t>(inputDimensions); |
| 2033 | if ((axis < -inputDimensions_int) || (inputDimensions_int <= axis)) |
Sadik Armagan | 8f397a1 | 2022-06-17 15:38:22 +0100 | [diff] [blame] | 2034 | { |
| 2035 | return Fail("%s: Operation has invalid axis: %d. It is out of bounds [-%d, %d))", __func__, axis, |
| 2036 | inputDimensions, inputDimensions); |
| 2037 | } |
| 2038 | |
| 2039 | GatherDescriptor desc; |
| 2040 | desc.m_Axis = axis; |
| 2041 | |
| 2042 | bool isSupported = false; |
Cathal Corbett | 5383767 | 2022-09-01 11:34:37 +0100 | [diff] [blame] | 2043 | armnn::BackendId setBackend; |
Sadik Armagan | 8f397a1 | 2022-06-17 15:38:22 +0100 | [diff] [blame] | 2044 | auto validateFunc = [&](const armnn::TensorInfo& outputInfo, bool& isSupported) |
| 2045 | { |
| 2046 | FORWARD_LAYER_SUPPORT_FUNC(__func__, |
| 2047 | IsGatherSupported, |
| 2048 | data.m_Backends, |
| 2049 | isSupported, |
Cathal Corbett | 5383767 | 2022-09-01 11:34:37 +0100 | [diff] [blame] | 2050 | setBackend, |
Sadik Armagan | 8f397a1 | 2022-06-17 15:38:22 +0100 | [diff] [blame] | 2051 | input.GetTensorInfo(), |
| 2052 | indices.GetTensorInfo(), |
| 2053 | outputInfo, |
| 2054 | desc); |
| 2055 | }; |
| 2056 | |
| 2057 | if(!IsDynamicTensor(outputInfo)) |
| 2058 | { |
| 2059 | validateFunc(outputInfo, isSupported); |
| 2060 | } |
| 2061 | else |
| 2062 | { |
| 2063 | isSupported = AreDynamicTensorsSupported(); |
| 2064 | } |
| 2065 | |
| 2066 | if (!isSupported) |
| 2067 | { |
| 2068 | return false; |
| 2069 | } |
| 2070 | |
| 2071 | IConnectableLayer* layer = data.m_Network->AddGatherLayer(desc); |
Cathal Corbett | 5383767 | 2022-09-01 11:34:37 +0100 | [diff] [blame] | 2072 | layer->SetBackendId(setBackend); |
Sadik Armagan | 8f397a1 | 2022-06-17 15:38:22 +0100 | [diff] [blame] | 2073 | assert(layer != nullptr); |
| 2074 | input.Connect(layer->GetInputSlot(0)); |
| 2075 | indices.Connect(layer->GetInputSlot(1)); |
| 2076 | |
| 2077 | return SetupAndTrackLayerOutputSlot(operation, 0, *layer, model, data, nullptr, validateFunc); |
| 2078 | } |
| 2079 | |
| 2080 | bool Converter::ConvertGroupedConv2d(const Operation& operation, const Model& model, ConversionData& data) |
| 2081 | { |
| 2082 | VLOG(DRIVER) << "Converter::ConvertGroupedConv2d()"; |
| 2083 | // |
| 2084 | // Parse data |
| 2085 | // |
| 2086 | LayerInputHandle input = ConvertToLayerInputHandle(operation, 0, model, data); |
| 2087 | if (!input.IsValid()) |
| 2088 | { |
| 2089 | return Fail("%s: Operation has invalid inputs", __func__); |
| 2090 | } |
| 2091 | const TensorInfo& inputInfo = input.GetTensorInfo(); |
| 2092 | |
| 2093 | const Operand* output = GetOutputOperand(operation, 0, model); |
| 2094 | if (!output) |
| 2095 | { |
| 2096 | return Fail("%s: Could not read output 0", __func__); |
| 2097 | } |
| 2098 | TensorInfo outputInfo = GetTensorInfoForOperand(*output); |
| 2099 | |
| 2100 | // Look ahead to determine data layout |
| 2101 | DataLayout dataLayout = DataLayout::NHWC; |
| 2102 | if (operation.inputs.size() == 12) |
| 2103 | { |
| 2104 | dataLayout = OptionalDataLayout(operation, 11, model, data); |
| 2105 | } |
| 2106 | else |
| 2107 | { |
| 2108 | dataLayout = OptionalDataLayout(operation, 8, model, data); |
| 2109 | } |
| 2110 | |
| 2111 | // NOTE: |
| 2112 | // NNAPI weights are always OHWI, i.e. [depth_out, filter_height, filter_width, depth_group], |
| 2113 | // but Arm NN expects the filter's height and width indices to match the input's height and |
| 2114 | // width indices so when the DataLayout is NCHW, we need to permute the weights to OIHW |
| 2115 | const PermutationVector ohwiToOihw = { 0u, 2u, 3u, 1u }; |
| 2116 | const ConstTensorPin weightsPin = (dataLayout == DataLayout::NCHW) ? |
| 2117 | ConvertOperationInputToConstTensorPin(operation, 1, |
| 2118 | model, data, ohwiToOihw) : |
| 2119 | ConvertOperationInputToConstTensorPin(operation, 1, model, data); |
| 2120 | const ConstTensorPin biasesPin = |
| 2121 | ConvertOperationInputToConstTensorPin(operation, 2, model, data); |
| 2122 | if (!weightsPin.IsValid() || !biasesPin.IsValid()) |
| 2123 | { |
| 2124 | return Fail("%s: Operation has invalid inputs", __func__); |
| 2125 | } |
| 2126 | |
| 2127 | ConstTensor weights = weightsPin.GetConstTensor(); |
| 2128 | ConstTensor biases = biasesPin.GetConstTensor(); |
| 2129 | SanitizeBiasQuantizationScale(biases.GetInfo(), weights.GetInfo(), inputInfo); |
| 2130 | |
| 2131 | const TensorShape& inputShape = inputInfo.GetShape(); |
| 2132 | const TensorShape& outputShape = outputInfo.GetShape(); |
| 2133 | const TensorShape& weightsShape = weights.GetShape(); |
| 2134 | const TensorShape& biasesShape = biases.GetShape(); |
| 2135 | |
| 2136 | armnnUtils::DataLayoutIndexed dataLayoutIndexed(dataLayout); |
| 2137 | const unsigned int channelsIndex = dataLayoutIndexed.GetChannelsIndex(); |
| 2138 | const unsigned int heightIndex = dataLayoutIndexed.GetHeightIndex(); |
| 2139 | const unsigned int widthIndex = dataLayoutIndexed.GetWidthIndex(); |
| 2140 | |
| 2141 | Convolution2dDescriptor desc; |
| 2142 | desc.m_DataLayout = dataLayout; |
| 2143 | desc.m_BiasEnabled = true; |
| 2144 | |
| 2145 | int numGroups; |
| 2146 | ActivationFn activation; |
| 2147 | |
| 2148 | if (operation.inputs.size() == 12) |
| 2149 | { |
| 2150 | if (!GetInputScalar(operation, 3, OperandType::INT32, desc.m_PadLeft, model, data) || |
| 2151 | !GetInputScalar(operation, 4, OperandType::INT32, desc.m_PadRight, model, data) || |
| 2152 | !GetInputScalar(operation, 5, OperandType::INT32, desc.m_PadTop, model, data) || |
| 2153 | !GetInputScalar(operation, 6, OperandType::INT32, desc.m_PadBottom, model, data) || |
| 2154 | !GetInputScalar(operation, 7, OperandType::INT32, desc.m_StrideX, model, data) || |
| 2155 | !GetInputScalar(operation, 8, OperandType::INT32, desc.m_StrideY, model, data) || |
| 2156 | !GetInputScalar(operation, 9, OperandType::INT32, numGroups, model, data) || |
| 2157 | !GetInputActivationFunction(operation, 10, activation, model, data)) |
| 2158 | { |
| 2159 | return Fail("%s: Operation has invalid inputs (explicit padding)", __func__); |
| 2160 | } |
| 2161 | |
| 2162 | } |
| 2163 | else if (operation.inputs.size() == 9) |
| 2164 | { |
| 2165 | ::android::nn::PaddingScheme paddingScheme; |
| 2166 | if (!GetInputPaddingScheme(operation, 3, paddingScheme, model, data) || |
| 2167 | !GetInputScalar(operation, 4, OperandType::INT32, desc.m_StrideX, model, data) || |
| 2168 | !GetInputScalar(operation, 5, OperandType::INT32, desc.m_StrideY, model, data) || |
| 2169 | !GetInputScalar(operation, 6, OperandType::INT32, numGroups, model, data) || |
| 2170 | !GetInputActivationFunction(operation, 7, activation, model, data)) |
| 2171 | { |
| 2172 | return Fail("%s: Operation has invalid inputs (implicit padding)", __func__); |
| 2173 | } |
| 2174 | |
| 2175 | const uint32_t inputX = inputInfo.GetShape()[widthIndex]; |
| 2176 | const uint32_t inputY = inputInfo.GetShape()[heightIndex]; |
| 2177 | |
| 2178 | const uint32_t kernelX = weightsShape[widthIndex]; |
| 2179 | const uint32_t kernelY = weightsShape[heightIndex]; |
| 2180 | |
| 2181 | CalcPadding(inputX, kernelX, desc.m_StrideX, desc.m_PadLeft, desc.m_PadRight, paddingScheme); |
| 2182 | CalcPadding(inputY, kernelY, desc.m_StrideY, desc.m_PadTop, desc.m_PadBottom, paddingScheme); |
| 2183 | } |
| 2184 | else |
| 2185 | { |
| 2186 | return Fail("%s: Unsupported number of operation inputs", __func__); |
| 2187 | } |
| 2188 | |
| 2189 | // Equivalent to outputShape[channelsIndex], but we can't know the outputShape in the case of dynamic tensors |
| 2190 | const unsigned int outputChannels = weightsShape[0]; |
| 2191 | |
| 2192 | const unsigned int channelsPerGroup = weightsShape[channelsIndex]; |
| 2193 | const unsigned int channelMultiplier = outputChannels / numGroups; |
| 2194 | |
| 2195 | // |
| 2196 | // Validate all relevant inputs |
| 2197 | // |
| 2198 | if (numGroups <= 0) |
| 2199 | { |
| 2200 | return Fail("%s: Number of groups must be greater than 0. Got: %d", __func__, numGroups); |
| 2201 | } |
| 2202 | |
| 2203 | if (outputChannels % numGroups != 0u) |
| 2204 | { |
| 2205 | return Fail("%s: Output channels must be divisible by the number of groups", __func__); |
| 2206 | } |
| 2207 | |
| 2208 | // |
| 2209 | // Set up Splitter layer |
| 2210 | // |
| 2211 | unsigned int splitterDimSizes[4] = { inputShape[0], inputShape[1], inputShape[2], inputShape[3] }; |
| 2212 | splitterDimSizes[channelsIndex] /= numGroups; // split in depth |
| 2213 | |
| 2214 | TensorInfo splitterOutputInfo(4, |
| 2215 | splitterDimSizes, |
| 2216 | inputInfo.GetDataType(), |
| 2217 | inputInfo.GetQuantizationScale(), |
| 2218 | inputInfo.GetQuantizationOffset()); |
| 2219 | |
| 2220 | std::vector<std::reference_wrapper<TensorInfo>> splitterOutputInfos(numGroups, std::ref(splitterOutputInfo)); |
| 2221 | |
| 2222 | ViewsDescriptor splitterDesc(numGroups); |
| 2223 | for (unsigned int group = 0u; group < numGroups; ++group) |
| 2224 | { |
| 2225 | splitterDesc.SetViewOriginCoord(group, channelsIndex, splitterDimSizes[channelsIndex] * group); |
| 2226 | for (unsigned int dimIdx = 0u; dimIdx < 4u; dimIdx++) |
| 2227 | { |
| 2228 | splitterDesc.SetViewSize(group, dimIdx, splitterDimSizes[dimIdx]); |
| 2229 | } |
| 2230 | } |
| 2231 | |
| 2232 | bool isSupported = false; |
Cathal Corbett | 5383767 | 2022-09-01 11:34:37 +0100 | [diff] [blame] | 2233 | armnn::BackendId setBackendSplit; |
Sadik Armagan | 8f397a1 | 2022-06-17 15:38:22 +0100 | [diff] [blame] | 2234 | FORWARD_LAYER_SUPPORT_FUNC(__func__, |
| 2235 | IsSplitterSupported, |
| 2236 | data.m_Backends, |
| 2237 | isSupported, |
Cathal Corbett | 5383767 | 2022-09-01 11:34:37 +0100 | [diff] [blame] | 2238 | setBackendSplit, |
Sadik Armagan | 8f397a1 | 2022-06-17 15:38:22 +0100 | [diff] [blame] | 2239 | inputInfo, |
| 2240 | splitterOutputInfos, |
| 2241 | splitterDesc); |
| 2242 | if (!isSupported) |
| 2243 | { |
| 2244 | return false; |
| 2245 | } |
| 2246 | |
| 2247 | IConnectableLayer* splitterLayer = data.m_Network->AddSplitterLayer(splitterDesc); |
Cathal Corbett | 5383767 | 2022-09-01 11:34:37 +0100 | [diff] [blame] | 2248 | splitterLayer->SetBackendId(setBackendSplit); |
Sadik Armagan | 8f397a1 | 2022-06-17 15:38:22 +0100 | [diff] [blame] | 2249 | if (!splitterLayer) |
| 2250 | { |
| 2251 | return Fail("%s: Failed to add SplitterLayer", __func__); |
| 2252 | } |
| 2253 | |
| 2254 | input.Connect(splitterLayer->GetInputSlot(0)); |
| 2255 | for (unsigned int group = 0u; group < splitterLayer->GetNumOutputSlots(); ++group) |
| 2256 | { |
| 2257 | splitterLayer->GetOutputSlot(group).SetTensorInfo(splitterOutputInfo); |
| 2258 | } |
| 2259 | |
| 2260 | // |
| 2261 | // Set up Convolution2d layers for each group |
| 2262 | // |
| 2263 | |
| 2264 | // Set up group tensor shapes |
| 2265 | TensorShape groupInputShape(inputShape); |
| 2266 | groupInputShape[channelsIndex] = channelsPerGroup; |
| 2267 | |
| 2268 | TensorShape groupWeightsShape(weightsShape); |
| 2269 | groupWeightsShape[0] /= channelMultiplier * numGroups; |
| 2270 | |
| 2271 | TensorShape groupBiasesShape({ 1 }); |
| 2272 | |
| 2273 | // Set up group tensor infos |
| 2274 | TensorInfo groupInputInfo(inputInfo); |
| 2275 | groupInputInfo.SetShape(groupInputShape); |
| 2276 | |
| 2277 | const TensorInfo& weightsInfo = weights.GetInfo(); |
| 2278 | TensorInfo groupWeightsInfo(weightsInfo); |
| 2279 | groupWeightsInfo.SetShape(groupWeightsShape); |
| 2280 | |
| 2281 | const TensorInfo& biasesInfo = biases.GetInfo(); |
| 2282 | TensorInfo groupBiasesInfo(biasesInfo); |
| 2283 | groupBiasesInfo.SetShape(groupBiasesShape); |
| 2284 | |
| 2285 | TensorInfo groupOutputInfo(outputInfo); |
| 2286 | |
| 2287 | TensorShape groupOutputShape(outputShape); |
| 2288 | const bool isDynamic = IsDynamicTensor(outputInfo); |
| 2289 | if (!isDynamic) |
| 2290 | { |
| 2291 | groupOutputShape[channelsIndex] = 1; |
| 2292 | } |
| 2293 | groupOutputInfo.SetShape(groupOutputShape); |
| 2294 | |
| 2295 | const unsigned int weightsDataTypeSize = GetDataTypeSize(groupWeightsInfo.GetDataType()); |
| 2296 | const unsigned int biasesDataTypeSize = GetDataTypeSize(groupBiasesInfo.GetDataType()); |
| 2297 | |
| 2298 | std::vector<IConnectableLayer*> convLayers(numGroups * channelMultiplier, nullptr); |
| 2299 | for (unsigned int group = 0u; group < numGroups; ++group) |
| 2300 | { |
| 2301 | for (unsigned int m = 0u; m < channelMultiplier; ++m) |
| 2302 | { |
| 2303 | auto index = group * channelMultiplier + m; |
| 2304 | |
| 2305 | const unsigned int weightsDataOffset = groupWeightsShape.GetNumElements() * index * weightsDataTypeSize; |
| 2306 | const unsigned int biasesDataOffset = groupBiasesShape.GetNumElements() * index * biasesDataTypeSize; |
| 2307 | |
| 2308 | if (weightsInfo.HasPerAxisQuantization()) |
| 2309 | { |
| 2310 | // Extract per-axis quantization scales for group weights |
| 2311 | const std::vector<float>& weightsQuantScales = weightsInfo.GetQuantizationScales(); |
| 2312 | groupWeightsInfo.SetQuantizationScales( |
| 2313 | std::vector<float>(weightsQuantScales.begin() + index, |
| 2314 | weightsQuantScales.begin() + index + groupWeightsShape[0])); |
| 2315 | |
| 2316 | // Extract per-axis quantization scales for group biases |
| 2317 | const std::vector<float>& biasesQuantScales = biasesInfo.GetQuantizationScales(); |
| 2318 | groupBiasesInfo.SetQuantizationScales( |
| 2319 | std::vector<float>(biasesQuantScales.begin() + index, |
| 2320 | biasesQuantScales.begin() + index + groupWeightsShape[0])); |
| 2321 | } |
| 2322 | |
| 2323 | // Extract weights and biases data for current group convolution |
| 2324 | ConstTensor groupWeights(groupWeightsInfo, |
| 2325 | static_cast<const void *>(reinterpret_cast<const char *>(weights.GetMemoryArea()) + |
| 2326 | weightsDataOffset)); |
| 2327 | ConstTensor groupBiases(groupBiasesInfo, |
| 2328 | static_cast<const void *>(reinterpret_cast<const char *>(biases.GetMemoryArea()) + |
| 2329 | biasesDataOffset)); |
| 2330 | |
| 2331 | isSupported = false; |
Cathal Corbett | 5383767 | 2022-09-01 11:34:37 +0100 | [diff] [blame] | 2332 | armnn::BackendId setBackendConv; |
Sadik Armagan | 8f397a1 | 2022-06-17 15:38:22 +0100 | [diff] [blame] | 2333 | auto validateFunc = [&](const armnn::TensorInfo& outputInfo, bool& isSupported) |
| 2334 | { |
| 2335 | FORWARD_LAYER_SUPPORT_FUNC(__func__, |
| 2336 | IsConvolution2dSupported, |
| 2337 | data.m_Backends, |
| 2338 | isSupported, |
Cathal Corbett | 5383767 | 2022-09-01 11:34:37 +0100 | [diff] [blame] | 2339 | setBackendConv, |
Sadik Armagan | 8f397a1 | 2022-06-17 15:38:22 +0100 | [diff] [blame] | 2340 | groupInputInfo, |
| 2341 | outputInfo, |
| 2342 | desc, |
| 2343 | groupWeightsInfo, |
| 2344 | Optional<TensorInfo>(groupBiasesInfo)); |
| 2345 | }; |
| 2346 | |
| 2347 | if(!isDynamic) |
| 2348 | { |
| 2349 | validateFunc(groupOutputInfo, isSupported); |
| 2350 | } |
| 2351 | else |
| 2352 | { |
| 2353 | isSupported = AreDynamicTensorsSupported(); |
| 2354 | } |
| 2355 | |
| 2356 | if (!isSupported) |
| 2357 | { |
| 2358 | return false; |
| 2359 | } |
Teresa Charlin | d936033 | 2022-08-30 14:27:10 +0100 | [diff] [blame] | 2360 | |
| 2361 | IConnectableLayer* weightsLayer = data.m_Network->AddConstantLayer(groupWeights); |
| 2362 | IConnectableLayer* biasLayer = data.m_Network->AddConstantLayer(groupBiases); |
| 2363 | IConnectableLayer* convLayer = data.m_Network->AddConvolution2dLayer(desc); |
| 2364 | |
Cathal Corbett | 5383767 | 2022-09-01 11:34:37 +0100 | [diff] [blame] | 2365 | convLayer->SetBackendId(setBackendConv); |
| 2366 | |
Sadik Armagan | 8f397a1 | 2022-06-17 15:38:22 +0100 | [diff] [blame] | 2367 | if (!convLayer) |
| 2368 | { |
| 2369 | return Fail("%s: AddConvolution2dLayer failed", __func__); |
| 2370 | } |
| 2371 | |
| 2372 | splitterLayer->GetOutputSlot(group).Connect(convLayer->GetInputSlot(0)); |
Teresa Charlin | d936033 | 2022-08-30 14:27:10 +0100 | [diff] [blame] | 2373 | weightsLayer->GetOutputSlot(0).Connect(convLayer->GetInputSlot(1)); |
| 2374 | biasLayer->GetOutputSlot(0).Connect(convLayer->GetInputSlot(2)); |
| 2375 | |
| 2376 | weightsLayer->GetOutputSlot(0).SetTensorInfo(groupWeightsInfo); |
| 2377 | biasLayer->GetOutputSlot(0).SetTensorInfo(groupBiasesInfo); |
Sadik Armagan | 8f397a1 | 2022-06-17 15:38:22 +0100 | [diff] [blame] | 2378 | convLayer->GetOutputSlot(0).SetTensorInfo(groupOutputInfo); |
| 2379 | |
| 2380 | if(isDynamic) |
| 2381 | { |
| 2382 | convLayer->GetOutputSlot(0).IsTensorInfoSet(); |
| 2383 | |
| 2384 | validateFunc(convLayer->GetOutputSlot(0).GetTensorInfo(), isSupported); |
| 2385 | |
| 2386 | outputInfo = convLayer->GetOutputSlot(0).GetTensorInfo(); |
| 2387 | |
| 2388 | if (!isSupported) |
| 2389 | { |
| 2390 | return false; |
| 2391 | } |
| 2392 | } |
| 2393 | |
| 2394 | convLayers[index] = convLayer; |
| 2395 | } |
| 2396 | } |
| 2397 | |
| 2398 | // |
| 2399 | // Set up Concat layer |
| 2400 | // |
| 2401 | ConcatDescriptor concatDescriptor; |
| 2402 | // Equivalent to outputShape[channelsIndex], but we can't know the outputShape in the case of dynamic tensors |
| 2403 | concatDescriptor = ConcatDescriptor(weightsShape[0]); |
| 2404 | for (unsigned int group = 0u; group < numGroups; ++group) |
| 2405 | { |
| 2406 | for (unsigned int m = 0u; m < channelMultiplier; ++m) |
| 2407 | { |
| 2408 | auto index = group * channelMultiplier + m; |
| 2409 | concatDescriptor.SetViewOriginCoord(index, channelsIndex, index); |
| 2410 | concatDescriptor.SetConcatAxis(channelsIndex); |
| 2411 | } |
| 2412 | } |
| 2413 | |
| 2414 | isSupported = false; |
Cathal Corbett | 5383767 | 2022-09-01 11:34:37 +0100 | [diff] [blame] | 2415 | armnn::BackendId setBackendConcat; |
Sadik Armagan | 8f397a1 | 2022-06-17 15:38:22 +0100 | [diff] [blame] | 2416 | FORWARD_LAYER_SUPPORT_FUNC(__func__, |
| 2417 | IsConcatSupported, |
| 2418 | data.m_Backends, |
| 2419 | isSupported, |
Cathal Corbett | 5383767 | 2022-09-01 11:34:37 +0100 | [diff] [blame] | 2420 | setBackendConcat, |
Sadik Armagan | 8f397a1 | 2022-06-17 15:38:22 +0100 | [diff] [blame] | 2421 | std::vector<const TensorInfo*>(numGroups * channelMultiplier, &groupOutputInfo), |
| 2422 | outputInfo, |
| 2423 | concatDescriptor); |
| 2424 | |
| 2425 | if (!isSupported) |
| 2426 | { |
| 2427 | return false; |
| 2428 | } |
| 2429 | |
| 2430 | IConnectableLayer* concatLayer = data.m_Network->AddConcatLayer(concatDescriptor); |
Cathal Corbett | 5383767 | 2022-09-01 11:34:37 +0100 | [diff] [blame] | 2431 | concatLayer->SetBackendId(setBackendConcat); |
Sadik Armagan | 8f397a1 | 2022-06-17 15:38:22 +0100 | [diff] [blame] | 2432 | if (!concatLayer) |
| 2433 | { |
| 2434 | return Fail("%s: AddConcatLayer failed", __func__); |
| 2435 | } |
| 2436 | |
| 2437 | for (unsigned int group = 0u; group < numGroups; ++group) |
| 2438 | { |
| 2439 | for (unsigned int m = 0u; m < channelMultiplier; ++m) |
| 2440 | { |
| 2441 | auto index = group * channelMultiplier + m; |
| 2442 | convLayers[index]->GetOutputSlot(0).Connect(concatLayer->GetInputSlot(index)); |
| 2443 | } |
| 2444 | } |
| 2445 | concatLayer->GetOutputSlot(0).SetTensorInfo(outputInfo); |
| 2446 | |
| 2447 | return SetupAndTrackLayerOutputSlot(operation, 0, *concatLayer, model, |
| 2448 | data, nullptr, nullptr, activation); |
| 2449 | } |
| 2450 | |
| 2451 | bool Converter::ConvertHardSwish(const Operation& operation, const Model& model, ConversionData& data) |
| 2452 | { |
| 2453 | VLOG(DRIVER) << "Converter::ConvertHardSwish()"; |
| 2454 | ActivationDescriptor desc; |
| 2455 | desc.m_Function = ActivationFunction::HardSwish; |
| 2456 | |
| 2457 | return ::ConvertToActivation(operation, __func__, desc, model, data); |
| 2458 | } |
| 2459 | |
| 2460 | bool Converter::ConvertInstanceNormalization(const Operation& operation, const Model& model, ConversionData& data) |
| 2461 | { |
| 2462 | VLOG(DRIVER) << "Converter::ConvertInstanceNormalization()"; |
| 2463 | |
| 2464 | LayerInputHandle input = ConvertToLayerInputHandle(operation, 0, model, data); |
| 2465 | if (!input.IsValid()) |
| 2466 | { |
| 2467 | return Fail("%s: Operation has an invalid input 0", __func__); |
| 2468 | } |
| 2469 | |
| 2470 | const Operand* output = GetOutputOperand(operation, 0, model); |
| 2471 | if (!output) |
| 2472 | { |
| 2473 | return Fail("%s: Operation has an invalid output", __func__); |
| 2474 | } |
| 2475 | |
| 2476 | const TensorInfo& outputInfo = GetTensorInfoForOperand(*output); |
| 2477 | |
| 2478 | // Determine data type of input tensor |
| 2479 | OperandType inputType; |
| 2480 | if (!GetOperandType(operation, 0, model, inputType)) |
| 2481 | { |
| 2482 | return Fail("%s: Operation has invalid inputs", __func__); |
| 2483 | } |
| 2484 | |
| 2485 | InstanceNormalizationDescriptor desc; |
| 2486 | |
| 2487 | // Read gamma, beta & epsilon |
| 2488 | if (inputType == OperandType::TENSOR_FLOAT16) |
| 2489 | { |
| 2490 | Half fp16Gamma; |
| 2491 | Half fp16Beta; |
| 2492 | Half fp16Epsilon; |
| 2493 | |
| 2494 | if (!GetInputScalar(operation, 1, OperandType::FLOAT16, fp16Gamma, model, data) || |
| 2495 | !GetInputScalar(operation, 2, OperandType::FLOAT16, fp16Beta, model, data) || |
| 2496 | !GetInputScalar(operation, 3, OperandType::FLOAT16, fp16Epsilon, model, data)) |
| 2497 | { |
| 2498 | return Fail("%s: Operation has invalid inputs (FLOAT16)", __func__); |
| 2499 | } |
| 2500 | |
| 2501 | desc.m_Gamma = static_cast<float>(fp16Gamma); |
| 2502 | desc.m_Beta = static_cast<float>(fp16Beta); |
| 2503 | desc.m_Eps = static_cast<float>(fp16Epsilon); |
| 2504 | } |
| 2505 | else if (inputType == OperandType::TENSOR_FLOAT32) |
| 2506 | { |
| 2507 | if (!GetInputScalar(operation, 1, OperandType::FLOAT32, desc.m_Gamma, model, data) || |
| 2508 | !GetInputScalar(operation, 2, OperandType::FLOAT32, desc.m_Beta, model, data) || |
| 2509 | !GetInputScalar(operation, 3, OperandType::FLOAT32, desc.m_Eps, model, data)) |
| 2510 | { |
| 2511 | return Fail("%s: Operation has invalid inputs (FLOAT32)", __func__); |
| 2512 | } |
| 2513 | } |
| 2514 | else |
| 2515 | { |
| 2516 | return Fail("%s: Unsupported input tensor type: %d", __func__, inputType); |
| 2517 | } |
| 2518 | |
| 2519 | desc.m_DataLayout = OptionalDataLayout(operation, 4, model, data); |
| 2520 | |
| 2521 | bool isSupported = false; |
Cathal Corbett | 5383767 | 2022-09-01 11:34:37 +0100 | [diff] [blame] | 2522 | armnn::BackendId setBackend; |
Sadik Armagan | 8f397a1 | 2022-06-17 15:38:22 +0100 | [diff] [blame] | 2523 | auto validateFunc = [&](const armnn::TensorInfo& outputInfo, bool& isSupported) |
| 2524 | { |
| 2525 | FORWARD_LAYER_SUPPORT_FUNC(__func__, |
| 2526 | IsInstanceNormalizationSupported, |
| 2527 | data.m_Backends, |
| 2528 | isSupported, |
Cathal Corbett | 5383767 | 2022-09-01 11:34:37 +0100 | [diff] [blame] | 2529 | setBackend, |
Sadik Armagan | 8f397a1 | 2022-06-17 15:38:22 +0100 | [diff] [blame] | 2530 | input.GetTensorInfo(), |
| 2531 | outputInfo, |
| 2532 | desc); |
| 2533 | }; |
| 2534 | |
| 2535 | if(IsDynamicTensor(outputInfo)) |
| 2536 | { |
| 2537 | isSupported = AreDynamicTensorsSupported(); |
| 2538 | } |
| 2539 | else |
| 2540 | { |
| 2541 | validateFunc(outputInfo, isSupported); |
| 2542 | } |
| 2543 | |
| 2544 | if (!isSupported) |
| 2545 | { |
| 2546 | return false; |
| 2547 | } |
| 2548 | |
| 2549 | IConnectableLayer* layer = data.m_Network->AddInstanceNormalizationLayer(desc); |
Cathal Corbett | 5383767 | 2022-09-01 11:34:37 +0100 | [diff] [blame] | 2550 | layer->SetBackendId(setBackend); |
Sadik Armagan | 8f397a1 | 2022-06-17 15:38:22 +0100 | [diff] [blame] | 2551 | input.Connect(layer->GetInputSlot(0)); |
| 2552 | |
| 2553 | return SetupAndTrackLayerOutputSlot(operation, 0, *layer, model, data, nullptr, validateFunc); |
| 2554 | } |
| 2555 | |
| 2556 | bool Converter::ConvertL2Normalization(const Operation& operation, const Model& model, ConversionData& data) |
| 2557 | { |
| 2558 | VLOG(DRIVER) << "Converter::ConvertL2Normalization()"; |
| 2559 | |
| 2560 | if (operation.inputs.size() != 1) |
| 2561 | { |
| 2562 | return Fail("%s: Optional inputs are not supported", __func__); |
| 2563 | } |
| 2564 | |
| 2565 | LayerInputHandle input = ConvertToLayerInputHandle(operation, 0, model, data); |
| 2566 | if (!input.IsValid()) |
| 2567 | { |
| 2568 | return Fail("%s: Operation has invalid inputs", __func__); |
| 2569 | } |
| 2570 | |
| 2571 | const Operand* output = GetOutputOperand(operation, 0, model); |
| 2572 | if (!output) |
| 2573 | { |
| 2574 | return Fail("%s: Could not read output 0", __func__); |
| 2575 | } |
| 2576 | |
| 2577 | const armnn::TensorInfo& inputInfo = input.GetTensorInfo(); |
| 2578 | const armnn::TensorInfo& outputInfo = GetTensorInfoForOperand(*output); |
| 2579 | |
| 2580 | if (outputInfo.GetNumDimensions() != 4u) |
| 2581 | { |
| 2582 | return Fail("%s: Tensor Rank other than 4 is not supported", __func__); |
| 2583 | } |
| 2584 | |
| 2585 | armnn::L2NormalizationDescriptor desc; |
| 2586 | desc.m_DataLayout = armnn::DataLayout::NHWC; |
| 2587 | |
| 2588 | bool isSupported = false; |
Cathal Corbett | 5383767 | 2022-09-01 11:34:37 +0100 | [diff] [blame] | 2589 | armnn::BackendId setBackend; |
Sadik Armagan | 8f397a1 | 2022-06-17 15:38:22 +0100 | [diff] [blame] | 2590 | auto validateFunc = [&](const armnn::TensorInfo& outputInfo, bool& isSupported) |
| 2591 | { |
| 2592 | FORWARD_LAYER_SUPPORT_FUNC(__func__, |
| 2593 | IsL2NormalizationSupported, |
| 2594 | data.m_Backends, |
| 2595 | isSupported, |
Cathal Corbett | 5383767 | 2022-09-01 11:34:37 +0100 | [diff] [blame] | 2596 | setBackend, |
Sadik Armagan | 8f397a1 | 2022-06-17 15:38:22 +0100 | [diff] [blame] | 2597 | inputInfo, |
| 2598 | outputInfo, |
| 2599 | desc); |
| 2600 | }; |
| 2601 | |
| 2602 | if(!IsDynamicTensor(outputInfo)) |
| 2603 | { |
| 2604 | validateFunc(outputInfo, isSupported); |
| 2605 | } |
| 2606 | else |
| 2607 | { |
| 2608 | isSupported = AreDynamicTensorsSupported(); |
| 2609 | } |
| 2610 | |
| 2611 | if (!isSupported) |
| 2612 | { |
| 2613 | return false; |
| 2614 | } |
| 2615 | |
| 2616 | armnn::IConnectableLayer* layer = data.m_Network->AddL2NormalizationLayer(desc); |
Cathal Corbett | 5383767 | 2022-09-01 11:34:37 +0100 | [diff] [blame] | 2617 | layer->SetBackendId(setBackend); |
Sadik Armagan | 8f397a1 | 2022-06-17 15:38:22 +0100 | [diff] [blame] | 2618 | assert(layer != nullptr); |
| 2619 | input.Connect(layer->GetInputSlot(0)); |
| 2620 | |
| 2621 | return SetupAndTrackLayerOutputSlot(operation, 0, *layer, model, data, nullptr, validateFunc); |
| 2622 | } |
| 2623 | |
| 2624 | bool Converter::ConvertL2Pool2d(const Operation& operation, const Model& model, ConversionData& data) |
| 2625 | { |
| 2626 | VLOG(DRIVER) << "Converter::ConvertL2Pool2d()"; |
| 2627 | return ConvertPooling2d(operation, __func__, PoolingAlgorithm::L2, model, data); |
| 2628 | } |
| 2629 | |
| 2630 | bool Converter::ConvertLocalResponseNormalization(const Operation& operation, |
| 2631 | const Model& model, |
| 2632 | ConversionData& data) |
| 2633 | { |
| 2634 | VLOG(DRIVER) << "Converter::ConvertLocalResponseNormalization()"; |
| 2635 | |
| 2636 | if (operation.inputs.size() != 5) |
| 2637 | { |
| 2638 | return Fail("%s: Optional inputs are not supported", __func__); |
| 2639 | } |
| 2640 | |
| 2641 | LayerInputHandle input = ConvertToLayerInputHandle(operation, 0, model, data); |
| 2642 | if (!input.IsValid()) |
| 2643 | { |
| 2644 | return Fail("%s: Operation has invalid inputs", __func__); |
| 2645 | } |
| 2646 | |
| 2647 | const Operand* output = GetOutputOperand(operation, 0, model); |
| 2648 | if (!output) |
| 2649 | { |
| 2650 | return Fail("%s: Could not read output 0", __func__); |
| 2651 | } |
| 2652 | |
| 2653 | const armnn::TensorInfo& inputInfo = input.GetTensorInfo(); |
| 2654 | const armnn::TensorInfo& outputInfo = GetTensorInfoForOperand(*output); |
| 2655 | |
| 2656 | if (outputInfo.GetNumDimensions() != 4u) |
| 2657 | { |
| 2658 | return Fail("%s: Tensor Rank other than 4 is not supported", __func__); |
| 2659 | } |
| 2660 | |
| 2661 | armnn::NormalizationDescriptor descriptor; |
| 2662 | descriptor.m_DataLayout = armnn::DataLayout::NHWC; |
| 2663 | descriptor.m_NormChannelType = armnn::NormalizationAlgorithmChannel::Across; |
| 2664 | descriptor.m_NormMethodType = armnn::NormalizationAlgorithmMethod::LocalBrightness; |
| 2665 | |
| 2666 | if (!input.IsValid() || |
| 2667 | !GetInputScalar(operation, 1, OperandType::INT32, descriptor.m_NormSize, model, data) || |
| 2668 | !GetInputFloat32(operation, 2, descriptor.m_K, model, data) || |
| 2669 | !GetInputFloat32(operation, 3, descriptor.m_Alpha, model, data) || |
| 2670 | !GetInputFloat32(operation, 4, descriptor.m_Beta, model, data)) |
| 2671 | { |
| 2672 | return Fail("%s: Operation has invalid inputs", __func__); |
| 2673 | } |
| 2674 | |
| 2675 | // ArmNN expects normSize to be the full size of the normalization |
| 2676 | // window rather than the radius as in AndroidNN. |
| 2677 | descriptor.m_NormSize = 1 + (2 * descriptor.m_NormSize); |
| 2678 | |
| 2679 | bool isSupported = false; |
Cathal Corbett | 5383767 | 2022-09-01 11:34:37 +0100 | [diff] [blame] | 2680 | armnn::BackendId setBackend; |
Sadik Armagan | 8f397a1 | 2022-06-17 15:38:22 +0100 | [diff] [blame] | 2681 | auto validateFunc = [&](const armnn::TensorInfo& outputInfo, bool& isSupported) |
| 2682 | { |
| 2683 | FORWARD_LAYER_SUPPORT_FUNC(__func__, |
| 2684 | IsNormalizationSupported, |
| 2685 | data.m_Backends, |
| 2686 | isSupported, |
Cathal Corbett | 5383767 | 2022-09-01 11:34:37 +0100 | [diff] [blame] | 2687 | setBackend, |
Sadik Armagan | 8f397a1 | 2022-06-17 15:38:22 +0100 | [diff] [blame] | 2688 | inputInfo, |
| 2689 | outputInfo, |
| 2690 | descriptor); |
| 2691 | }; |
| 2692 | |
| 2693 | if(!IsDynamicTensor(outputInfo)) |
| 2694 | { |
| 2695 | validateFunc(outputInfo, isSupported); |
| 2696 | } |
| 2697 | else |
| 2698 | { |
| 2699 | isSupported = AreDynamicTensorsSupported(); |
| 2700 | } |
| 2701 | |
| 2702 | if (!isSupported) |
| 2703 | { |
| 2704 | return false; |
| 2705 | } |
| 2706 | |
| 2707 | |
| 2708 | armnn::IConnectableLayer* layer = data.m_Network->AddNormalizationLayer(descriptor); |
Cathal Corbett | 5383767 | 2022-09-01 11:34:37 +0100 | [diff] [blame] | 2709 | layer->SetBackendId(setBackend); |
Sadik Armagan | 8f397a1 | 2022-06-17 15:38:22 +0100 | [diff] [blame] | 2710 | assert(layer != nullptr); |
| 2711 | input.Connect(layer->GetInputSlot(0)); |
| 2712 | |
| 2713 | return SetupAndTrackLayerOutputSlot(operation, 0, *layer, model, data, nullptr, validateFunc); |
| 2714 | } |
| 2715 | |
| 2716 | bool Converter::ConvertLogicalBinary(const Operation& operation, |
| 2717 | const Model& model, |
| 2718 | ConversionData& data, |
| 2719 | armnn::LogicalBinaryOperation logicalOperation) |
| 2720 | { |
| 2721 | VLOG(DRIVER) << "Converter::ConvertLogicalBinary()"; |
| 2722 | VLOG(DRIVER) << "ConvertLogicalBinary()"; |
| 2723 | VLOG(DRIVER) << "logicalOperation = " << GetLogicalBinaryOperationAsCString(logicalOperation); |
| 2724 | |
| 2725 | LayerInputHandle input0 = ConvertToLayerInputHandle(operation, 0, model, data); |
| 2726 | LayerInputHandle input1 = ConvertToLayerInputHandle(operation, 1, model, data); |
| 2727 | |
| 2728 | if (!(input0.IsValid() && input1.IsValid())) |
| 2729 | { |
| 2730 | return Fail("%s: Operation has invalid inputs", __func__); |
| 2731 | } |
| 2732 | |
| 2733 | const Operand* output = GetOutputOperand(operation, 0, model); |
| 2734 | if (!output) |
| 2735 | { |
| 2736 | return Fail("%s: Could not read output 0", __func__); |
| 2737 | } |
| 2738 | |
| 2739 | const TensorInfo& inputInfo0 = input0.GetTensorInfo(); |
| 2740 | const TensorInfo& inputInfo1 = input1.GetTensorInfo(); |
| 2741 | const TensorInfo& outputInfo = GetTensorInfoForOperand(*output); |
| 2742 | |
| 2743 | LogicalBinaryDescriptor descriptor(logicalOperation); |
| 2744 | |
| 2745 | bool isSupported = false; |
Cathal Corbett | 5383767 | 2022-09-01 11:34:37 +0100 | [diff] [blame] | 2746 | armnn::BackendId setBackend; |
Sadik Armagan | 8f397a1 | 2022-06-17 15:38:22 +0100 | [diff] [blame] | 2747 | auto validateFunc = [&](const armnn::TensorInfo& outputInfo, bool& isSupported) |
| 2748 | { |
| 2749 | FORWARD_LAYER_SUPPORT_FUNC(__func__, |
| 2750 | IsLogicalBinarySupported, |
| 2751 | data.m_Backends, |
| 2752 | isSupported, |
Cathal Corbett | 5383767 | 2022-09-01 11:34:37 +0100 | [diff] [blame] | 2753 | setBackend, |
Sadik Armagan | 8f397a1 | 2022-06-17 15:38:22 +0100 | [diff] [blame] | 2754 | inputInfo0, |
| 2755 | inputInfo1, |
| 2756 | outputInfo, |
| 2757 | descriptor); |
| 2758 | }; |
| 2759 | |
| 2760 | if(!IsDynamicTensor(outputInfo)) |
| 2761 | { |
| 2762 | validateFunc(outputInfo, isSupported); |
| 2763 | } |
| 2764 | else |
| 2765 | { |
| 2766 | isSupported = AreDynamicTensorsSupported(); |
| 2767 | } |
| 2768 | |
| 2769 | if (!isSupported) |
| 2770 | { |
| 2771 | return false; |
| 2772 | } |
| 2773 | |
| 2774 | IConnectableLayer* layer = data.m_Network->AddLogicalBinaryLayer(descriptor); |
Cathal Corbett | 5383767 | 2022-09-01 11:34:37 +0100 | [diff] [blame] | 2775 | layer->SetBackendId(setBackend); |
Sadik Armagan | 8f397a1 | 2022-06-17 15:38:22 +0100 | [diff] [blame] | 2776 | assert(layer != nullptr); |
| 2777 | |
| 2778 | bool isReshapeSupported = BroadcastTensor(input0, input1, layer, data); |
| 2779 | if (!isReshapeSupported) |
| 2780 | { |
| 2781 | return false; |
| 2782 | } |
| 2783 | |
| 2784 | return SetupAndTrackLayerOutputSlot(operation, 0, *layer, model, data, nullptr, validateFunc); |
| 2785 | } |
| 2786 | |
| 2787 | bool Converter::ConvertLogistic(const Operation& operation, const Model& model, ConversionData& data) |
| 2788 | { |
| 2789 | VLOG(DRIVER) << "Converter::ConvertLogistic()"; |
| 2790 | armnn::ActivationDescriptor desc; |
| 2791 | desc.m_Function = armnn::ActivationFunction::Sigmoid; |
| 2792 | |
| 2793 | return ConvertToActivation(operation, __func__, desc, model, data); |
| 2794 | } |
| 2795 | |
| 2796 | bool Converter::ConvertLogSoftmax(const Operation& operation, const Model& model, ConversionData& data) |
| 2797 | { |
| 2798 | VLOG(DRIVER) << "Converter::ConvertLogSoftmax()"; |
| 2799 | |
| 2800 | LayerInputHandle input = ConvertToLayerInputHandle(operation, 0, model, data); |
| 2801 | if (!input.IsValid()) |
| 2802 | { |
| 2803 | return Fail("%s: Failed to read input 0", __func__); |
| 2804 | } |
| 2805 | |
| 2806 | const Operand* output = GetOutputOperand(operation, 0, model); |
| 2807 | if (!output) |
| 2808 | { |
| 2809 | return Fail("%s: Failed to read output", __func__); |
| 2810 | } |
| 2811 | |
| 2812 | const TensorInfo& outputInfo = GetTensorInfoForOperand(*output); |
| 2813 | |
| 2814 | // Determine data type of input tensor |
| 2815 | OperandType inputType; |
| 2816 | if (!GetOperandType(operation, 0, model, inputType)) |
| 2817 | { |
| 2818 | return Fail("%s: Operation has invalid inputs", __func__); |
| 2819 | } |
| 2820 | |
| 2821 | LogSoftmaxDescriptor descriptor; |
| 2822 | |
| 2823 | // Read beta |
| 2824 | if (inputType == OperandType::TENSOR_FLOAT16) |
| 2825 | { |
| 2826 | Half fp16Beta; |
| 2827 | if (!GetInputScalar(operation, 1, OperandType::FLOAT16, fp16Beta, model, data)) |
| 2828 | { |
| 2829 | return Fail("%s: Failed to read input 1 (FLOAT16)", __func__); |
| 2830 | } |
| 2831 | |
| 2832 | descriptor.m_Beta = static_cast<float>(fp16Beta); |
| 2833 | } |
| 2834 | else if (inputType == OperandType::TENSOR_FLOAT32) |
| 2835 | { |
| 2836 | if (!GetInputScalar(operation, 1, OperandType::FLOAT32, descriptor.m_Beta, model, data)) |
| 2837 | { |
| 2838 | return Fail("%s: Failed to read input 1 (FLOAT32)", __func__); |
| 2839 | } |
| 2840 | } |
| 2841 | else |
| 2842 | { |
| 2843 | return Fail("%s: Unsupported input tensor type: %d", __func__, inputType); |
| 2844 | } |
| 2845 | |
| 2846 | // Read axis |
| 2847 | if (!GetInputInt32(operation, 2, descriptor.m_Axis, model, data)) |
| 2848 | { |
| 2849 | return Fail("%s: Failed to read input 2", __func__); |
| 2850 | } |
| 2851 | |
| 2852 | bool isSupported = false; |
Cathal Corbett | 5383767 | 2022-09-01 11:34:37 +0100 | [diff] [blame] | 2853 | armnn::BackendId setBackend; |
Sadik Armagan | 8f397a1 | 2022-06-17 15:38:22 +0100 | [diff] [blame] | 2854 | auto validateFunc = [&](const armnn::TensorInfo& outputInfo, bool& isSupported) |
| 2855 | { |
| 2856 | FORWARD_LAYER_SUPPORT_FUNC(__func__, |
| 2857 | IsLogSoftmaxSupported, |
| 2858 | data.m_Backends, |
| 2859 | isSupported, |
Cathal Corbett | 5383767 | 2022-09-01 11:34:37 +0100 | [diff] [blame] | 2860 | setBackend, |
Sadik Armagan | 8f397a1 | 2022-06-17 15:38:22 +0100 | [diff] [blame] | 2861 | input.GetTensorInfo(), |
| 2862 | outputInfo, |
| 2863 | descriptor); |
| 2864 | }; |
| 2865 | |
| 2866 | if(IsDynamicTensor(outputInfo)) |
| 2867 | { |
| 2868 | isSupported = AreDynamicTensorsSupported(); |
| 2869 | } |
| 2870 | else |
| 2871 | { |
| 2872 | validateFunc(outputInfo, isSupported); |
| 2873 | } |
| 2874 | |
| 2875 | if (!isSupported) |
| 2876 | { |
| 2877 | return false; |
| 2878 | } |
| 2879 | |
| 2880 | IConnectableLayer* layer = data.m_Network->AddLogSoftmaxLayer(descriptor); |
Cathal Corbett | 5383767 | 2022-09-01 11:34:37 +0100 | [diff] [blame] | 2881 | layer->SetBackendId(setBackend); |
Sadik Armagan | 8f397a1 | 2022-06-17 15:38:22 +0100 | [diff] [blame] | 2882 | if (!layer) |
| 2883 | { |
| 2884 | return Fail("%s: AddLogSoftmaxLayer() returned nullptr", __func__); |
| 2885 | } |
| 2886 | |
| 2887 | input.Connect(layer->GetInputSlot(0)); |
| 2888 | |
| 2889 | return SetupAndTrackLayerOutputSlot(operation, 0, *layer, model, data, nullptr, validateFunc); |
| 2890 | } |
| 2891 | |
| 2892 | bool Converter::ConvertLstm(const Operation& operation, const Model& model, ConversionData& data) |
| 2893 | { |
| 2894 | VLOG(DRIVER) << "Converter::ConvertLstm()"; |
| 2895 | |
| 2896 | // Inputs: |
| 2897 | // 00: The input: A 2-D tensor of ANEURALNETWORKS_TENSOR_FLOAT32, of shape [batch_size, input_size], where |
| 2898 | // “batch_size” corresponds to the batching dimension, and “input_size” is the size of the input. |
| 2899 | LayerInputHandle input = ConvertToLayerInputHandle(operation, 0, model, data); |
| 2900 | if (!input.IsValid()) |
| 2901 | { |
| 2902 | return Fail("%s: Could not read input 0: input", __func__); |
| 2903 | } |
| 2904 | // 18: The output state: A 2-D tensor of ANEURALNETWORKS_TENSOR_FLOAT32, of shape [batch_size, output_size]. |
| 2905 | LayerInputHandle outputStateIn = ConvertToLayerInputHandle(operation, 18, model, data); |
| 2906 | if (!outputStateIn.IsValid()) |
| 2907 | { |
| 2908 | return Fail("%s: Could not read input 18: outputStateIn", __func__); |
| 2909 | } |
| 2910 | // 19: The cell state: A 2-D tensor of ANEURALNETWORKS_TENSOR_FLOAT32, of shape [batch_size, num_units]. |
| 2911 | LayerInputHandle cellStateIn = ConvertToLayerInputHandle(operation, 19, model, data); |
| 2912 | if (!cellStateIn.IsValid()) |
| 2913 | { |
| 2914 | return Fail("%s: Could not read input 19: cellStateIn", __func__); |
| 2915 | } |
| 2916 | |
| 2917 | // Get the mandatory input tensors: |
| 2918 | // 02: The input-to-forget weights: A 2-D tensor of ANEURALNETWORKS_TENSOR_FLOAT32, of shape |
| 2919 | // [num_units, input_size]. |
| 2920 | const ConstTensorPin inputToForgetWeightsPin = |
| 2921 | (DequantizeAndMakeConstTensorPin(operation, model, data, 2)); |
| 2922 | // 03: The input-to-cell weights: A 2-D tensor of ANEURALNETWORKS_TENSOR_FLOAT32, of shape |
| 2923 | // [num_units, input_size]. |
| 2924 | const ConstTensorPin inputToCellWeightsPin = |
| 2925 | (DequantizeAndMakeConstTensorPin(operation, model, data, 3)); |
| 2926 | // 04: The input-to-output weights: A 2-D tensor of ANEURALNETWORKS_TENSOR_FLOAT32, of shape |
| 2927 | // [num_units, input_size]. |
| 2928 | const ConstTensorPin inputToOutputWeightsPin = |
| 2929 | (DequantizeAndMakeConstTensorPin(operation, model, data, 4)); |
| 2930 | // 06: The recurrent-to-forget weights: A 2-D tensor of ANEURALNETWORKS_TENSOR_FLOAT32, of shape |
| 2931 | // [num_units, output_size]. |
| 2932 | const ConstTensorPin recurrentToForgetWeightsPin = |
| 2933 | (DequantizeAndMakeConstTensorPin(operation, model, data, 6)); |
| 2934 | // 07: The recurrent-to-cell weights: A 2-D tensor of ANEURALNETWORKS_TENSOR_FLOAT32, of shape |
| 2935 | // [num_units, output_size]. |
| 2936 | const ConstTensorPin recurrentToCellWeightsPin = |
| 2937 | (DequantizeAndMakeConstTensorPin(operation, model, data, 7)); |
| 2938 | // 08: The recurrent-to-output weights: A 2-D tensor of ANEURALNETWORKS_TENSOR_FLOAT32, of shape |
| 2939 | // [num_units, output_size]. |
| 2940 | const ConstTensorPin recurrentToOutputWeightsPin = |
| 2941 | (DequantizeAndMakeConstTensorPin(operation, model, data, 8)); |
| 2942 | // 13: The forget gate bias: A 1-D tensor of ANEURALNETWORKS_TENSOR_FLOAT32, of shape [num_units]. |
| 2943 | const ConstTensorPin forgetGateBiasPin = |
| 2944 | ConvertOperationInputToConstTensorPin(operation, 13, model, data); |
| 2945 | // 14: The cell bias: A 1-D tensor of ANEURALNETWORKS_TENSOR_FLOAT32, of shape [num_units]. |
| 2946 | const ConstTensorPin cellBiasPin = |
| 2947 | ConvertOperationInputToConstTensorPin(operation, 14, model, data); |
| 2948 | // 15: The output gate bias: A 1-D tensor of ANEURALNETWORKS_TENSOR_FLOAT32, of shape [num_units]. |
| 2949 | const ConstTensorPin outputGateBiasPin = |
| 2950 | ConvertOperationInputToConstTensorPin(operation, 15, model, data); |
| 2951 | |
| 2952 | if (!inputToForgetWeightsPin.IsValid() || |
| 2953 | !inputToCellWeightsPin.IsValid() || |
| 2954 | !inputToOutputWeightsPin.IsValid() || |
| 2955 | !recurrentToForgetWeightsPin.IsValid() || |
| 2956 | !recurrentToCellWeightsPin.IsValid() || |
| 2957 | !recurrentToOutputWeightsPin.IsValid() || |
| 2958 | !forgetGateBiasPin.IsValid() || |
| 2959 | !cellBiasPin.IsValid() || |
| 2960 | !outputGateBiasPin.IsValid()) |
| 2961 | { |
| 2962 | return Fail("%s: Operation has invalid tensor inputs", __func__); |
| 2963 | } |
| 2964 | |
| 2965 | // Get the optional input tensors: |
| 2966 | // 01: The input-to-input weights: Optional. A 2-D tensor of ANEURALNETWORKS_TENSOR_FLOAT32, of shape |
| 2967 | // [num_units, input_size], where “num_units” corresponds to the number of cell units. |
| 2968 | const ConstTensorPin inputToInputWeightsPin = |
| 2969 | (DequantizeAndMakeConstTensorPin(operation, model, data, 1, true)); |
| 2970 | // 05: The recurrent-to-input weights: Optional. A 2-D tensor of ANEURALNETWORKS_TENSOR_FLOAT32, of shape |
| 2971 | // [num_units, output_size], where “output_size” corresponds to either the number of cell units (i.e., |
| 2972 | // “num_units”), or the second dimension of the “projection_weights”, if defined. |
| 2973 | const ConstTensorPin recurrentToInputWeightsPin = |
| 2974 | (DequantizeAndMakeConstTensorPin(operation, model, data, 5, true)); |
| 2975 | // 09: The cell-to-input weights: Optional. A 1-D tensor of ANEURALNETWORKS_TENSOR_FLOAT32, of shape [num_units]. |
| 2976 | const ConstTensorPin cellToInputWeightsPin = |
| 2977 | (DequantizeAndMakeConstTensorPin(operation, model, data, 9, true)); |
| 2978 | // 10: The cell-to-forget weights: Optional. A 1-D tensor of ANEURALNETWORKS_TENSOR_FLOAT32, of shape [num_units]. |
| 2979 | const ConstTensorPin cellToForgetWeightsPin = |
| 2980 | (DequantizeAndMakeConstTensorPin(operation, model, data, 10, true)); |
| 2981 | // 11: The cell-to-output weights: Optional. A 1-D tensor of ANEURALNETWORKS_TENSOR_FLOAT32, of shape [num_units]. |
| 2982 | const ConstTensorPin cellToOutputWeightsPin = |
| 2983 | (DequantizeAndMakeConstTensorPin(operation, model, data, 11, true)); |
| 2984 | // 12: The input gate bias: Optional. A 1-D tensor of ANEURALNETWORKS_TENSOR_FLOAT32, of shape [num_units]. |
| 2985 | const ConstTensorPin inputGateBiasPin = |
| 2986 | ConvertOperationInputToConstTensorPin(operation, |
| 2987 | 12, |
| 2988 | model, |
| 2989 | data, |
| 2990 | g_DontPermute, |
| 2991 | nullptr, |
| 2992 | true); |
| 2993 | |
| 2994 | // 16: The projection weights: Optional. A 2-D tensor of ANEURALNETWORKS_TENSOR_FLOAT32, of shape |
| 2995 | // [output_size, num_units]. |
| 2996 | const ConstTensorPin projectionWeightsPin = |
| 2997 | (DequantizeAndMakeConstTensorPin(operation, model, data, 16, true)); |
| 2998 | // 17: The projection bias: Optional. A 1-D tensor of ANEURALNETWORKS_TENSOR_FLOAT32, of shape [output_size]. |
| 2999 | const ConstTensorPin projectionBiasPin = |
| 3000 | ConvertOperationInputToConstTensorPin(operation, |
| 3001 | 17, |
| 3002 | model, |
| 3003 | data, |
| 3004 | g_DontPermute, |
| 3005 | nullptr, |
| 3006 | true); |
| 3007 | |
| 3008 | if ((!inputToInputWeightsPin.IsValid() && !inputToInputWeightsPin.IsOptional()) || |
| 3009 | (!recurrentToInputWeightsPin.IsValid() && !recurrentToInputWeightsPin.IsOptional()) || |
| 3010 | (!cellToInputWeightsPin.IsValid() && !cellToInputWeightsPin.IsOptional()) || |
| 3011 | (!cellToForgetWeightsPin.IsValid() && !cellToForgetWeightsPin.IsOptional()) || |
| 3012 | (!cellToOutputWeightsPin.IsValid() && !cellToOutputWeightsPin.IsOptional()) || |
| 3013 | (!inputGateBiasPin.IsValid() && !inputGateBiasPin.IsOptional()) || |
| 3014 | (!projectionWeightsPin.IsValid() && !projectionWeightsPin.IsOptional()) || |
| 3015 | (!projectionBiasPin.IsValid() && !projectionBiasPin.IsOptional())) |
| 3016 | { |
| 3017 | return Fail("%s: Operation has invalid tensor inputs", __func__); |
| 3018 | } |
| 3019 | |
| 3020 | // Get the mandatory input scalars (actually 1-D tensors of size 1): |
| 3021 | // 20: The activation function: A value indicating the activation function: |
| 3022 | // 0: None; 1: Relu; 3: Relu6; 4: Tanh; 6: Sigmoid. |
| 3023 | // 21: The clipping threshold: for the cell state, such that values are bound within [-cell_clip, cell_clip]. |
| 3024 | // If set to 0.0 then clipping is disabled. |
| 3025 | // 22: The clipping threshold: for the output from the projection layer, such that values are bound within |
| 3026 | // [-proj_clip, proj_clip]. If set to 0.0 then clipping is disabled. |
| 3027 | ActivationFn activation = ActivationFn::kActivationNone; |
| 3028 | float cellClip; |
| 3029 | float projClip; |
| 3030 | if (!GetInputActivationFunctionFromTensor(operation, 20, activation, model, data) || |
| 3031 | !GetInputScalar(operation, 21, OperandType::FLOAT32, cellClip, model, data) || |
| 3032 | !GetInputScalar(operation, 22, OperandType::FLOAT32, projClip, model, data)) |
| 3033 | { |
| 3034 | return Fail("%s: Operation has invalid scalar inputs", __func__); |
| 3035 | } |
| 3036 | |
| 3037 | // Get the normalization tensors |
| 3038 | // 23: The input layer normalization weights. A 1-D tensor of shape [num_units]. |
| 3039 | // Used to rescale normalized inputs to activation at input gate. |
| 3040 | const ConstTensorPin inputLayerNormWeightsPin |
| 3041 | (DequantizeAndMakeConstTensorPin(operation, model, data, 23, true)); |
| 3042 | |
| 3043 | // 24: The forget layer normalization weights. A 1-D tensor of shape [num_units]. |
| 3044 | // Used to rescale normalized inputs to activation at forget gate. |
| 3045 | const ConstTensorPin forgetLayerNormWeightsPin = |
| 3046 | ConvertOperationInputToConstTensorPin(operation, |
| 3047 | 24, |
| 3048 | model, |
| 3049 | data, |
| 3050 | g_DontPermute, |
| 3051 | nullptr, |
| 3052 | true); |
| 3053 | |
| 3054 | // 25: The cell layer normalization weights. A 1-D tensor of shape [num_units]. |
| 3055 | // Used to rescale normalized inputs to activation at cell gate. |
| 3056 | const ConstTensorPin cellLayerNormWeightsPin = |
| 3057 | ConvertOperationInputToConstTensorPin(operation, |
| 3058 | 25, |
| 3059 | model, |
| 3060 | data, |
| 3061 | g_DontPermute, |
| 3062 | nullptr, |
| 3063 | true); |
| 3064 | |
| 3065 | // 26: The output layer normalization weights. A 1-D tensor of shape [num_units]. |
| 3066 | // Used to rescale normalized inputs to activation at output gate. |
| 3067 | const ConstTensorPin outputLayerNormWeightsPin = |
| 3068 | ConvertOperationInputToConstTensorPin(operation, |
| 3069 | 26, |
| 3070 | model, |
| 3071 | data, |
| 3072 | g_DontPermute, |
| 3073 | nullptr, |
| 3074 | true); |
| 3075 | |
| 3076 | // Outputs: |
| 3077 | // 00: The scratch buffer: A 2-D tensor of ANEURALNETWORKS_TENSOR_FLOAT32, of shape [batch_size, num_units * 4] |
| 3078 | // with CIFG, or [batch_size, num_units * 3] without CIFG. |
| 3079 | const Operand* scratchBuffer = GetOutputOperand(operation, 0, model); |
| 3080 | if (!scratchBuffer) |
| 3081 | { |
| 3082 | return Fail("%s: Could not read output 0: scratchBuffer", __func__); |
| 3083 | } |
| 3084 | // 01: The output state (out): A 2-D tensor of ANEURALNETWORKS_TENSOR_FLOAT32, of shape [batch_size, output_size]. |
| 3085 | const Operand* outputStateOut = GetOutputOperand(operation, 1, model); |
| 3086 | if (!outputStateOut) |
| 3087 | { |
| 3088 | return Fail("%s: Could not read output 1: outputStateOut", __func__); |
| 3089 | } |
| 3090 | // 02: The cell state (out): A 2-D tensor of ANEURALNETWORKS_TENSOR_FLOAT32, of shape [batch_size, num_units]. |
| 3091 | const Operand* cellStateOut = GetOutputOperand(operation, 2, model); |
| 3092 | if (!cellStateOut) |
| 3093 | { |
| 3094 | return Fail("%s: Could not read output 2: cellStateOut", __func__); |
| 3095 | } |
| 3096 | // 03: The output: A 2-D tensor of ANEURALNETWORKS_TENSOR_FLOAT32, of shape [batch_size, output_size]. This is |
| 3097 | // effectively the same as the current “output state (out)” value. |
| 3098 | const Operand* output = GetOutputOperand(operation, 3, model); |
| 3099 | if (!output) |
| 3100 | { |
| 3101 | return Fail("%s: Could not read output 3: output", __func__); |
| 3102 | } |
| 3103 | |
| 3104 | // set the params structure for the AddLstmLayer call |
| 3105 | LstmInputParams params; |
| 3106 | params.m_InputToInputWeights = inputToInputWeightsPin.GetConstTensorPtr(); |
| 3107 | params.m_InputToForgetWeights = inputToForgetWeightsPin.GetConstTensorPtr(); |
| 3108 | params.m_InputToCellWeights = inputToCellWeightsPin.GetConstTensorPtr(); |
| 3109 | params.m_InputToOutputWeights = inputToOutputWeightsPin.GetConstTensorPtr(); |
| 3110 | params.m_RecurrentToInputWeights = recurrentToInputWeightsPin.GetConstTensorPtr(); |
| 3111 | params.m_RecurrentToForgetWeights = recurrentToForgetWeightsPin.GetConstTensorPtr(); |
| 3112 | params.m_RecurrentToCellWeights = recurrentToCellWeightsPin.GetConstTensorPtr(); |
| 3113 | params.m_RecurrentToOutputWeights = recurrentToOutputWeightsPin.GetConstTensorPtr(); |
| 3114 | params.m_CellToInputWeights = cellToInputWeightsPin.GetConstTensorPtr(); |
| 3115 | params.m_CellToForgetWeights = cellToForgetWeightsPin.GetConstTensorPtr(); |
| 3116 | params.m_CellToOutputWeights = cellToOutputWeightsPin.GetConstTensorPtr(); |
| 3117 | params.m_InputGateBias = inputGateBiasPin.GetConstTensorPtr(); |
| 3118 | params.m_ForgetGateBias = forgetGateBiasPin.GetConstTensorPtr(); |
| 3119 | params.m_CellBias = cellBiasPin.GetConstTensorPtr(); |
| 3120 | params.m_OutputGateBias = outputGateBiasPin.GetConstTensorPtr(); |
| 3121 | params.m_ProjectionWeights = projectionWeightsPin.GetConstTensorPtr(); |
| 3122 | params.m_ProjectionBias = projectionBiasPin.GetConstTensorPtr(); |
| 3123 | params.m_InputLayerNormWeights = inputLayerNormWeightsPin.GetConstTensorPtr(); |
| 3124 | params.m_ForgetLayerNormWeights = forgetLayerNormWeightsPin.GetConstTensorPtr(); |
| 3125 | params.m_CellLayerNormWeights = cellLayerNormWeightsPin.GetConstTensorPtr(); |
| 3126 | params.m_OutputLayerNormWeights = outputLayerNormWeightsPin.GetConstTensorPtr(); |
| 3127 | |
| 3128 | // set the layer descriptor |
| 3129 | LstmDescriptor desc; |
| 3130 | desc.m_ActivationFunc = activation; |
| 3131 | desc.m_ClippingThresCell = cellClip; |
| 3132 | desc.m_ClippingThresProj = projClip; |
| 3133 | desc.m_CifgEnabled = (params.m_InputToInputWeights == nullptr || |
| 3134 | params.m_RecurrentToInputWeights == nullptr || |
| 3135 | params.m_InputGateBias == nullptr); |
| 3136 | desc.m_PeepholeEnabled = (params.m_CellToForgetWeights != nullptr || |
| 3137 | params.m_CellToOutputWeights != nullptr); |
| 3138 | desc.m_ProjectionEnabled = (params.m_ProjectionWeights != nullptr); |
| 3139 | desc.m_LayerNormEnabled = (params.m_InputLayerNormWeights != nullptr || |
| 3140 | params.m_ForgetLayerNormWeights != nullptr || |
| 3141 | params.m_CellLayerNormWeights != nullptr || |
| 3142 | params.m_OutputLayerNormWeights != nullptr); |
| 3143 | |
| 3144 | // validate the optional input groups |
| 3145 | if (desc.m_CifgEnabled && |
| 3146 | (params.m_InputToInputWeights != nullptr || |
| 3147 | params.m_RecurrentToInputWeights != nullptr || |
| 3148 | params.m_InputGateBias != nullptr)) |
| 3149 | { |
| 3150 | return Fail("%s: All, or none, of input-to-input weights, recurrent-to-input weights," |
| 3151 | " and input gate bias must be provided", __func__); |
| 3152 | } |
| 3153 | |
| 3154 | if (!desc.m_ProjectionEnabled && params.m_ProjectionBias != nullptr) |
| 3155 | { |
| 3156 | return Fail("%s: projection bias should not be provided without projection weights", __func__); |
| 3157 | } |
| 3158 | |
| 3159 | if (desc.m_PeepholeEnabled && |
| 3160 | (params.m_CellToForgetWeights == nullptr || |
| 3161 | params.m_CellToOutputWeights == nullptr || |
| 3162 | (!desc.m_CifgEnabled && params.m_CellToInputWeights == nullptr))) |
| 3163 | { |
| 3164 | return Fail("%s: All, or none, of cell-to-forget weights and cell-to-output weights must be provided" |
| 3165 | " and, if CIFG is not enabled, cell-to-input weights must also be provided", __func__); |
| 3166 | } |
| 3167 | |
| 3168 | if (desc.m_LayerNormEnabled && |
| 3169 | (params.m_ForgetLayerNormWeights == nullptr || |
| 3170 | params.m_CellLayerNormWeights == nullptr || |
| 3171 | params.m_OutputLayerNormWeights == nullptr || |
| 3172 | (!desc.m_CifgEnabled && params.m_InputLayerNormWeights == nullptr))) |
| 3173 | { |
| 3174 | return Fail("%s: All, or none, of forget-norm weights, cell-norm weights and output-norm weights must be" |
| 3175 | " provided and, if CIFG is not enabled, input-norm weights must also be provided", __func__); |
| 3176 | } |
| 3177 | |
| 3178 | // Check if the layer is supported |
| 3179 | // Inputs |
| 3180 | const TensorInfo& inputInfo = input.GetTensorInfo(); |
| 3181 | const TensorInfo& outputStateInInfo = outputStateIn.GetTensorInfo(); |
| 3182 | const TensorInfo& cellStateInInfo = cellStateIn.GetTensorInfo(); |
| 3183 | |
| 3184 | // Outputs |
| 3185 | const TensorInfo& scratchBufferInfo = GetTensorInfoForOperand(*scratchBuffer); |
| 3186 | const TensorInfo& outputStateOutInfo = GetTensorInfoForOperand(*outputStateOut); |
| 3187 | const TensorInfo& cellStateOutInfo = GetTensorInfoForOperand(*cellStateOut); |
| 3188 | const TensorInfo& outputInfo = GetTensorInfoForOperand(*output); |
| 3189 | |
| 3190 | // Basic parameters |
| 3191 | LstmInputParamsInfo paramsInfo; |
| 3192 | paramsInfo.m_InputToForgetWeights = &(params.m_InputToForgetWeights->GetInfo()); |
| 3193 | paramsInfo.m_InputToCellWeights = &(params.m_InputToCellWeights->GetInfo()); |
| 3194 | paramsInfo.m_InputToOutputWeights = &(params.m_InputToOutputWeights->GetInfo()); |
| 3195 | paramsInfo.m_RecurrentToForgetWeights = &(params.m_RecurrentToForgetWeights->GetInfo()); |
| 3196 | paramsInfo.m_RecurrentToCellWeights = &(params.m_RecurrentToCellWeights->GetInfo()); |
| 3197 | paramsInfo.m_RecurrentToOutputWeights = &(params.m_RecurrentToOutputWeights->GetInfo()); |
| 3198 | paramsInfo.m_ForgetGateBias = &(params.m_ForgetGateBias->GetInfo()); |
| 3199 | paramsInfo.m_CellBias = &(params.m_CellBias->GetInfo()); |
| 3200 | paramsInfo.m_OutputGateBias = &(params.m_OutputGateBias->GetInfo()); |
| 3201 | |
| 3202 | // Optional parameters |
| 3203 | if (!desc.m_CifgEnabled) |
| 3204 | { |
| 3205 | paramsInfo.m_InputToInputWeights = &(params.m_InputToInputWeights->GetInfo()); |
| 3206 | paramsInfo.m_RecurrentToInputWeights = &(params.m_RecurrentToInputWeights->GetInfo()); |
| 3207 | if (params.m_CellToInputWeights != nullptr) |
| 3208 | { |
| 3209 | paramsInfo.m_CellToInputWeights = &(params.m_CellToInputWeights->GetInfo()); |
| 3210 | } |
| 3211 | paramsInfo.m_InputGateBias = &(params.m_InputGateBias->GetInfo()); |
| 3212 | } |
| 3213 | |
| 3214 | if (desc.m_ProjectionEnabled) |
| 3215 | { |
| 3216 | paramsInfo.m_ProjectionWeights = &(params.m_ProjectionWeights->GetInfo()); |
| 3217 | if (params.m_ProjectionBias != nullptr) |
| 3218 | { |
| 3219 | paramsInfo.m_ProjectionBias = &(params.m_ProjectionBias->GetInfo()); |
| 3220 | } |
| 3221 | } |
| 3222 | |
| 3223 | if (desc.m_PeepholeEnabled) |
| 3224 | { |
| 3225 | paramsInfo.m_CellToForgetWeights = &(params.m_CellToForgetWeights->GetInfo()); |
| 3226 | paramsInfo.m_CellToOutputWeights = &(params.m_CellToOutputWeights->GetInfo()); |
| 3227 | } |
| 3228 | |
| 3229 | if (desc.m_LayerNormEnabled) |
| 3230 | { |
| 3231 | if(!desc.m_CifgEnabled) |
| 3232 | { |
| 3233 | paramsInfo.m_InputLayerNormWeights = &(params.m_InputLayerNormWeights->GetInfo()); |
| 3234 | } |
| 3235 | paramsInfo.m_ForgetLayerNormWeights = &(params.m_ForgetLayerNormWeights->GetInfo()); |
| 3236 | paramsInfo.m_CellLayerNormWeights = &(params.m_CellLayerNormWeights->GetInfo()); |
| 3237 | paramsInfo.m_OutputLayerNormWeights = &(params.m_OutputLayerNormWeights->GetInfo()); |
| 3238 | } |
| 3239 | |
| 3240 | bool isSupported = false; |
Cathal Corbett | 5383767 | 2022-09-01 11:34:37 +0100 | [diff] [blame] | 3241 | armnn::BackendId setBackend; |
Sadik Armagan | 8f397a1 | 2022-06-17 15:38:22 +0100 | [diff] [blame] | 3242 | auto validateFunc = [&](const armnn::TensorInfo& outputInfo, bool& isSupported) |
| 3243 | { |
| 3244 | FORWARD_LAYER_SUPPORT_FUNC(__func__, |
| 3245 | IsLstmSupported, |
| 3246 | data.m_Backends, |
| 3247 | isSupported, |
Cathal Corbett | 5383767 | 2022-09-01 11:34:37 +0100 | [diff] [blame] | 3248 | setBackend, |
Sadik Armagan | 8f397a1 | 2022-06-17 15:38:22 +0100 | [diff] [blame] | 3249 | inputInfo, |
| 3250 | outputStateInInfo, |
| 3251 | cellStateInInfo, |
| 3252 | scratchBufferInfo, |
| 3253 | outputStateOutInfo, |
| 3254 | cellStateOutInfo, |
| 3255 | outputInfo, |
| 3256 | desc, |
| 3257 | paramsInfo); |
| 3258 | }; |
| 3259 | |
| 3260 | bool isDynamic = false; |
| 3261 | if (!IsDynamicTensor(outputStateOutInfo) && |
| 3262 | !IsDynamicTensor(scratchBufferInfo) && |
| 3263 | !IsDynamicTensor(cellStateOutInfo) && |
| 3264 | !IsDynamicTensor(outputInfo)) |
| 3265 | { |
| 3266 | validateFunc(outputInfo, isSupported); |
| 3267 | } |
| 3268 | else |
| 3269 | { |
| 3270 | isDynamic = true; |
| 3271 | isSupported = AreDynamicTensorsSupported(); |
| 3272 | } |
| 3273 | |
| 3274 | if (!isSupported) |
| 3275 | { |
| 3276 | return false; |
| 3277 | } |
| 3278 | |
| 3279 | // Add the layer |
| 3280 | IConnectableLayer* layer = data.m_Network->AddLstmLayer(desc, params, "Lstm"); |
Cathal Corbett | 5383767 | 2022-09-01 11:34:37 +0100 | [diff] [blame] | 3281 | layer->SetBackendId(setBackend); |
Sadik Armagan | 8f397a1 | 2022-06-17 15:38:22 +0100 | [diff] [blame] | 3282 | |
| 3283 | input.Connect(layer->GetInputSlot(0)); |
| 3284 | outputStateIn.Connect(layer->GetInputSlot(1)); |
| 3285 | cellStateIn.Connect(layer->GetInputSlot(2)); |
| 3286 | |
| 3287 | if (!isDynamic) |
| 3288 | { |
| 3289 | return ( |
| 3290 | SetupAndTrackLayerOutputSlot(operation, 0, *layer, 0, model, data) && |
| 3291 | SetupAndTrackLayerOutputSlot(operation, 1, *layer, 1, model, data) && |
| 3292 | SetupAndTrackLayerOutputSlot(operation, 2, *layer, 2, model, data) && |
| 3293 | SetupAndTrackLayerOutputSlot(operation, 3, *layer, 3, model, data)); |
| 3294 | } |
| 3295 | else |
| 3296 | { |
| 3297 | return ( |
| 3298 | SetupAndTrackLayerOutputSlot(operation, 0, *layer, 0, model, data) && |
| 3299 | SetupAndTrackLayerOutputSlot(operation, 1, *layer, 1, model, data) && |
| 3300 | SetupAndTrackLayerOutputSlot(operation, 2, *layer, 2, model, data) && |
| 3301 | SetupAndTrackLayerOutputSlot( |
| 3302 | operation, 3, *layer, 3, model, data, nullptr, validateFunc, ActivationFn::kActivationNone, true)); |
| 3303 | } |
| 3304 | |
| 3305 | } |
| 3306 | |
| 3307 | bool Converter::ConvertMaxPool2d(const Operation& operation, const Model& model, ConversionData& data) |
| 3308 | { |
| 3309 | VLOG(DRIVER) << "Converter::ConvertMaxPool2d()"; |
| 3310 | return ConvertPooling2d(operation, __func__, PoolingAlgorithm::Max, model, data); |
| 3311 | } |
| 3312 | |
Sadik Armagan | 8f397a1 | 2022-06-17 15:38:22 +0100 | [diff] [blame] | 3313 | bool Converter::ConvertMean(const Operation& operation, const Model& model, ConversionData& data) |
| 3314 | { |
| 3315 | VLOG(DRIVER) << "Converter::ConvertMean()"; |
| 3316 | |
| 3317 | LayerInputHandle input = ConvertToLayerInputHandle(operation, 0, model, data); |
| 3318 | if (!input.IsValid()) |
| 3319 | { |
| 3320 | return Fail("%s: Operation has invalid inputs", __func__); |
| 3321 | } |
| 3322 | |
| 3323 | const Operand* output = GetOutputOperand(operation, 0, model); |
| 3324 | if (!output) |
| 3325 | { |
| 3326 | return Fail("%s: Could not read output 0", __func__); |
| 3327 | } |
| 3328 | |
| 3329 | const armnn::TensorInfo& outputInfo = GetTensorInfoForOperand(*output); |
| 3330 | |
| 3331 | const Operand* axisOperand = GetInputOperand(operation, 1, model); |
| 3332 | if (!axisOperand) |
| 3333 | { |
| 3334 | return Fail("%s: Could not read input 1", __func__); |
| 3335 | } |
| 3336 | |
| 3337 | std::vector<int32_t> axis; |
| 3338 | if (!GetTensorInt32Values(*axisOperand, axis, model, data)) |
| 3339 | { |
| 3340 | return Fail("%s: Input 1 has invalid values", __func__); |
| 3341 | } |
| 3342 | |
| 3343 | const armnn::TensorInfo& inputInfo = input.GetTensorInfo(); |
| 3344 | |
| 3345 | // Convert the axis to unsigned int and remove duplicates. |
| 3346 | unsigned int rank = inputInfo.GetNumDimensions(); |
| 3347 | std::set<unsigned int> uniqueAxis; |
| 3348 | std::transform(axis.begin(), axis.end(), |
| 3349 | std::inserter(uniqueAxis, uniqueAxis.begin()), |
| 3350 | [rank](int i) -> unsigned int { return (i + rank) % rank; }); |
| 3351 | |
| 3352 | // Get the "keep dims" flag. |
| 3353 | int32_t keepDims = 0; |
| 3354 | if (!GetInputInt32(operation, 2, keepDims, model, data)) |
| 3355 | { |
| 3356 | return Fail("%s: Could not read input 2", __func__); |
| 3357 | } |
| 3358 | |
| 3359 | armnn::MeanDescriptor descriptor; |
| 3360 | descriptor.m_Axis.assign(uniqueAxis.begin(), uniqueAxis.end()); |
| 3361 | descriptor.m_KeepDims = keepDims > 0; |
| 3362 | |
| 3363 | bool isSupported = false; |
Cathal Corbett | 5383767 | 2022-09-01 11:34:37 +0100 | [diff] [blame] | 3364 | armnn::BackendId setBackend; |
Sadik Armagan | 8f397a1 | 2022-06-17 15:38:22 +0100 | [diff] [blame] | 3365 | auto validateFunc = [&](const armnn::TensorInfo& outputInfo, bool& isSupported) |
| 3366 | { |
| 3367 | FORWARD_LAYER_SUPPORT_FUNC(__func__, |
| 3368 | IsMeanSupported, |
| 3369 | data.m_Backends, |
| 3370 | isSupported, |
Cathal Corbett | 5383767 | 2022-09-01 11:34:37 +0100 | [diff] [blame] | 3371 | setBackend, |
Sadik Armagan | 8f397a1 | 2022-06-17 15:38:22 +0100 | [diff] [blame] | 3372 | inputInfo, |
| 3373 | outputInfo, |
| 3374 | descriptor); |
| 3375 | }; |
| 3376 | |
| 3377 | if(!IsDynamicTensor(outputInfo)) |
| 3378 | { |
| 3379 | validateFunc(outputInfo, isSupported); |
| 3380 | } |
| 3381 | else |
| 3382 | { |
| 3383 | isSupported = AreDynamicTensorsSupported(); |
| 3384 | } |
| 3385 | |
| 3386 | if (!isSupported) |
| 3387 | { |
| 3388 | return false; |
| 3389 | } |
| 3390 | |
| 3391 | armnn::IConnectableLayer* const layer = data.m_Network->AddMeanLayer(descriptor); |
Cathal Corbett | 5383767 | 2022-09-01 11:34:37 +0100 | [diff] [blame] | 3392 | layer->SetBackendId(setBackend); |
Sadik Armagan | 8f397a1 | 2022-06-17 15:38:22 +0100 | [diff] [blame] | 3393 | assert(layer != nullptr); |
| 3394 | input.Connect(layer->GetInputSlot(0)); |
| 3395 | |
| 3396 | return SetupAndTrackLayerOutputSlot(operation, 0, *layer, model, data, nullptr, validateFunc); |
| 3397 | } |
| 3398 | |
Sadik Armagan | 8f397a1 | 2022-06-17 15:38:22 +0100 | [diff] [blame] | 3399 | bool Converter::ConvertPad(const Operation& operation, const Model& model, ConversionData& data) |
| 3400 | { |
| 3401 | VLOG(DRIVER) << "Converter::ConvertPad()"; |
| 3402 | |
| 3403 | LayerInputHandle input = ConvertToLayerInputHandle(operation, 0, model, data); |
| 3404 | if (!input.IsValid()) |
| 3405 | { |
| 3406 | return Fail("%s: Operation has invalid inputs", __func__); |
| 3407 | } |
| 3408 | |
| 3409 | const armnn::TensorInfo& inputInfo = input.GetTensorInfo(); |
| 3410 | unsigned int rank = inputInfo.GetNumDimensions(); |
| 3411 | |
| 3412 | armnn::PadDescriptor descriptor; |
| 3413 | if (!ConvertPaddings(operation, model, data, rank, descriptor)) |
| 3414 | { |
| 3415 | return Fail("%s: Could not convert paddings", __func__); |
| 3416 | } |
| 3417 | |
| 3418 | // For a ANEURALNETWORKS_TENSOR_QUANT8_ASYMM and ANEURALNETWORKS_TENSOR_QUANT8_ASYMM_SIGNED tensor, |
| 3419 | // the scale and zeroPoint must be the same as input0 |
| 3420 | // Before Android Q, the pad value for ANEURALNETWORKS_TENSOR_QUANT8_ASYMM was undefined. Since Android Q the pad |
| 3421 | // value must be "logical zero" we set it to be equal to the QuantizationOffset so effectively it ends up as |
| 3422 | // (QuantizationOffset - QuantizationOffset) * scale = 0. |
| 3423 | if (inputInfo.GetDataType() == armnn::DataType::QAsymmU8 || inputInfo.GetDataType() == armnn::DataType::QAsymmS8) |
| 3424 | { |
| 3425 | descriptor.m_PadValue = inputInfo.GetQuantizationOffset(); |
| 3426 | } |
| 3427 | |
| 3428 | const Operand* output = GetOutputOperand(operation, 0, model); |
| 3429 | if (!output) |
| 3430 | { |
| 3431 | return Fail("%s: Could not read output", __func__); |
| 3432 | } |
| 3433 | |
| 3434 | const armnn::TensorInfo& outputInfo = GetTensorInfoForOperand(*output); |
| 3435 | |
| 3436 | bool isSupported = false; |
Cathal Corbett | 5383767 | 2022-09-01 11:34:37 +0100 | [diff] [blame] | 3437 | armnn::BackendId setBackend; |
Sadik Armagan | 8f397a1 | 2022-06-17 15:38:22 +0100 | [diff] [blame] | 3438 | auto validateFunc = [&](const armnn::TensorInfo& outputInfo, bool& isSupported) |
| 3439 | { |
| 3440 | FORWARD_LAYER_SUPPORT_FUNC(__func__, |
| 3441 | IsPadSupported, |
| 3442 | data.m_Backends, |
| 3443 | isSupported, |
Cathal Corbett | 5383767 | 2022-09-01 11:34:37 +0100 | [diff] [blame] | 3444 | setBackend, |
Sadik Armagan | 8f397a1 | 2022-06-17 15:38:22 +0100 | [diff] [blame] | 3445 | inputInfo, |
| 3446 | outputInfo, |
| 3447 | descriptor); |
| 3448 | }; |
| 3449 | |
| 3450 | if(!IsDynamicTensor(outputInfo)) |
| 3451 | { |
| 3452 | validateFunc(outputInfo, isSupported); |
| 3453 | } |
| 3454 | else |
| 3455 | { |
| 3456 | isSupported = AreDynamicTensorsSupported(); |
| 3457 | } |
| 3458 | |
| 3459 | if (!isSupported) |
| 3460 | { |
| 3461 | return false; |
| 3462 | } |
| 3463 | |
| 3464 | armnn::IConnectableLayer* const layer = data.m_Network->AddPadLayer(descriptor); |
Cathal Corbett | 5383767 | 2022-09-01 11:34:37 +0100 | [diff] [blame] | 3465 | layer->SetBackendId(setBackend); |
Sadik Armagan | 8f397a1 | 2022-06-17 15:38:22 +0100 | [diff] [blame] | 3466 | assert(layer != nullptr); |
| 3467 | input.Connect(layer->GetInputSlot(0)); |
| 3468 | |
| 3469 | return SetupAndTrackLayerOutputSlot(operation, 0, *layer, model, data, nullptr, validateFunc); |
| 3470 | } |
| 3471 | |
| 3472 | bool Converter::ConvertPadV2(const Operation& operation, const Model& model, ConversionData& data) |
| 3473 | { |
| 3474 | VLOG(DRIVER) << "Converter::ConvertPadV2()"; |
| 3475 | |
| 3476 | LayerInputHandle input = ConvertToLayerInputHandle(operation, 0, model, data); |
| 3477 | if (!input.IsValid()) |
| 3478 | { |
| 3479 | return Fail("%s: Could not read input 0", __func__); |
| 3480 | } |
| 3481 | |
| 3482 | const Operand* output = GetOutputOperand(operation, 0, model); |
| 3483 | if (!output) |
| 3484 | { |
| 3485 | return Fail("%s: Could not read output", __func__); |
| 3486 | } |
| 3487 | |
| 3488 | const TensorInfo& inputInfo = input.GetTensorInfo(); |
| 3489 | unsigned int rank = inputInfo.GetNumDimensions(); |
| 3490 | |
| 3491 | PadDescriptor descriptor; |
| 3492 | if (!ConvertPaddings(operation, model, data, rank, descriptor)) |
| 3493 | { |
| 3494 | return Fail("%s: Could not convert paddings", __func__); |
| 3495 | } |
| 3496 | |
| 3497 | const TensorInfo& outputInfo = GetTensorInfoForOperand(*output); |
| 3498 | |
| 3499 | // Determine type of padding value |
| 3500 | OperandType operandType0; |
| 3501 | OperandType operandType2; |
| 3502 | |
| 3503 | if (!GetOperandType(operation, 0, model, operandType0) || |
| 3504 | !GetOperandType(operation, 2, model, operandType2)) |
| 3505 | { |
| 3506 | return Fail("%s: Operation has invalid inputs", __func__); |
| 3507 | } |
| 3508 | |
| 3509 | // Read value to use for padding |
| 3510 | if (operandType0 == OperandType::TENSOR_FLOAT16 && operandType2 == OperandType::FLOAT16) |
| 3511 | { |
| 3512 | Half f16PadValue; |
| 3513 | if (!GetInputScalar(operation, 2, operandType2, f16PadValue, model, data)) |
| 3514 | { |
| 3515 | return Fail("%s: Could not read input 2 (FLOAT16)", __func__); |
| 3516 | } |
| 3517 | |
| 3518 | descriptor.m_PadValue = f16PadValue; |
| 3519 | } |
| 3520 | else if (operandType0 == OperandType::TENSOR_FLOAT32 && operandType2 == OperandType::FLOAT32) |
| 3521 | { |
| 3522 | if (!GetInputFloat32(operation, 2, descriptor.m_PadValue, model, data)) |
| 3523 | { |
| 3524 | return Fail("%s: Could not read input 2 (FLOAT32)", __func__); |
| 3525 | } |
| 3526 | } |
| 3527 | else if (isQuantizedOperand(operandType0) && operandType2 == OperandType::INT32) |
| 3528 | { |
| 3529 | int32_t intPadValue = 0; |
| 3530 | if (!GetInputInt32(operation, 2, intPadValue, model, data)) |
| 3531 | { |
| 3532 | return Fail("%s: Could not read input 2 (INT32)", __func__); |
| 3533 | } |
| 3534 | descriptor.m_PadValue = intPadValue; |
| 3535 | } |
| 3536 | else |
| 3537 | { |
| 3538 | return Fail("%s: Operation has invalid inputs: type mismatch", __func__); |
| 3539 | } |
| 3540 | |
| 3541 | bool isSupported = false; |
Cathal Corbett | 5383767 | 2022-09-01 11:34:37 +0100 | [diff] [blame] | 3542 | armnn::BackendId setBackend; |
Sadik Armagan | 8f397a1 | 2022-06-17 15:38:22 +0100 | [diff] [blame] | 3543 | auto validateFunc = [&](const armnn::TensorInfo& outputInfo, bool& isSupported) |
| 3544 | { |
| 3545 | FORWARD_LAYER_SUPPORT_FUNC(__func__, |
| 3546 | IsPadSupported, |
| 3547 | data.m_Backends, |
| 3548 | isSupported, |
Cathal Corbett | 5383767 | 2022-09-01 11:34:37 +0100 | [diff] [blame] | 3549 | setBackend, |
Sadik Armagan | 8f397a1 | 2022-06-17 15:38:22 +0100 | [diff] [blame] | 3550 | inputInfo, |
| 3551 | outputInfo, |
| 3552 | descriptor); |
| 3553 | }; |
| 3554 | |
| 3555 | if(IsDynamicTensor(outputInfo)) |
| 3556 | { |
| 3557 | isSupported = AreDynamicTensorsSupported(); |
| 3558 | } |
| 3559 | else |
| 3560 | { |
| 3561 | validateFunc(outputInfo, isSupported); |
| 3562 | } |
| 3563 | |
| 3564 | if (!isSupported) |
| 3565 | { |
| 3566 | return false; |
| 3567 | } |
| 3568 | |
| 3569 | IConnectableLayer* const layer = data.m_Network->AddPadLayer(descriptor); |
Cathal Corbett | 5383767 | 2022-09-01 11:34:37 +0100 | [diff] [blame] | 3570 | layer->SetBackendId(setBackend); |
Sadik Armagan | 8f397a1 | 2022-06-17 15:38:22 +0100 | [diff] [blame] | 3571 | assert(layer != nullptr); |
| 3572 | input.Connect(layer->GetInputSlot(0)); |
| 3573 | |
| 3574 | return SetupAndTrackLayerOutputSlot(operation, 0, *layer, model, data, nullptr, validateFunc); |
| 3575 | } |
| 3576 | |
| 3577 | bool Converter::ConvertPrelu(const Operation& operation, const Model& model, ConversionData& data) |
| 3578 | { |
| 3579 | VLOG(DRIVER) << "Converter::ConvertPrelu()"; |
| 3580 | |
| 3581 | LayerInputHandle input = ConvertToLayerInputHandle(operation, 0, model, data); |
| 3582 | LayerInputHandle alpha = ConvertToLayerInputHandle(operation, 1, model, data); |
| 3583 | |
| 3584 | if (!input.IsValid() || !alpha.IsValid()) |
| 3585 | { |
| 3586 | return Fail("%s: Operation has invalid inputs", __func__); |
| 3587 | } |
| 3588 | |
| 3589 | const Operand* output = GetOutputOperand(operation, 0, model); |
| 3590 | |
| 3591 | if (!output) |
| 3592 | { |
| 3593 | return Fail("%s: Could not read output", __func__); |
| 3594 | } |
| 3595 | |
| 3596 | const TensorInfo& inputInfo = input.GetTensorInfo(); |
| 3597 | const TensorInfo& alphaInfo = alpha.GetTensorInfo(); |
| 3598 | const TensorInfo& outputInfo = GetTensorInfoForOperand(*output); |
| 3599 | |
| 3600 | bool isSupported = false; |
Cathal Corbett | 5383767 | 2022-09-01 11:34:37 +0100 | [diff] [blame] | 3601 | armnn::BackendId setBackend; |
Sadik Armagan | 8f397a1 | 2022-06-17 15:38:22 +0100 | [diff] [blame] | 3602 | auto validateFunc = [&](const armnn::TensorInfo& outputInfo, bool& isSupported) |
| 3603 | { |
| 3604 | FORWARD_LAYER_SUPPORT_FUNC(__func__, |
| 3605 | IsPreluSupported, |
| 3606 | data.m_Backends, |
| 3607 | isSupported, |
Cathal Corbett | 5383767 | 2022-09-01 11:34:37 +0100 | [diff] [blame] | 3608 | setBackend, |
Sadik Armagan | 8f397a1 | 2022-06-17 15:38:22 +0100 | [diff] [blame] | 3609 | inputInfo, |
| 3610 | alphaInfo, |
| 3611 | outputInfo); |
| 3612 | }; |
| 3613 | |
| 3614 | if(IsDynamicTensor(outputInfo)) |
| 3615 | { |
| 3616 | isSupported = AreDynamicTensorsSupported(); |
| 3617 | } |
| 3618 | else |
| 3619 | { |
| 3620 | validateFunc(outputInfo, isSupported); |
| 3621 | } |
| 3622 | |
| 3623 | if (!isSupported) |
| 3624 | { |
| 3625 | return false; |
| 3626 | } |
| 3627 | |
| 3628 | IConnectableLayer* const layer = data.m_Network->AddPreluLayer(); |
Cathal Corbett | 5383767 | 2022-09-01 11:34:37 +0100 | [diff] [blame] | 3629 | layer->SetBackendId(setBackend); |
Sadik Armagan | 8f397a1 | 2022-06-17 15:38:22 +0100 | [diff] [blame] | 3630 | |
| 3631 | if (!layer) |
| 3632 | { |
| 3633 | return Fail("%s: AddPreluLayer failed", __func__); |
| 3634 | } |
| 3635 | |
| 3636 | bool isReshapeSupported = BroadcastTensor(input, alpha, layer, data); |
| 3637 | if (!isReshapeSupported) |
| 3638 | { |
| 3639 | return false; |
| 3640 | } |
| 3641 | |
| 3642 | return SetupAndTrackLayerOutputSlot(operation, 0, *layer, model, data, nullptr, validateFunc); |
| 3643 | } |
| 3644 | |
| 3645 | bool Converter::ConvertQuantize(const Operation& operation, const Model& model, ConversionData& data) |
| 3646 | { |
| 3647 | VLOG(DRIVER) << "Converter::ConvertQuantize()"; |
| 3648 | |
| 3649 | LayerInputHandle input = ConvertToLayerInputHandle(operation, 0, model, data); |
| 3650 | if (!input.IsValid()) |
| 3651 | { |
| 3652 | return Fail("%s: Operation has invalid input", __func__); |
| 3653 | } |
| 3654 | |
| 3655 | const Operand* const outputOperand = GetOutputOperand(operation, 0, model); |
| 3656 | if (!outputOperand) |
| 3657 | { |
| 3658 | return Fail("%s: Operation has invalid outputs", __func__); |
| 3659 | } |
| 3660 | |
| 3661 | const TensorInfo& outputInfo = GetTensorInfoForOperand(*outputOperand); |
| 3662 | |
| 3663 | bool isSupported = false; |
Cathal Corbett | 5383767 | 2022-09-01 11:34:37 +0100 | [diff] [blame] | 3664 | armnn::BackendId setBackend; |
Sadik Armagan | 8f397a1 | 2022-06-17 15:38:22 +0100 | [diff] [blame] | 3665 | auto validateFunc = [&](const armnn::TensorInfo& outputInfo, bool& isSupported) |
| 3666 | { |
| 3667 | FORWARD_LAYER_SUPPORT_FUNC(__func__, |
| 3668 | IsQuantizeSupported, |
| 3669 | data.m_Backends, |
| 3670 | isSupported, |
Cathal Corbett | 5383767 | 2022-09-01 11:34:37 +0100 | [diff] [blame] | 3671 | setBackend, |
Sadik Armagan | 8f397a1 | 2022-06-17 15:38:22 +0100 | [diff] [blame] | 3672 | input.GetTensorInfo(), |
| 3673 | outputInfo); |
| 3674 | }; |
| 3675 | |
| 3676 | if(IsDynamicTensor(outputInfo)) |
| 3677 | { |
| 3678 | isSupported = AreDynamicTensorsSupported(); |
| 3679 | } |
| 3680 | else |
| 3681 | { |
| 3682 | validateFunc(outputInfo, isSupported); |
| 3683 | } |
| 3684 | |
| 3685 | if (!isSupported) |
| 3686 | { |
| 3687 | return false; |
| 3688 | } |
| 3689 | |
| 3690 | IConnectableLayer* const layer = data.m_Network->AddQuantizeLayer(); |
Cathal Corbett | 5383767 | 2022-09-01 11:34:37 +0100 | [diff] [blame] | 3691 | layer->SetBackendId(setBackend); |
Sadik Armagan | 8f397a1 | 2022-06-17 15:38:22 +0100 | [diff] [blame] | 3692 | assert(layer != nullptr); |
| 3693 | input.Connect(layer->GetInputSlot(0)); |
| 3694 | |
| 3695 | return SetupAndTrackLayerOutputSlot(operation, 0, *layer, model, data, nullptr, validateFunc); |
| 3696 | } |
| 3697 | |
| 3698 | bool Converter::ConvertQuantizedLstm(const Operation& operation, const Model& model, ConversionData& data) |
| 3699 | { |
| 3700 | VLOG(DRIVER) << "Converter::ConvertQuantizedLstm()"; |
| 3701 | |
| 3702 | VLOG(DRIVER) << "ConvertQuantizedLstm()"; |
| 3703 | |
| 3704 | //Inputs: |
| 3705 | // 0: The input: A 2-D tensor of type ANEURALNETWORKS_TENSOR_QUANT8_ASYMM and shape [numBatches, inputSize] |
| 3706 | // specifying the input to the LSTM cell. Tensor is quantized with a fixed quantization range of -1, 127/128. |
| 3707 | LayerInputHandle input = ConvertToLayerInputHandle(operation, 0, model, data); |
| 3708 | if (!input.IsValid()) |
| 3709 | { |
| 3710 | return Fail("%s: Could not read input 0: input", __func__); |
| 3711 | } |
| 3712 | |
| 3713 | // 18: The output state: A 2-D tensor of ANEURALNETWORKS_TENSOR_QUANT8_ASYMM, of shape [batch_size, output_size]. |
| 3714 | LayerInputHandle outputStatePrevTimeStep = ConvertToLayerInputHandle(operation, 18, model, data); |
| 3715 | if (!outputStatePrevTimeStep.IsValid()) |
| 3716 | { |
| 3717 | return Fail("%s: Could not read input 18: outputStatePrevTimeStep", __func__); |
| 3718 | } |
| 3719 | |
| 3720 | // 19: The cell state: A 2-D tensor of ANEURALNETWORKS_TENSOR_QUANT16_SYMM, of shape [batch_size, num_units]. |
| 3721 | LayerInputHandle cellStatePrevTimeStep = ConvertToLayerInputHandle(operation, 19, model, data); |
| 3722 | if (!cellStatePrevTimeStep.IsValid()) |
| 3723 | { |
| 3724 | return Fail("%s: Could not read input 19: cellStatePrevTimeStep", __func__); |
| 3725 | } |
| 3726 | |
| 3727 | // Get the mandatory input tensors: |
| 3728 | |
| 3729 | // 02: The input-to-forget weights: A 2-D tensor of ANEURALNETWORKS_TENSOR_QUANT8_SYMM, of shape |
| 3730 | // [num_units, input_size]. |
| 3731 | const ConstTensorPin inputToForgetWeightsPin = |
| 3732 | ConvertOperationInputToConstTensorPin(operation, 2, model, data); |
| 3733 | |
| 3734 | // 03: The input-to-cell weights: A 2-D tensor of ANEURALNETWORKS_TENSOR_QUANT8_SYMM, of shape |
| 3735 | // [num_units, input_size]. |
| 3736 | const ConstTensorPin inputToCellWeightsPin = |
| 3737 | ConvertOperationInputToConstTensorPin(operation, 3, model, data); |
| 3738 | |
| 3739 | // 04: The input-to-output weights: A 2-D tensor of ANEURALNETWORKS_TENSOR_QUANT8_SYMM, of shape |
| 3740 | // [num_units, input_size]. |
| 3741 | const ConstTensorPin inputToOutputWeightsPin = |
| 3742 | ConvertOperationInputToConstTensorPin(operation, 4, model, data); |
| 3743 | |
| 3744 | // 06: The recurrent-to-forget weights: A 2-D tensor of ANEURALNETWORKS_TENSOR_QUANT8_SYMM, of shape |
| 3745 | // [num_units, output_size]. |
| 3746 | const ConstTensorPin recurrentToForgetWeightsPin = |
| 3747 | ConvertOperationInputToConstTensorPin(operation, 6, model, data); |
| 3748 | |
| 3749 | // 07: The recurrent-to-cell weights: A 2-D tensor of ANEURALNETWORKS_TENSOR_QUANT8_SYMM, of shape |
| 3750 | // [num_units, output_size]. |
| 3751 | const ConstTensorPin recurrentToCellWeightsPin = |
| 3752 | ConvertOperationInputToConstTensorPin(operation, 7, model, data); |
| 3753 | |
| 3754 | // 08: The recurrent-to-output weights: A 2-D tensor of ANEURALNETWORKS_TENSOR_QUANT8_SYMM, of shape |
| 3755 | // [num_units, output_size]. |
| 3756 | const ConstTensorPin recurrentToOutputWeightsPin = |
| 3757 | ConvertOperationInputToConstTensorPin(operation, 8, model, data); |
| 3758 | |
| 3759 | // 13: The forget gate bias: A 1-D tensor of ANEURALNETWORKS_TENSOR_INT32, of shape [num_units]. |
| 3760 | const ConstTensorPin forgetGateBiasPin = |
| 3761 | ConvertOperationInputToConstTensorPin(operation, 13, model, data); |
| 3762 | |
| 3763 | // 14: The cell bias: A 1-D tensor of ANEURALNETWORKS_TENSOR_INT32, of shape [num_units]. |
| 3764 | const ConstTensorPin cellBiasPin = |
| 3765 | ConvertOperationInputToConstTensorPin(operation, 14, model, data); |
| 3766 | |
| 3767 | // 15: The output gate bias: A 1-D tensor of ANEURALNETWORKS_TENSOR_INT32, of shape [num_units]. |
| 3768 | const ConstTensorPin outputGateBiasPin = |
| 3769 | ConvertOperationInputToConstTensorPin(operation, 15, model, data); |
| 3770 | |
| 3771 | if (!inputToForgetWeightsPin.IsValid() || |
| 3772 | !inputToCellWeightsPin.IsValid() || |
| 3773 | !inputToOutputWeightsPin.IsValid() || |
| 3774 | !recurrentToForgetWeightsPin.IsValid() || |
| 3775 | !recurrentToCellWeightsPin.IsValid() || |
| 3776 | !recurrentToOutputWeightsPin.IsValid() || |
| 3777 | !forgetGateBiasPin.IsValid() || |
| 3778 | !cellBiasPin.IsValid() || |
| 3779 | !outputGateBiasPin.IsValid()) |
| 3780 | { |
| 3781 | return Fail("%s: Operation has invalid tensor inputs", __func__); |
| 3782 | } |
| 3783 | |
| 3784 | // Get the optional input tensors: |
| 3785 | |
| 3786 | // 01: The input-to-input weights: Optional. A 2-D tensor of ANEURALNETWORKS_TENSOR_QUANT8_SYMM, of shape |
| 3787 | // [num_units, input_size], where “num_units” corresponds to the number of cell units. |
| 3788 | const ConstTensorPin inputToInputWeightsPin = |
| 3789 | ConvertOperationInputToConstTensorPin(operation, |
| 3790 | 1, |
| 3791 | model, |
| 3792 | data, |
| 3793 | g_DontPermute, |
| 3794 | nullptr, |
| 3795 | true); |
| 3796 | |
| 3797 | // 05: The recurrent-to-input weights: Optional. A 2-D tensor of ANEURALNETWORKS_TENSOR_QUANT8_SYMM, of shape |
| 3798 | // [num_units, output_size], where “output_size” corresponds to either the number of cell units (i.e., |
| 3799 | // “num_units”), or the second dimension of the “projection_weights”, if defined. |
| 3800 | const ConstTensorPin recurrentToInputWeightsPin = |
| 3801 | ConvertOperationInputToConstTensorPin(operation, |
| 3802 | 5, |
| 3803 | model, |
| 3804 | data, |
| 3805 | g_DontPermute, |
| 3806 | nullptr, |
| 3807 | true); |
| 3808 | |
| 3809 | // 09: The cell-to-input weights: Optional. A 1-D tensor of ANEURALNETWORKS_TENSOR_QUANT16_SYMM, of shape |
| 3810 | // [num_units]. |
| 3811 | const ConstTensorPin cellToInputWeightsPin = |
| 3812 | ConvertOperationInputToConstTensorPin(operation, |
| 3813 | 9, |
| 3814 | model, |
| 3815 | data, |
| 3816 | g_DontPermute, |
| 3817 | nullptr, |
| 3818 | true); |
| 3819 | |
| 3820 | // 10: The cell-to-forget weights: Optional. A 1-D tensor of ANEURALNETWORKS_TENSOR_QUANT16_SYMM, of shape |
| 3821 | // [num_units]. |
| 3822 | const ConstTensorPin cellToForgetWeightsPin = |
| 3823 | ConvertOperationInputToConstTensorPin(operation, |
| 3824 | 10, |
| 3825 | model, |
| 3826 | data, |
| 3827 | g_DontPermute, |
| 3828 | nullptr, |
| 3829 | true); |
| 3830 | |
| 3831 | // 11: The cell-to-output weights: Optional. A 1-D tensor of ANEURALNETWORKS_TENSOR_QUANT16_SYMM, of shape |
| 3832 | // [num_units]. |
| 3833 | const ConstTensorPin cellToOutputWeightsPin = |
| 3834 | ConvertOperationInputToConstTensorPin(operation, |
| 3835 | 11, |
| 3836 | model, |
| 3837 | data, |
| 3838 | g_DontPermute, |
| 3839 | nullptr, |
| 3840 | true); |
| 3841 | |
| 3842 | // 12: The input gate bias: Optional. A 1-D tensor of ANEURALNETWORKS_TENSOR_INT32, of shape [num_units]. |
| 3843 | const ConstTensorPin inputGateBiasPin = |
| 3844 | ConvertOperationInputToConstTensorPin(operation, |
| 3845 | 12, |
| 3846 | model, |
| 3847 | data, |
| 3848 | g_DontPermute, |
| 3849 | nullptr, |
| 3850 | true); |
| 3851 | |
| 3852 | // 16: The projection weights: Optional. A 2-D tensor of ANEURALNETWORKS_TENSOR_QUANT8_SYMM, of shape |
| 3853 | // [output_size, num_units]. |
| 3854 | const ConstTensorPin projectionWeightsPin = |
| 3855 | ConvertOperationInputToConstTensorPin(operation, |
| 3856 | 16, |
| 3857 | model, |
| 3858 | data, |
| 3859 | g_DontPermute, |
| 3860 | nullptr, |
| 3861 | true); |
| 3862 | |
| 3863 | // 17: The projection bias: Optional. A 1-D tensor of ANEURALNETWORKS_TENSOR_INT32, of shape [output_size]. |
| 3864 | const ConstTensorPin projectionBiasPin = |
| 3865 | ConvertOperationInputToConstTensorPin(operation, |
| 3866 | 17, |
| 3867 | model, |
| 3868 | data, |
| 3869 | g_DontPermute, |
| 3870 | nullptr, |
| 3871 | true); |
| 3872 | |
| 3873 | if ((!inputToInputWeightsPin.IsValid() && !inputToInputWeightsPin.IsOptional()) |
| 3874 | || (!recurrentToInputWeightsPin.IsValid() && !recurrentToInputWeightsPin.IsOptional()) |
| 3875 | || (!cellToInputWeightsPin.IsValid() && !cellToInputWeightsPin.IsOptional()) |
| 3876 | || (!cellToForgetWeightsPin.IsValid() && !cellToForgetWeightsPin.IsOptional()) |
| 3877 | || (!cellToOutputWeightsPin.IsValid() && !cellToOutputWeightsPin.IsOptional()) |
| 3878 | || (!inputGateBiasPin.IsValid() && !inputGateBiasPin.IsOptional()) |
| 3879 | || (!projectionWeightsPin.IsValid() && !projectionWeightsPin.IsOptional()) |
| 3880 | || (!projectionBiasPin.IsValid() && !projectionBiasPin.IsOptional())) |
| 3881 | { |
| 3882 | return Fail("%s: Operation has invalid tensor inputs", __func__); |
| 3883 | } |
| 3884 | |
| 3885 | |
| 3886 | // Get the optional normalization tensors |
| 3887 | |
| 3888 | // 20: The input layer normalization weights. A 1-D tensor of shape [num_units] ANEURALNETWORKS_TENSOR_QUANT16_SYMM. |
| 3889 | // Used to rescale normalized inputs to activation at input gate. |
| 3890 | const ConstTensorPin inputLayerNormWeightsPin = |
| 3891 | ConvertOperationInputToConstTensorPin(operation, |
| 3892 | 20, |
| 3893 | model, |
| 3894 | data, |
| 3895 | g_DontPermute, |
| 3896 | nullptr, |
| 3897 | true); |
| 3898 | |
| 3899 | // 21: The forget layer normalization weights. A 1-D tensor of shape [num_units] ANEURALNETWORKS_TENSOR_QUANT16_SYMM |
| 3900 | // Used to rescale normalized inputs to activation at forget gate. |
| 3901 | const ConstTensorPin forgetLayerNormWeightsPin = |
| 3902 | ConvertOperationInputToConstTensorPin(operation, |
| 3903 | 21, |
| 3904 | model, |
| 3905 | data, |
| 3906 | g_DontPermute, |
| 3907 | nullptr, |
| 3908 | true); |
| 3909 | |
| 3910 | // 22: The cell layer normalization weights. A 1-D tensor of shape [num_units] ANEURALNETWORKS_TENSOR_QUANT16_SYMM. |
| 3911 | // Used to rescale normalized inputs to activation at cell gate. |
| 3912 | const ConstTensorPin cellLayerNormWeightsPin = |
| 3913 | ConvertOperationInputToConstTensorPin(operation, |
| 3914 | 22, |
| 3915 | model, |
| 3916 | data, |
| 3917 | g_DontPermute, |
| 3918 | nullptr, |
| 3919 | true); |
| 3920 | |
| 3921 | // 23: The output layer normalization weights. A 1-D tensor of shape [num_units]. |
| 3922 | // Used to rescale normalized inputs to activation at output gate. |
| 3923 | const ConstTensorPin outputLayerNormWeightsPin = |
| 3924 | ConvertOperationInputToConstTensorPin(operation, |
| 3925 | 23, |
| 3926 | model, |
| 3927 | data, |
| 3928 | g_DontPermute, |
| 3929 | nullptr, |
| 3930 | true); |
| 3931 | |
| 3932 | if ((!inputLayerNormWeightsPin.IsValid() && !inputLayerNormWeightsPin.IsOptional()) |
| 3933 | || (!forgetLayerNormWeightsPin.IsValid() && !forgetLayerNormWeightsPin.IsOptional()) |
| 3934 | || (!cellLayerNormWeightsPin.IsValid() && !cellLayerNormWeightsPin.IsOptional()) |
| 3935 | || (!outputLayerNormWeightsPin.IsValid() && !outputLayerNormWeightsPin.IsOptional())) |
| 3936 | { |
| 3937 | return Fail("%s: Operation has invalid tensor inputs", __func__); |
| 3938 | } |
| 3939 | |
| 3940 | // Get the optional input scalars: |
| 3941 | // 24: The cell clip: If provided the cell state is clipped by this value prior to the cell output activation. |
| 3942 | // 25: The projection clip: If provided and projection is enabled, this is used for clipping the projected values. |
| 3943 | |
| 3944 | // Get the mandatory input scalars: |
| 3945 | // 26: The scale of the intermediate result of matmul, i.e. input to layer normalization, at input gate. |
| 3946 | // 27: The scale of the intermediate result of matmul, i.e. input to layer normalization, at forget gate. |
| 3947 | // 28: The scale of the intermediate result of matmul, i.e. input to layer normalization, at cell gate. |
| 3948 | // 29: The scale of the intermediate result of matmul, i.e. input to layer normalization, at output gate. |
| 3949 | // 30: The zero point of the hidden state, i.e. input to projection. |
| 3950 | // 31: The scale of the hidden state, i.e. input to projection. |
| 3951 | float cellClip, projClip, matMulInputGate, matMulForgetGate, matMulCellGate, matMulOutputGate, projInputScale; |
| 3952 | int projInputZeroPoint; |
| 3953 | |
| 3954 | if (!GetInputScalar(operation, 24, OperandType::FLOAT32, cellClip, model, data, true) || |
| 3955 | !GetInputScalar(operation, 25, OperandType::FLOAT32, projClip, model, data, true) || |
| 3956 | !GetInputScalar(operation, 26, OperandType::FLOAT32, matMulInputGate, model, data) || |
| 3957 | !GetInputScalar(operation, 27, OperandType::FLOAT32, matMulForgetGate, model, data) || |
| 3958 | !GetInputScalar(operation, 28, OperandType::FLOAT32, matMulCellGate, model, data) || |
| 3959 | !GetInputScalar(operation, 29, OperandType::FLOAT32, matMulOutputGate, model, data) || |
| 3960 | !GetInputScalar(operation, 30, OperandType::INT32, projInputZeroPoint, model, data) || |
| 3961 | !GetInputScalar(operation, 31, OperandType::FLOAT32, projInputScale, model, data)) |
| 3962 | { |
| 3963 | return Fail("%s: Operation has invalid scalar inputs", __func__); |
| 3964 | } |
| 3965 | |
| 3966 | // Outputs: |
| 3967 | // 0: The output state (out): A 2-D tensor of ANEURALNETWORKS_TENSOR_QUANT8_ASYMM_SIGNED, of shape [batch_size, |
| 3968 | // output_size]. |
| 3969 | const Operand* outputStateOut = GetOutputOperand(operation, 0, model); |
| 3970 | if (!outputStateOut) |
| 3971 | { |
| 3972 | return Fail("%s: Could not read output 0: outputStateOut", __func__); |
| 3973 | } |
| 3974 | |
| 3975 | // 1: The cell state (out): A 2-D tensor of ANEURALNETWORKS_TENSOR_QUANT16_SYMM, of shape [batch_size, num_units]. |
| 3976 | const Operand* cellStateOut = GetOutputOperand(operation, 1, model); |
| 3977 | if (!cellStateOut) |
| 3978 | { |
| 3979 | return Fail("%s: Could not read output 1: cellStateOut", __func__); |
| 3980 | } |
| 3981 | |
| 3982 | // 2: The output: A 2-D tensor of ANEURALNETWORKS_TENSOR_QUANT8_ASYMM_SIGNED, of shape [batch_size, output_size]. |
| 3983 | // This is effectively the same as the current “output state (out)” value. |
| 3984 | const Operand* output = GetOutputOperand(operation, 2, model); |
| 3985 | if (!output) |
| 3986 | { |
| 3987 | return Fail("%s: Could not read output 2: output", __func__); |
| 3988 | } |
| 3989 | |
| 3990 | // set the params structure for the AddLstmLayer call |
| 3991 | LstmInputParams params; |
| 3992 | params.m_InputToInputWeights = inputToInputWeightsPin.GetConstTensorPtr(); |
| 3993 | params.m_InputToForgetWeights = inputToForgetWeightsPin.GetConstTensorPtr(); |
| 3994 | params.m_InputToCellWeights = inputToCellWeightsPin.GetConstTensorPtr(); |
| 3995 | params.m_InputToOutputWeights = inputToOutputWeightsPin.GetConstTensorPtr(); |
| 3996 | params.m_RecurrentToInputWeights = recurrentToInputWeightsPin.GetConstTensorPtr(); |
| 3997 | params.m_RecurrentToForgetWeights = recurrentToForgetWeightsPin.GetConstTensorPtr(); |
| 3998 | params.m_RecurrentToCellWeights = recurrentToCellWeightsPin.GetConstTensorPtr(); |
| 3999 | params.m_RecurrentToOutputWeights = recurrentToOutputWeightsPin.GetConstTensorPtr(); |
| 4000 | params.m_CellToInputWeights = cellToInputWeightsPin.GetConstTensorPtr(); |
| 4001 | params.m_CellToForgetWeights = cellToForgetWeightsPin.GetConstTensorPtr(); |
| 4002 | params.m_CellToOutputWeights = cellToOutputWeightsPin.GetConstTensorPtr(); |
| 4003 | params.m_InputGateBias = inputGateBiasPin.GetConstTensorPtr(); |
| 4004 | params.m_ForgetGateBias = forgetGateBiasPin.GetConstTensorPtr(); |
| 4005 | params.m_CellBias = cellBiasPin.GetConstTensorPtr(); |
| 4006 | params.m_OutputGateBias = outputGateBiasPin.GetConstTensorPtr(); |
| 4007 | params.m_ProjectionWeights = projectionWeightsPin.GetConstTensorPtr(); |
| 4008 | params.m_ProjectionBias = projectionBiasPin.GetConstTensorPtr(); |
| 4009 | params.m_InputLayerNormWeights = inputLayerNormWeightsPin.GetConstTensorPtr(); |
| 4010 | params.m_ForgetLayerNormWeights = forgetLayerNormWeightsPin.GetConstTensorPtr(); |
| 4011 | params.m_CellLayerNormWeights = cellLayerNormWeightsPin.GetConstTensorPtr(); |
| 4012 | params.m_OutputLayerNormWeights = outputLayerNormWeightsPin.GetConstTensorPtr(); |
| 4013 | |
| 4014 | // set the layer descriptor |
| 4015 | QLstmDescriptor desc; |
| 4016 | desc.m_CellClip = cellClip; |
| 4017 | desc.m_ProjectionClip = projClip; |
| 4018 | desc.m_CifgEnabled = (params.m_InputToInputWeights == nullptr || |
| 4019 | params.m_RecurrentToInputWeights == nullptr || |
| 4020 | params.m_InputGateBias == nullptr); |
| 4021 | desc.m_PeepholeEnabled = (params.m_CellToForgetWeights != nullptr || |
| 4022 | params.m_CellToOutputWeights != nullptr); |
| 4023 | desc.m_ProjectionEnabled = (params.m_ProjectionWeights != nullptr); |
| 4024 | desc.m_LayerNormEnabled = (params.m_InputLayerNormWeights != nullptr || |
| 4025 | params.m_ForgetLayerNormWeights != nullptr || |
| 4026 | params.m_CellLayerNormWeights != nullptr || |
| 4027 | params.m_OutputLayerNormWeights != nullptr); |
| 4028 | desc.m_InputIntermediateScale = matMulInputGate; |
| 4029 | desc.m_ForgetIntermediateScale = matMulForgetGate; |
| 4030 | desc.m_CellIntermediateScale = matMulCellGate; |
| 4031 | desc.m_OutputIntermediateScale = matMulOutputGate; |
| 4032 | desc.m_HiddenStateScale = projInputScale; |
| 4033 | desc.m_HiddenStateZeroPoint = projInputZeroPoint; |
| 4034 | |
| 4035 | // validate the optional input groups |
| 4036 | if (desc.m_CifgEnabled && |
| 4037 | (params.m_InputToInputWeights != nullptr || |
| 4038 | params.m_RecurrentToInputWeights != nullptr || |
| 4039 | params.m_InputGateBias != nullptr)) |
| 4040 | { |
| 4041 | return Fail("%s: All, or none, of input-to-input weights, recurrent-to-input weights," |
| 4042 | " and input gate bias must be provided", __func__); |
| 4043 | } |
| 4044 | |
| 4045 | if (!desc.m_ProjectionEnabled && params.m_ProjectionBias != nullptr) |
| 4046 | { |
| 4047 | return Fail("%s: projection bias should not be provided without projection weights", __func__); |
| 4048 | } |
| 4049 | |
| 4050 | if (desc.m_PeepholeEnabled && |
| 4051 | (params.m_CellToForgetWeights == nullptr || |
| 4052 | params.m_CellToOutputWeights == nullptr || |
| 4053 | (!desc.m_CifgEnabled && params.m_CellToInputWeights == nullptr))) |
| 4054 | { |
| 4055 | return Fail("%s: All, or none, of cell-to-forget weights and cell-to-output weights must be provided" |
| 4056 | " and, if CIFG is not enabled, cell-to-input weights must also be provided", __func__); |
| 4057 | } |
| 4058 | |
| 4059 | if (desc.m_LayerNormEnabled && |
| 4060 | (params.m_ForgetLayerNormWeights == nullptr || |
| 4061 | params.m_CellLayerNormWeights == nullptr || |
| 4062 | params.m_OutputLayerNormWeights == nullptr || |
| 4063 | (!desc.m_CifgEnabled && params.m_InputLayerNormWeights == nullptr))) |
| 4064 | { |
| 4065 | return Fail("%s: All, or none, of forget-norm weights, cell-norm weights and output-norm weights must be" |
| 4066 | " provided and, if CIFG is not enabled, input-norm weights must also be provided", __func__); |
| 4067 | } |
| 4068 | |
| 4069 | // Basic parameters |
| 4070 | LstmInputParamsInfo paramsInfo; |
| 4071 | paramsInfo.m_InputToForgetWeights = &(params.m_InputToForgetWeights->GetInfo()); |
| 4072 | paramsInfo.m_InputToCellWeights = &(params.m_InputToCellWeights->GetInfo()); |
| 4073 | paramsInfo.m_InputToOutputWeights = &(params.m_InputToOutputWeights->GetInfo()); |
| 4074 | paramsInfo.m_RecurrentToForgetWeights = &(params.m_RecurrentToForgetWeights->GetInfo()); |
| 4075 | paramsInfo.m_RecurrentToCellWeights = &(params.m_RecurrentToCellWeights->GetInfo()); |
| 4076 | paramsInfo.m_RecurrentToOutputWeights = &(params.m_RecurrentToOutputWeights->GetInfo()); |
| 4077 | paramsInfo.m_ForgetGateBias = &(params.m_ForgetGateBias->GetInfo()); |
| 4078 | paramsInfo.m_CellBias = &(params.m_CellBias->GetInfo()); |
| 4079 | paramsInfo.m_OutputGateBias = &(params.m_OutputGateBias->GetInfo()); |
| 4080 | |
| 4081 | // Inputs |
| 4082 | const TensorInfo& inputInfo = input.GetTensorInfo(); |
| 4083 | const TensorInfo& outputStatePrevTimeStepInfo = outputStatePrevTimeStep.GetTensorInfo(); |
| 4084 | const TensorInfo& cellStatePrevTimeStepInfo = cellStatePrevTimeStep.GetTensorInfo(); |
| 4085 | |
| 4086 | // Outputs |
| 4087 | TensorInfo outputStateOutInfo = GetTensorInfoForOperand(*outputStateOut); |
| 4088 | TensorInfo outputInfo = GetTensorInfoForOperand(*output); |
| 4089 | const TensorInfo& cellStateOutInfo = GetTensorInfoForOperand(*cellStateOut); |
| 4090 | |
| 4091 | // Optional parameters |
| 4092 | if (!desc.m_CifgEnabled) |
| 4093 | { |
| 4094 | paramsInfo.m_InputToInputWeights = &(params.m_InputToInputWeights->GetInfo()); |
| 4095 | paramsInfo.m_RecurrentToInputWeights = &(params.m_RecurrentToInputWeights->GetInfo()); |
| 4096 | if (desc.m_PeepholeEnabled) |
| 4097 | { |
| 4098 | paramsInfo.m_CellToInputWeights = &(params.m_CellToInputWeights->GetInfo()); |
| 4099 | } |
| 4100 | paramsInfo.m_InputGateBias = &(params.m_InputGateBias->GetInfo()); |
| 4101 | } |
| 4102 | |
| 4103 | |
| 4104 | if (desc.m_ProjectionEnabled) |
| 4105 | { |
| 4106 | paramsInfo.m_ProjectionWeights = &(params.m_ProjectionWeights->GetInfo()); |
| 4107 | if (params.m_ProjectionBias != nullptr) |
| 4108 | { |
| 4109 | paramsInfo.m_ProjectionBias = &(params.m_ProjectionBias->GetInfo()); |
| 4110 | } |
| 4111 | } |
| 4112 | else |
| 4113 | { |
| 4114 | // If Projection is disabled, override non-const outputs to change the quant info with hidden params, then |
| 4115 | // create a new const TensorInfo based on this |
| 4116 | outputStateOutInfo.SetQuantizationScale(projInputScale); |
| 4117 | outputStateOutInfo.SetQuantizationOffset(projInputZeroPoint); |
| 4118 | outputInfo.SetQuantizationScale(projInputScale); |
| 4119 | outputInfo.SetQuantizationOffset(projInputZeroPoint); |
| 4120 | } |
| 4121 | |
| 4122 | const TensorInfo constOutputStateOutInfo(outputStateOutInfo); |
| 4123 | const TensorInfo constOutputInfo(outputInfo); |
| 4124 | |
| 4125 | if (desc.m_PeepholeEnabled) |
| 4126 | { |
| 4127 | paramsInfo.m_CellToForgetWeights = &(params.m_CellToForgetWeights->GetInfo()); |
| 4128 | paramsInfo.m_CellToOutputWeights = &(params.m_CellToOutputWeights->GetInfo()); |
| 4129 | } |
| 4130 | |
| 4131 | if (desc.m_LayerNormEnabled) |
| 4132 | { |
| 4133 | if(!desc.m_CifgEnabled) |
| 4134 | { |
| 4135 | paramsInfo.m_InputLayerNormWeights = &(params.m_InputLayerNormWeights->GetInfo()); |
| 4136 | } |
| 4137 | paramsInfo.m_ForgetLayerNormWeights = &(params.m_ForgetLayerNormWeights->GetInfo()); |
| 4138 | paramsInfo.m_CellLayerNormWeights = &(params.m_CellLayerNormWeights->GetInfo()); |
| 4139 | paramsInfo.m_OutputLayerNormWeights = &(params.m_OutputLayerNormWeights->GetInfo()); |
| 4140 | } |
| 4141 | |
| 4142 | // Check if the layer is supported |
| 4143 | bool isSupported = false; |
Cathal Corbett | 5383767 | 2022-09-01 11:34:37 +0100 | [diff] [blame] | 4144 | armnn::BackendId setBackend; |
Sadik Armagan | 8f397a1 | 2022-06-17 15:38:22 +0100 | [diff] [blame] | 4145 | auto validateFunc = [&](const armnn::TensorInfo& cellStateOutInfo, bool& isSupported) |
| 4146 | { |
| 4147 | FORWARD_LAYER_SUPPORT_FUNC(__func__, |
| 4148 | IsQLstmSupported, |
| 4149 | data.m_Backends, |
| 4150 | isSupported, |
Cathal Corbett | 5383767 | 2022-09-01 11:34:37 +0100 | [diff] [blame] | 4151 | setBackend, |
Sadik Armagan | 8f397a1 | 2022-06-17 15:38:22 +0100 | [diff] [blame] | 4152 | inputInfo, |
| 4153 | outputStatePrevTimeStepInfo, |
| 4154 | cellStatePrevTimeStepInfo, |
| 4155 | constOutputStateOutInfo, |
| 4156 | cellStateOutInfo, |
| 4157 | constOutputInfo, |
| 4158 | desc, |
| 4159 | paramsInfo); |
| 4160 | }; |
| 4161 | |
| 4162 | bool isDynamic = false; |
| 4163 | if (!IsDynamicTensor(constOutputStateOutInfo) && |
| 4164 | !IsDynamicTensor(cellStateOutInfo) && |
| 4165 | !IsDynamicTensor(constOutputInfo)) |
| 4166 | { |
| 4167 | validateFunc(outputInfo, isSupported); |
| 4168 | } |
| 4169 | else |
| 4170 | { |
| 4171 | isDynamic = true; |
| 4172 | isSupported = AreDynamicTensorsSupported(); |
| 4173 | } |
| 4174 | |
| 4175 | if (!isSupported) |
| 4176 | { |
| 4177 | return false; |
| 4178 | } |
| 4179 | |
| 4180 | // Add the layer |
| 4181 | IConnectableLayer* layer = data.m_Network->AddQLstmLayer(desc, params, "QLstm"); |
Cathal Corbett | 5383767 | 2022-09-01 11:34:37 +0100 | [diff] [blame] | 4182 | layer->SetBackendId(setBackend); |
Sadik Armagan | 8f397a1 | 2022-06-17 15:38:22 +0100 | [diff] [blame] | 4183 | |
| 4184 | input.Connect(layer->GetInputSlot(0)); |
| 4185 | outputStatePrevTimeStep.Connect(layer->GetInputSlot(1)); |
| 4186 | cellStatePrevTimeStep.Connect(layer->GetInputSlot(2)); |
| 4187 | |
| 4188 | if (!isDynamic) |
| 4189 | { |
| 4190 | return ( SetupAndTrackLayerOutputSlot( |
| 4191 | operation, 0, *layer, 0, model, data, &constOutputStateOutInfo) && |
| 4192 | SetupAndTrackLayerOutputSlot(operation, 1, *layer, 1, model, data) && |
| 4193 | SetupAndTrackLayerOutputSlot(operation, 2, *layer, 2, model, data, &constOutputInfo)); |
| 4194 | } |
| 4195 | else |
| 4196 | { |
| 4197 | return ( SetupAndTrackLayerOutputSlot( |
| 4198 | operation, 0, *layer, 0, model, data, &constOutputStateOutInfo) && |
| 4199 | SetupAndTrackLayerOutputSlot( |
| 4200 | operation, 1, *layer, 1, model, data, nullptr, validateFunc, |
| 4201 | ActivationFn::kActivationNone, true) && |
| 4202 | SetupAndTrackLayerOutputSlot(operation, 2, *layer, 2, model, data, &constOutputInfo)); |
| 4203 | } |
| 4204 | } |
| 4205 | |
| 4206 | bool Converter::ConvertQuantized16BitLstm(const Operation& operation, const Model& model, ConversionData& data) |
| 4207 | { |
| 4208 | VLOG(DRIVER) << "Converter::ConvertQuantized16BitLstm()"; |
| 4209 | VLOG(DRIVER) << "Policy::ConvertQuantized16BitLstm()"; |
| 4210 | |
| 4211 | //Inputs: |
| 4212 | // 0: The input: A 2-D tensor of type ANEURALNETWORKS_TENSOR_QUANT8_ASYMM and shape [numBatches, inputSize] |
| 4213 | // specifying the input to the LSTM cell. Tensor is quantized with a fixed quantization range of -1, 127/128. |
| 4214 | LayerInputHandle input = ConvertToLayerInputHandle(operation, 0, model, data); |
| 4215 | if (!input.IsValid()) |
| 4216 | { |
| 4217 | return Fail("%s: Could not read input 0: input", __func__); |
| 4218 | } |
| 4219 | |
| 4220 | //13: The previous cell state: A 2-D tensor of type ANEURALNETWORKS_TENSOR_QUANT16_SYMM and shape |
| 4221 | // [numBatches, outputSize] specifying the cell state from the previous time step of the LSTM cell. |
| 4222 | // It is quantized using a quantization range of -2^4, 2^4 * 32767/32768. |
| 4223 | LayerInputHandle previousCellStateIn = ConvertToLayerInputHandle(operation, 13, model, data); |
| 4224 | if (!previousCellStateIn.IsValid()) |
| 4225 | { |
| 4226 | return Fail("%s: Could not read input 13: previousCellStateIn", __func__); |
| 4227 | } |
| 4228 | |
| 4229 | // 14: The previous output state: A 2-D tensor of type ANEURALNETWORKS_TENSOR_QUANT8_ASYMM and shape |
| 4230 | // [numBathes, outputSize] specifying the output of the LSTM cell from previous time-step. Tensor |
| 4231 | // is quantized with a fixed quantization range of -1, 127/128. |
| 4232 | LayerInputHandle previousOutputIn = ConvertToLayerInputHandle(operation, 14, model, data); |
| 4233 | if (!previousOutputIn.IsValid()) |
| 4234 | { |
| 4235 | return Fail("%s: Could not read input 14: previousOutputIn", __func__); |
| 4236 | } |
| 4237 | |
| 4238 | // Get the input tensors: |
| 4239 | // 1: The input-to-input weights. A 2-D tensor of type ANEURALNETWORKS_TENSOR_QUANT8_ASYMM and shape |
| 4240 | // [outputSize, inputSize] specifying input-to-input part of weights for fully-connected layer inside the |
| 4241 | // LSTM cell. Quantization zero point and scale must be the same across all the weights. |
| 4242 | const ConstTensorPin inputToInputWeightsPin = |
| 4243 | ConvertOperationInputToConstTensorPin(operation, 1, model, data); |
| 4244 | |
| 4245 | // 2: The input-to-forget weights. A 2-D tensor of type ANEURALNETWORKS_TENSOR_QUANT8_ASYMM and shape |
| 4246 | // [outputSize, inputSize] specifying input-to-forget part of weights for fully-connected layer inside the |
| 4247 | // LSTM cell. Quantization zero point and scale must be the same across all the weights. |
| 4248 | const ConstTensorPin inputToForgetWeightsPin = |
| 4249 | ConvertOperationInputToConstTensorPin(operation, 2, model, data); |
| 4250 | |
| 4251 | // 3: The input-to-cell weights. A 2-D tensor of type ANEURALNETWORKS_TENSOR_QUANT8_ASYMM and shape |
| 4252 | // [outputSize, inputSize] specifying input-to-cell part of weights for fully-connected layer inside the |
| 4253 | // LSTM cell. Quantization zero point and scale must be the same across all the weights. |
| 4254 | const ConstTensorPin inputToCellWeightsPin = |
| 4255 | ConvertOperationInputToConstTensorPin(operation, 3, model, data); |
| 4256 | |
| 4257 | // 4: The input-to-output weights. A 2-D tensor of type ANEURALNETWORKS_TENSOR_QUANT8_ASYMM and shape |
| 4258 | // [outputSize, inputSize] specifying input-to-output part of weights for fully-connected layer inside the |
| 4259 | // LSTM cell. Quantization zero point and scale must be the same across all the weights. |
| 4260 | const ConstTensorPin inputToOutputWeightsPin = |
| 4261 | ConvertOperationInputToConstTensorPin(operation, 4, model, data); |
| 4262 | |
| 4263 | // 5: The recurrent-to-input weights. A 2-D tensor of type ANEURALNETWORKS_TENSOR_QUANT8_ASYMM and shape |
| 4264 | // [outputSize, outputSize] specifying recurrent-to-input part of weights for fully-connected layer inside |
| 4265 | // the LSTM cell. Quantization zero point and scale must be the same across all the weights. |
| 4266 | const ConstTensorPin recurrentToInputWeightsPin = |
| 4267 | ConvertOperationInputToConstTensorPin(operation, 5, model, data); |
| 4268 | |
| 4269 | // 6: The recurrent-to-forget weights. A 2-D tensor of type ANEURALNETWORKS_TENSOR_QUANT8_ASYMM and shape |
| 4270 | // [outputSize, outputSize] specifying recurrent-to-forget part of weights for fully-connected layer inside |
| 4271 | // the LSTM cell. Quantization zero point and scale must be the same across all the weights. |
| 4272 | const ConstTensorPin recurrentToForgetWeightsPin = |
| 4273 | ConvertOperationInputToConstTensorPin(operation, 6, model, data); |
| 4274 | |
| 4275 | // 7: The recurrent-to-cell weights. A 2-D tensor of type ANEURALNETWORKS_TENSOR_QUANT8_ASYMM and shape |
| 4276 | // [outputSize, outputSize] specifying recurrent-to-cell part of weights for fully-connected layer inside |
| 4277 | // the LSTM cell. Quantization zero point and scale must be the same across all the weights. |
| 4278 | const ConstTensorPin recurrentToCellWeightsPin = |
| 4279 | ConvertOperationInputToConstTensorPin(operation, 7, model, data); |
| 4280 | |
| 4281 | // 8: The recurrent-to-output weights. A 2-D tensor of type ANEURALNETWORKS_TENSOR_QUANT8_ASYMM and shape |
| 4282 | // [outputSize, outputSize] specifying recurrent-to-output part of weights for fully-connected layer inside |
| 4283 | // the LSTM cell. Quantization zero point and scale must be the same across all the weights. |
| 4284 | const ConstTensorPin recurrentToOutputWeightsPin = |
| 4285 | ConvertOperationInputToConstTensorPin(operation, 8, model, data); |
| 4286 | |
| 4287 | // 9: The input gate bias. A 1-D tensor of type ANEURALNETWORKS_TENSOR_INT32 and shape [outputSize] specifying the |
| 4288 | // bias for the fully-connected layer inside the LSTM cell. Bias is quantized with scale being a product |
| 4289 | // of input and weights scales and zeroPoint equal to 0. |
| 4290 | const ConstTensorPin inputGateBiasPin = |
| 4291 | ConvertOperationInputToConstTensorPin(operation, 9, model, data); |
| 4292 | |
| 4293 | // 10: The forget gate bias. A 1-D tensor of type ANEURALNETWORKS_TENSOR_INT32 and shape [outputSize] specifying |
| 4294 | // the bias for the fully-connected layer inside the LSTM cell. Bias is quantized with scale being a product |
| 4295 | // of input and weights scales and zeroPoint equal to 0. |
| 4296 | const ConstTensorPin forgetGateBiasPin = |
| 4297 | ConvertOperationInputToConstTensorPin(operation, 10, model, data); |
| 4298 | |
| 4299 | // 11:The cell bias. A 1-D tensor of type ANEURALNETWORKS_TENSOR_INT32 and shape [outputSize] specifying the bias |
| 4300 | // for the fully-connected layer inside the LSTM cell. Bias is quantized with scale being a product of input |
| 4301 | // and weights scales and zeroPoint equal to 0. |
| 4302 | const ConstTensorPin cellBiasPin = |
| 4303 | ConvertOperationInputToConstTensorPin(operation, 11, model, data); |
| 4304 | |
| 4305 | // 12:The output gate bias. A 1-D tensor of type ANEURALNETWORKS_TENSOR_INT32 and shape [outputSize] specifying |
| 4306 | // the bias for the fully-connected layer inside the LSTM cell. Bias is quantized with scale being a product |
| 4307 | // of input and weights scales and zeroPoint equal to 0. |
| 4308 | const ConstTensorPin outputGateBiasPin = |
| 4309 | ConvertOperationInputToConstTensorPin(operation, 12, model, data); |
| 4310 | |
| 4311 | if (!inputToInputWeightsPin.IsValid() || |
| 4312 | !inputToForgetWeightsPin.IsValid() || |
| 4313 | !inputToCellWeightsPin.IsValid() || |
| 4314 | !inputToOutputWeightsPin.IsValid() || |
| 4315 | !recurrentToInputWeightsPin.IsValid() || |
| 4316 | !recurrentToForgetWeightsPin.IsValid() || |
| 4317 | !recurrentToCellWeightsPin.IsValid() || |
| 4318 | !recurrentToOutputWeightsPin.IsValid() || |
| 4319 | !inputGateBiasPin.IsValid() || |
| 4320 | !forgetGateBiasPin.IsValid() || |
| 4321 | !cellBiasPin.IsValid() || |
| 4322 | !outputGateBiasPin.IsValid()) |
| 4323 | { |
| 4324 | return Fail("%s: Operation has invalid tensor inputs", __func__); |
| 4325 | } |
| 4326 | |
| 4327 | // Outputs: |
| 4328 | // 0: The cell state: A 2-D tensor of type ANEURALNETWORKS_TENSOR_QUANT16_SYMM and shape [numBatches, outputSize] |
| 4329 | // which contains a cell state from the current time step. Tensor is quantized using a quantization range |
| 4330 | // of -2^4, 2^4 * 32767/32768. |
| 4331 | const Operand* cellStateOut = GetOutputOperand(operation, 0, model); |
| 4332 | if (!cellStateOut) |
| 4333 | { |
| 4334 | return Fail("%s: Could not read output 0: cellStateOut", __func__); |
| 4335 | } |
| 4336 | |
| 4337 | // 1: The output: A 2-D tensor of type ANEURALNETWORKS_TENSOR_QUANT8_ASYMM and shape [numBathes, outputSize] which |
| 4338 | // contains the output value. Tensor is quantized with a fixed quantization range of -1, 127/128. |
| 4339 | const Operand* output = GetOutputOperand(operation, 1, model); |
| 4340 | if (!output) |
| 4341 | { |
| 4342 | return Fail("%s: Could not read output 1: output", __func__); |
| 4343 | } |
| 4344 | |
| 4345 | // Inputs |
| 4346 | const TensorInfo& inputInfo = input.GetTensorInfo(); |
| 4347 | const TensorInfo& previousCellStateInInfo = previousCellStateIn.GetTensorInfo(); |
| 4348 | const TensorInfo& previousOutputInInfo = previousOutputIn.GetTensorInfo(); |
| 4349 | |
| 4350 | // Outputs |
| 4351 | const TensorInfo& cellStateOutInfo = GetTensorInfoForOperand(*cellStateOut); |
| 4352 | const TensorInfo& outputInfo = GetTensorInfoForOperand(*output); |
| 4353 | |
| 4354 | // Dynamic tensors currently not supported |
| 4355 | if (IsDynamicTensor(cellStateOutInfo) || IsDynamicTensor(outputInfo)) |
| 4356 | { |
| 4357 | return Fail("%s: Dynamic output tensors are not supported", __func__); |
| 4358 | } |
| 4359 | |
| 4360 | QuantizedLstmInputParams params; |
| 4361 | |
| 4362 | params.m_InputToInputWeights = inputToInputWeightsPin.GetConstTensorPtr(); |
| 4363 | params.m_InputToForgetWeights = inputToForgetWeightsPin.GetConstTensorPtr(); |
| 4364 | params.m_InputToCellWeights = inputToCellWeightsPin.GetConstTensorPtr(); |
| 4365 | params.m_InputToOutputWeights = inputToOutputWeightsPin.GetConstTensorPtr(); |
| 4366 | params.m_RecurrentToInputWeights = recurrentToInputWeightsPin.GetConstTensorPtr(); |
| 4367 | params.m_RecurrentToForgetWeights = recurrentToForgetWeightsPin.GetConstTensorPtr(); |
| 4368 | params.m_RecurrentToCellWeights = recurrentToCellWeightsPin.GetConstTensorPtr(); |
| 4369 | params.m_RecurrentToOutputWeights = recurrentToOutputWeightsPin.GetConstTensorPtr(); |
| 4370 | params.m_InputGateBias = inputGateBiasPin.GetConstTensorPtr(); |
| 4371 | params.m_ForgetGateBias = forgetGateBiasPin.GetConstTensorPtr(); |
| 4372 | params.m_CellBias = cellBiasPin.GetConstTensorPtr(); |
| 4373 | params.m_OutputGateBias = outputGateBiasPin.GetConstTensorPtr(); |
| 4374 | |
| 4375 | QuantizedLstmInputParamsInfo paramsInfo; |
| 4376 | paramsInfo.m_InputToInputWeights = &(params.m_InputToInputWeights->GetInfo()); |
| 4377 | paramsInfo.m_InputToForgetWeights = &(params.m_InputToForgetWeights->GetInfo()); |
| 4378 | paramsInfo.m_InputToCellWeights = &(params.m_InputToCellWeights->GetInfo()); |
| 4379 | paramsInfo.m_InputToOutputWeights = &(params.m_InputToOutputWeights->GetInfo()); |
| 4380 | paramsInfo.m_RecurrentToInputWeights = &(params.m_RecurrentToInputWeights->GetInfo()); |
| 4381 | paramsInfo.m_RecurrentToForgetWeights = &(params.m_RecurrentToForgetWeights->GetInfo()); |
| 4382 | paramsInfo.m_RecurrentToCellWeights = &(params.m_RecurrentToCellWeights->GetInfo()); |
| 4383 | paramsInfo.m_RecurrentToOutputWeights = &(params.m_RecurrentToOutputWeights->GetInfo()); |
| 4384 | paramsInfo.m_InputGateBias = &(params.m_InputGateBias->GetInfo()); |
| 4385 | paramsInfo.m_ForgetGateBias = &(params.m_ForgetGateBias->GetInfo()); |
| 4386 | paramsInfo.m_CellBias = &(params.m_CellBias->GetInfo()); |
| 4387 | paramsInfo.m_OutputGateBias = &(params.m_OutputGateBias->GetInfo()); |
| 4388 | |
| 4389 | bool isSupported = false; |
Cathal Corbett | 5383767 | 2022-09-01 11:34:37 +0100 | [diff] [blame] | 4390 | armnn::BackendId setBackend; |
Sadik Armagan | 8f397a1 | 2022-06-17 15:38:22 +0100 | [diff] [blame] | 4391 | auto validateFunc = [&](const armnn::TensorInfo& outputInfo, bool& isSupported) |
| 4392 | { |
| 4393 | FORWARD_LAYER_SUPPORT_FUNC(__func__, |
| 4394 | IsQuantizedLstmSupported, |
| 4395 | data.m_Backends, |
| 4396 | isSupported, |
Cathal Corbett | 5383767 | 2022-09-01 11:34:37 +0100 | [diff] [blame] | 4397 | setBackend, |
Sadik Armagan | 8f397a1 | 2022-06-17 15:38:22 +0100 | [diff] [blame] | 4398 | inputInfo, |
| 4399 | previousCellStateInInfo, |
| 4400 | previousOutputInInfo, |
| 4401 | cellStateOutInfo, |
| 4402 | outputInfo, |
| 4403 | paramsInfo); |
| 4404 | }; |
| 4405 | |
| 4406 | bool isDynamic = false; |
| 4407 | if (!IsDynamicTensor(cellStateOutInfo) && |
| 4408 | !IsDynamicTensor(outputInfo)) |
| 4409 | { |
| 4410 | validateFunc(outputInfo, isSupported); |
| 4411 | } |
| 4412 | else |
| 4413 | { |
| 4414 | isDynamic = true; |
| 4415 | isSupported = AreDynamicTensorsSupported(); |
| 4416 | } |
| 4417 | |
| 4418 | if (!isSupported) |
| 4419 | { |
| 4420 | return false; |
| 4421 | } |
| 4422 | |
| 4423 | IConnectableLayer* const layer = data.m_Network->AddQuantizedLstmLayer(params, "QuantizedLstm"); |
Cathal Corbett | 5383767 | 2022-09-01 11:34:37 +0100 | [diff] [blame] | 4424 | layer->SetBackendId(setBackend); |
Sadik Armagan | 8f397a1 | 2022-06-17 15:38:22 +0100 | [diff] [blame] | 4425 | input.Connect(layer->GetInputSlot(0)); |
| 4426 | previousCellStateIn.Connect(layer->GetInputSlot(1)); |
| 4427 | previousOutputIn.Connect(layer->GetInputSlot(2)); |
| 4428 | |
| 4429 | if (!isDynamic) |
| 4430 | { |
| 4431 | return (SetupAndTrackLayerOutputSlot(operation, 0, *layer, 0, model, data) && |
| 4432 | SetupAndTrackLayerOutputSlot(operation, 1, *layer, 1, model, data)); |
| 4433 | } |
| 4434 | else |
| 4435 | { |
| 4436 | return (SetupAndTrackLayerOutputSlot(operation, 0, *layer, 0, model, data) && |
| 4437 | SetupAndTrackLayerOutputSlot( |
| 4438 | operation, 1, *layer, 1, model, data, nullptr, validateFunc, ActivationFn::kActivationNone, true)); |
| 4439 | } |
| 4440 | |
| 4441 | } |
| 4442 | |
| 4443 | bool Converter::ConvertRank(const Operation& operation, const Model& model, ConversionData& data) |
| 4444 | { |
| 4445 | VLOG(DRIVER) << "Converter::ConvertRank()"; |
| 4446 | |
| 4447 | const Operand* inputOperand = GetInputOperand(operation, 0, model); |
| 4448 | const Operand* outputOperand = GetOutputOperand(operation, 0, model); |
| 4449 | |
| 4450 | if (inputOperand == nullptr || outputOperand == nullptr) |
| 4451 | { |
| 4452 | return Fail("%s: Operation has invalid inputs", __func__); |
| 4453 | } |
| 4454 | |
| 4455 | const Shape inputOperandShape = GetOperandShape(*inputOperand); |
| 4456 | const Shape outputOperandShape = GetOperandShape(*outputOperand); |
| 4457 | |
| 4458 | LayerInputHandle input = ConvertToLayerInputHandle(operation, 0, model, data); |
| 4459 | if (!input.IsValid()) |
| 4460 | { |
| 4461 | return Fail("%s: Could not read input 0", __func__); |
| 4462 | } |
| 4463 | |
| 4464 | armnn::TensorInfo outInfo = GetTensorInfoForOperand(*outputOperand); |
| 4465 | if (IsDynamicTensor(outInfo)) |
| 4466 | { |
| 4467 | return Fail("%s: Dynamic output tensors are not supported", __func__); |
| 4468 | } |
| 4469 | |
| 4470 | bool isSupported = false; |
Cathal Corbett | 5383767 | 2022-09-01 11:34:37 +0100 | [diff] [blame] | 4471 | armnn::BackendId setBackend; |
Sadik Armagan | 8f397a1 | 2022-06-17 15:38:22 +0100 | [diff] [blame] | 4472 | FORWARD_LAYER_SUPPORT_FUNC(__func__, |
| 4473 | IsRankSupported, |
| 4474 | data.m_Backends, |
| 4475 | isSupported, |
Cathal Corbett | 5383767 | 2022-09-01 11:34:37 +0100 | [diff] [blame] | 4476 | setBackend, |
Sadik Armagan | 8f397a1 | 2022-06-17 15:38:22 +0100 | [diff] [blame] | 4477 | input.GetTensorInfo(), |
| 4478 | outInfo); |
| 4479 | if (!isSupported) |
| 4480 | { |
| 4481 | return false; |
| 4482 | } |
| 4483 | |
| 4484 | armnn::IConnectableLayer* layer = data.m_Network->AddRankLayer(); |
Cathal Corbett | 5383767 | 2022-09-01 11:34:37 +0100 | [diff] [blame] | 4485 | layer->SetBackendId(setBackend); |
Sadik Armagan | 8f397a1 | 2022-06-17 15:38:22 +0100 | [diff] [blame] | 4486 | assert(layer != nullptr); |
| 4487 | input.Connect(layer->GetInputSlot(0)); |
| 4488 | |
| 4489 | return SetupAndTrackLayerOutputSlot(operation, 0, *layer, model, data, &outInfo); |
| 4490 | } |
| 4491 | |
| 4492 | bool Converter::ConvertReLu(const Operation& operation, const Model& model, ConversionData& data) |
| 4493 | { |
| 4494 | VLOG(DRIVER) << "Converter::ConvertReLu()"; |
| 4495 | armnn::ActivationDescriptor desc; |
| 4496 | desc.m_Function = armnn::ActivationFunction::ReLu; |
| 4497 | |
| 4498 | |
| 4499 | LayerInputHandle input = ConvertToLayerInputHandle(operation, 0, model, data); |
| 4500 | if (!input.IsValid()) |
| 4501 | { |
Kevin May | 21e3999 | 2023-07-31 12:20:03 +0100 | [diff] [blame] | 4502 | return Fail("%s: Input 0 is invalid", "operationName", __func__); |
Sadik Armagan | 8f397a1 | 2022-06-17 15:38:22 +0100 | [diff] [blame] | 4503 | } |
| 4504 | |
| 4505 | const Operand* outputOperand = GetOutputOperand(operation, 0, model); |
| 4506 | if (!outputOperand) |
| 4507 | { |
| 4508 | return false; |
| 4509 | } |
| 4510 | |
| 4511 | const armnn::TensorInfo& outInfo = GetTensorInfoForOperand(*outputOperand); |
| 4512 | |
| 4513 | bool isSupported = false; |
Cathal Corbett | 5383767 | 2022-09-01 11:34:37 +0100 | [diff] [blame] | 4514 | armnn::BackendId setBackend; |
Sadik Armagan | 8f397a1 | 2022-06-17 15:38:22 +0100 | [diff] [blame] | 4515 | auto validateFunc = [&](const armnn::TensorInfo& outInfo, bool& isSupported) |
| 4516 | { |
| 4517 | FORWARD_LAYER_SUPPORT_FUNC(__func__, |
| 4518 | IsActivationSupported, |
| 4519 | data.m_Backends, |
| 4520 | isSupported, |
Cathal Corbett | 5383767 | 2022-09-01 11:34:37 +0100 | [diff] [blame] | 4521 | setBackend, |
Sadik Armagan | 8f397a1 | 2022-06-17 15:38:22 +0100 | [diff] [blame] | 4522 | input.GetTensorInfo(), |
| 4523 | outInfo, |
| 4524 | desc); |
| 4525 | }; |
| 4526 | |
| 4527 | if(IsDynamicTensor(outInfo)) |
| 4528 | { |
| 4529 | isSupported = AreDynamicTensorsSupported(); |
| 4530 | } |
| 4531 | else |
| 4532 | { |
| 4533 | validateFunc(outInfo, isSupported); |
| 4534 | } |
| 4535 | |
| 4536 | if (!isSupported) |
| 4537 | { |
| 4538 | return false; |
| 4539 | } |
| 4540 | |
| 4541 | armnn::IConnectableLayer* layer = data.m_Network->AddActivationLayer(desc); |
Jim Flynn | 987398f | 2023-09-16 18:53:52 +0100 | [diff] [blame] | 4542 | if (layer == nullptr) |
| 4543 | { |
| 4544 | throw armnn::NullPointerException("failed to add Activation Layer to network"); |
| 4545 | } |
Cathal Corbett | 5383767 | 2022-09-01 11:34:37 +0100 | [diff] [blame] | 4546 | layer->SetBackendId(setBackend); |
Sadik Armagan | 8f397a1 | 2022-06-17 15:38:22 +0100 | [diff] [blame] | 4547 | input.Connect(layer->GetInputSlot(0)); |
| 4548 | |
| 4549 | return SetupAndTrackLayerOutputSlot(operation, 0, *layer, model, data, nullptr, validateFunc); |
| 4550 | } |
| 4551 | |
| 4552 | bool Converter::ConvertReLu1(const Operation& operation, const Model& model, ConversionData& data) |
| 4553 | { |
| 4554 | VLOG(DRIVER) << "Converter::ConvertReLu1()"; |
| 4555 | armnn::ActivationDescriptor desc; |
| 4556 | desc.m_Function = armnn::ActivationFunction::BoundedReLu; |
| 4557 | desc.m_A = 1.0f; |
| 4558 | desc.m_B = -1.0f; |
| 4559 | |
| 4560 | return ConvertToActivation(operation, __func__, desc, model, data); |
| 4561 | } |
| 4562 | |
| 4563 | bool Converter::ConvertReLu6(const Operation& operation, const Model& model, ConversionData& data) |
| 4564 | { |
| 4565 | VLOG(DRIVER) << "Converter::ConvertReLu6()"; |
| 4566 | armnn::ActivationDescriptor desc; |
| 4567 | desc.m_Function = armnn::ActivationFunction::BoundedReLu; |
| 4568 | desc.m_A = 6.0f; |
| 4569 | |
| 4570 | return ConvertToActivation(operation, __func__, desc, model, data); |
| 4571 | } |
| 4572 | |
| 4573 | bool Converter::ConvertReshape(const Operation& operation, const Model& model, ConversionData& data) |
| 4574 | { |
| 4575 | VLOG(DRIVER) << "Converter::ConvertReshape()"; |
| 4576 | |
| 4577 | const Operand* inputOperand = GetInputOperand(operation, 0, model); |
| 4578 | const Operand* requestedShapeOperand = GetInputOperand(operation, 1, model); |
| 4579 | const Operand* outputOperand = GetOutputOperand(operation, 0, model); |
| 4580 | |
| 4581 | if (inputOperand == nullptr |
| 4582 | || requestedShapeOperand == nullptr |
| 4583 | || outputOperand == nullptr) |
| 4584 | { |
| 4585 | return Fail("%s: Operation has invalid inputs", __func__); |
| 4586 | } |
| 4587 | |
| 4588 | if (requestedShapeOperand->dimensions.size() != 1) |
| 4589 | { |
| 4590 | return Fail("%s: Input 1 expected to be one-dimensional (found %i dimensions)", |
| 4591 | __func__, requestedShapeOperand->dimensions.size()); |
| 4592 | } |
| 4593 | |
| 4594 | std::vector<int32_t> targetDimensions; |
| 4595 | if (!GetTensorInt32Values(*requestedShapeOperand, targetDimensions, model, data)) |
| 4596 | { |
| 4597 | return Fail("%s: Could not read values of input 1", __func__); |
| 4598 | } |
| 4599 | |
| 4600 | const Shape inputOperandShape = GetOperandShape(*inputOperand); |
| 4601 | |
| 4602 | Shape requestedShape; |
| 4603 | // targetDimensions may contain special values (e.g. -1). reshapePrepare() is an AndroidNN provided utility |
| 4604 | // function that resolves these values into a fully specified tensor shape. |
| 4605 | if (!reshapePrepare(inputOperandShape, targetDimensions.data(), targetDimensions.size(), &requestedShape)) |
| 4606 | { |
| 4607 | return Fail("%s: Failed to resolve the requested shape", __func__); |
| 4608 | } |
| 4609 | |
| 4610 | LayerInputHandle input = ConvertToLayerInputHandle(operation, 0, model, data); |
| 4611 | if (!input.IsValid()) |
| 4612 | { |
| 4613 | return Fail("%s: Could not read input 0", __func__); |
| 4614 | } |
| 4615 | |
| 4616 | armnn::ReshapeDescriptor reshapeDescriptor; |
| 4617 | reshapeDescriptor.m_TargetShape = armnn::TensorShape(requestedShape.dimensions.size(), |
| 4618 | requestedShape.dimensions.data()); |
| 4619 | |
| 4620 | const armnn::TensorInfo& outputInfo = GetTensorInfoForOperand(*outputOperand); |
| 4621 | |
| 4622 | bool isSupported = false; |
Cathal Corbett | 5383767 | 2022-09-01 11:34:37 +0100 | [diff] [blame] | 4623 | armnn::BackendId setBackend; |
Sadik Armagan | 8f397a1 | 2022-06-17 15:38:22 +0100 | [diff] [blame] | 4624 | auto validateFunc = [&](const armnn::TensorInfo& outputInfo, bool& isSupported) |
| 4625 | { |
| 4626 | FORWARD_LAYER_SUPPORT_FUNC(__func__, |
| 4627 | IsReshapeSupported, |
| 4628 | data.m_Backends, |
| 4629 | isSupported, |
Cathal Corbett | 5383767 | 2022-09-01 11:34:37 +0100 | [diff] [blame] | 4630 | setBackend, |
Sadik Armagan | 8f397a1 | 2022-06-17 15:38:22 +0100 | [diff] [blame] | 4631 | input.GetTensorInfo(), |
| 4632 | outputInfo, |
| 4633 | reshapeDescriptor); |
| 4634 | }; |
| 4635 | |
| 4636 | if(!IsDynamicTensor(outputInfo)) |
| 4637 | { |
| 4638 | validateFunc(outputInfo, isSupported); |
| 4639 | } |
| 4640 | else |
| 4641 | { |
| 4642 | isSupported = AreDynamicTensorsSupported(); |
| 4643 | } |
| 4644 | |
| 4645 | if (!isSupported) |
| 4646 | { |
| 4647 | return false; |
| 4648 | } |
| 4649 | |
| 4650 | armnn::IConnectableLayer* layer = data.m_Network->AddReshapeLayer(reshapeDescriptor); |
Cathal Corbett | 5383767 | 2022-09-01 11:34:37 +0100 | [diff] [blame] | 4651 | layer->SetBackendId(setBackend); |
Sadik Armagan | 8f397a1 | 2022-06-17 15:38:22 +0100 | [diff] [blame] | 4652 | assert(layer != nullptr); |
| 4653 | input.Connect(layer->GetInputSlot(0)); |
| 4654 | |
| 4655 | return SetupAndTrackLayerOutputSlot(operation, 0, *layer, model, data, nullptr, validateFunc); |
| 4656 | } |
| 4657 | |
| 4658 | bool Converter::ConvertResize(const Operation& operation, |
| 4659 | const Model& model, |
| 4660 | ConversionData& data, |
| 4661 | ResizeMethod resizeMethod) |
| 4662 | { |
| 4663 | VLOG(DRIVER) << "Converter::ConvertResize()"; |
| 4664 | VLOG(DRIVER) << "resizeMethod = " << GetResizeMethodAsCString(resizeMethod); |
| 4665 | |
| 4666 | LayerInputHandle input = ConvertToLayerInputHandle(operation, 0, model, data); |
| 4667 | if (!input.IsValid()) |
| 4668 | { |
| 4669 | return Fail("%s: Could not read input 0", __func__); |
| 4670 | } |
| 4671 | |
| 4672 | const Operand* output = GetOutputOperand(operation, 0, model); |
| 4673 | if (!output) |
| 4674 | { |
| 4675 | return Fail("%s: Could not read output 0", __func__); |
| 4676 | } |
| 4677 | |
| 4678 | const TensorInfo& inputInfo = input.GetTensorInfo(); |
| 4679 | const TensorInfo& outputInfo = GetTensorInfoForOperand(*output); |
| 4680 | |
| 4681 | ResizeDescriptor descriptor; |
| 4682 | descriptor.m_Method = resizeMethod; |
| 4683 | descriptor.m_DataLayout = OptionalDataLayout(operation, 3, model, data); |
| 4684 | |
| 4685 | OperandType operandType1; |
| 4686 | OperandType operandType2; |
| 4687 | |
| 4688 | if (!GetOperandType(operation, 1, model, operandType1) || |
| 4689 | !GetOperandType(operation, 2, model, operandType2)) |
| 4690 | { |
| 4691 | return Fail("%s: Operation has invalid inputs", __func__); |
| 4692 | } |
| 4693 | |
| 4694 | if (operandType1 != operandType2) |
| 4695 | { |
| 4696 | return Fail("%s: Operation has invalid inputs. Type of input 1 and 2 should be the same", __func__); |
| 4697 | } |
| 4698 | |
| 4699 | if (operandType1 == OperandType::INT32) |
| 4700 | { |
| 4701 | // Case 1: resizing by shape |
| 4702 | int32_t targetWidth = 0; |
| 4703 | int32_t targetHeight = 0; |
| 4704 | |
| 4705 | if (!GetInputInt32(operation, 1, targetWidth, model, data) || |
| 4706 | !GetInputInt32(operation, 2, targetHeight, model, data)) |
| 4707 | { |
| 4708 | return Fail("%s: Operation has invalid inputs for resizing by shape", __func__); |
| 4709 | } |
| 4710 | |
| 4711 | if (targetWidth < 0 || targetHeight < 0) |
| 4712 | { |
| 4713 | return Fail("%s: Operation has invalid inputs for resizing by shape. " |
| 4714 | "Target width/height cannot be < 0", __func__); |
| 4715 | } |
| 4716 | |
| 4717 | descriptor.m_TargetWidth = static_cast<uint32_t>(targetWidth); |
| 4718 | descriptor.m_TargetHeight = static_cast<uint32_t>(targetHeight); |
| 4719 | } |
| 4720 | else if (operandType1 == OperandType::FLOAT32) |
| 4721 | { |
| 4722 | // Case 2: resizing by scale |
| 4723 | float widthScale = 1.0f; |
| 4724 | float heightScale = 1.0f; |
| 4725 | |
| 4726 | if (!GetInputFloat32(operation, 1, widthScale, model, data) || |
| 4727 | !GetInputFloat32(operation, 2, heightScale, model, data)) |
| 4728 | { |
| 4729 | return Fail("%s: Operation has invalid inputs for resizing by scale", __func__); |
| 4730 | } |
| 4731 | |
| 4732 | const TensorShape& inputShape = inputInfo.GetShape(); |
| 4733 | armnnUtils::DataLayoutIndexed dataLayoutIndexed(descriptor.m_DataLayout); |
| 4734 | |
| 4735 | float width = inputShape[dataLayoutIndexed.GetWidthIndex()]; |
| 4736 | float height = inputShape[dataLayoutIndexed.GetHeightIndex()]; |
| 4737 | |
| 4738 | descriptor.m_TargetWidth = std::floor(width * widthScale); |
| 4739 | descriptor.m_TargetHeight = std::floor(height * heightScale); |
| 4740 | } |
| 4741 | else if (operandType1 == OperandType::FLOAT16) |
| 4742 | { |
| 4743 | Half widthScale; |
| 4744 | Half heightScale; |
| 4745 | |
| 4746 | if (!GetInputScalar(operation, 1, OperandType::FLOAT16, widthScale, model, data) || |
| 4747 | !GetInputScalar(operation, 2, OperandType::FLOAT16, heightScale, model, data)) |
| 4748 | { |
| 4749 | return Fail("%s: Operation has invalid inputs for resizing by scale", __func__); |
| 4750 | } |
| 4751 | |
| 4752 | const TensorShape& inputShape = inputInfo.GetShape(); |
| 4753 | armnnUtils::DataLayoutIndexed dataLayoutIndexed(descriptor.m_DataLayout); |
| 4754 | |
| 4755 | Half width = static_cast<Half>(inputShape[dataLayoutIndexed.GetWidthIndex()]); |
| 4756 | Half height = static_cast<Half>(inputShape[dataLayoutIndexed.GetHeightIndex()]); |
| 4757 | |
| 4758 | descriptor.m_TargetWidth = std::floor(width * widthScale); |
| 4759 | descriptor.m_TargetHeight = std::floor(height * heightScale); |
| 4760 | } |
| 4761 | else |
| 4762 | { |
| 4763 | return Fail("%s: Operand has invalid data type for resizing by scale", __func__); |
| 4764 | } |
| 4765 | |
| 4766 | descriptor.m_AlignCorners = GetOptionalBool(operation, 4, model, data); |
| 4767 | descriptor.m_HalfPixelCenters = GetOptionalBool(operation, 5, model, data); |
| 4768 | |
| 4769 | bool isSupported = false; |
Cathal Corbett | 5383767 | 2022-09-01 11:34:37 +0100 | [diff] [blame] | 4770 | armnn::BackendId setBackend; |
Sadik Armagan | 8f397a1 | 2022-06-17 15:38:22 +0100 | [diff] [blame] | 4771 | auto validateFunc = [&](const armnn::TensorInfo& outputInfo, bool& isSupported) |
| 4772 | { |
| 4773 | FORWARD_LAYER_SUPPORT_FUNC(__func__, |
| 4774 | IsResizeSupported, |
| 4775 | data.m_Backends, |
| 4776 | isSupported, |
Cathal Corbett | 5383767 | 2022-09-01 11:34:37 +0100 | [diff] [blame] | 4777 | setBackend, |
Sadik Armagan | 8f397a1 | 2022-06-17 15:38:22 +0100 | [diff] [blame] | 4778 | inputInfo, |
| 4779 | outputInfo, |
| 4780 | descriptor); |
| 4781 | }; |
| 4782 | |
| 4783 | if(IsDynamicTensor(outputInfo)) |
| 4784 | { |
| 4785 | isSupported = AreDynamicTensorsSupported(); |
| 4786 | } |
| 4787 | else |
| 4788 | { |
| 4789 | validateFunc(outputInfo, isSupported); |
| 4790 | } |
| 4791 | |
| 4792 | if (!isSupported) |
| 4793 | { |
| 4794 | return false; |
| 4795 | } |
| 4796 | |
| 4797 | IConnectableLayer* layer = data.m_Network->AddResizeLayer(descriptor); |
Cathal Corbett | 5383767 | 2022-09-01 11:34:37 +0100 | [diff] [blame] | 4798 | layer->SetBackendId(setBackend); |
Sadik Armagan | 8f397a1 | 2022-06-17 15:38:22 +0100 | [diff] [blame] | 4799 | assert(layer != nullptr); |
| 4800 | input.Connect(layer->GetInputSlot(0)); |
| 4801 | |
| 4802 | return SetupAndTrackLayerOutputSlot(operation, 0, *layer, model, data, nullptr, validateFunc); |
| 4803 | } |
| 4804 | |
Teresa Charlin | ec0959b | 2023-07-27 13:23:15 +0100 | [diff] [blame] | 4805 | bool Converter::ConvertReverseV2(const Operation& operation, const Model& model, ConversionData& data) |
| 4806 | { |
| 4807 | VLOG(DRIVER) << "Converter::ConvertReverseV2()"; |
| 4808 | |
| 4809 | LayerInputHandle input0 = ConvertToLayerInputHandle(operation, 0, model, data); |
| 4810 | LayerInputHandle input1 = ConvertToLayerInputHandle(operation, 1, model, data); |
| 4811 | if (!input0.IsValid() || !input1.IsValid()) |
| 4812 | { |
| 4813 | return Fail("%s: Operation has invalid inputs", __func__); |
| 4814 | } |
| 4815 | const armnn::TensorInfo& inputInfo0 = input0.GetTensorInfo(); |
| 4816 | const armnn::TensorInfo& inputInfo1 = input1.GetTensorInfo(); |
| 4817 | |
| 4818 | const Operand* outputOperand = GetOutputOperand(operation, 0, model); |
| 4819 | if (!outputOperand) |
| 4820 | { |
| 4821 | return Fail("%s: Could not read output 0", __func__); |
| 4822 | } |
| 4823 | const armnn::TensorInfo& outputInfo = GetTensorInfoForOperand(*outputOperand); |
| 4824 | |
| 4825 | bool isSupported = false; |
| 4826 | armnn::BackendId setBackend; |
| 4827 | auto validateFunc = [&](const armnn::TensorInfo& outputInfo, bool& isSupported) |
| 4828 | { |
| 4829 | FORWARD_LAYER_SUPPORT_FUNC(__func__, |
| 4830 | IsReverseV2Supported, |
| 4831 | data.m_Backends, |
| 4832 | isSupported, |
| 4833 | setBackend, |
| 4834 | inputInfo0, |
| 4835 | inputInfo1, |
| 4836 | outputInfo); |
| 4837 | }; |
| 4838 | |
| 4839 | if(!IsDynamicTensor(outputInfo)) |
| 4840 | { |
| 4841 | validateFunc(outputInfo, isSupported); |
| 4842 | } |
| 4843 | else |
| 4844 | { |
| 4845 | isSupported = AreDynamicTensorsSupported(); |
| 4846 | } |
| 4847 | |
| 4848 | if (!isSupported) |
| 4849 | { |
| 4850 | return false; |
| 4851 | } |
| 4852 | |
| 4853 | armnn::IConnectableLayer* const layer = data.m_Network->AddReverseV2Layer(); |
| 4854 | layer->SetBackendId(setBackend); |
| 4855 | assert(layer != nullptr); |
| 4856 | input0.Connect(layer->GetInputSlot(0)); |
| 4857 | input1.Connect(layer->GetInputSlot(1)); |
| 4858 | |
| 4859 | return SetupAndTrackLayerOutputSlot(operation, 0, *layer, model, data, nullptr, validateFunc); |
| 4860 | } |
| 4861 | |
Sadik Armagan | 8f397a1 | 2022-06-17 15:38:22 +0100 | [diff] [blame] | 4862 | bool Converter::ConvertSpaceToBatchNd(const Operation& operation, const Model& model, ConversionData& data) |
| 4863 | { |
| 4864 | VLOG(DRIVER) << "Converter::ConvertSpaceToBatchNd()"; |
| 4865 | |
| 4866 | LayerInputHandle input = ConvertToLayerInputHandle(operation, 0, model, data); |
| 4867 | if(!input.IsValid()) |
| 4868 | { |
| 4869 | return Fail("%s: Operation has invalid inputs", __func__); |
| 4870 | } |
| 4871 | |
| 4872 | const armnn::TensorInfo &inputInfo = input.GetTensorInfo(); |
| 4873 | unsigned int rank = inputInfo.GetNumDimensions(); |
| 4874 | unsigned int spatialDim = rank - 2; |
| 4875 | |
| 4876 | if(rank != 4) |
| 4877 | { |
| 4878 | Fail("%s: Only inputs with rank 4 are supported", __func__); |
| 4879 | } |
| 4880 | |
| 4881 | const Operand *output = GetOutputOperand(operation, 0, model); |
| 4882 | if(!output) |
| 4883 | { |
| 4884 | return Fail("%s: Could not read output 0", __func__); |
| 4885 | } |
| 4886 | |
| 4887 | const armnn::TensorInfo &outputInfo = GetTensorInfoForOperand(*output); |
| 4888 | |
| 4889 | const Operand *blockShapeOperand = GetInputOperand(operation, 1, model); |
| 4890 | const Operand *paddingsOperand = GetInputOperand(operation, 2, model); |
| 4891 | |
| 4892 | armnn::TensorShape blockShapeOperandShape = GetTensorShapeForOperand(*blockShapeOperand); |
| 4893 | if(blockShapeOperandShape.GetNumDimensions() != 1 || blockShapeOperandShape.GetNumElements() != spatialDim) |
| 4894 | { |
| 4895 | return Fail("%s: Operation has invalid block shape operand: expected shape [%d]", __func__, spatialDim); |
| 4896 | } |
| 4897 | |
| 4898 | std::vector<int32_t> blockShape; |
| 4899 | if(!GetTensorInt32Values(*blockShapeOperand, blockShape, model, data)) |
| 4900 | { |
| 4901 | return Fail("%s: Operation has an invalid or unsupported block size operand", __func__); |
| 4902 | } |
| 4903 | if(std::any_of(blockShape.cbegin(), blockShape.cend(), [](int32_t i) |
| 4904 | { return i < 1; })) |
| 4905 | { |
| 4906 | return Fail("%s: Block shape must be at least 1 in all dimensions.", __func__); |
| 4907 | } |
| 4908 | |
| 4909 | armnn::TensorShape paddingsOperandShape = GetTensorShapeForOperand(*paddingsOperand); |
| 4910 | if(paddingsOperandShape.GetNumDimensions() != 2 || paddingsOperandShape.GetNumElements() != 2 * spatialDim) |
| 4911 | { |
| 4912 | return Fail("%s: Operation has invalid paddings operand: expected shape [%d, 2]", __func__, spatialDim); |
| 4913 | } |
| 4914 | |
| 4915 | std::vector<std::pair<unsigned int, unsigned int>> paddingList; |
| 4916 | std::vector<int32_t> paddings; |
| 4917 | if(!GetTensorInt32Values(*paddingsOperand, paddings, model, data)) |
| 4918 | { |
| 4919 | return Fail("%s: Operation has an invalid or unsupported paddings operand", __func__); |
| 4920 | } |
| 4921 | for (unsigned int i = 0; i < paddings.size() - 1; i += 2) |
| 4922 | { |
| 4923 | int paddingBeforeInput = paddings[i]; |
| 4924 | int paddingAfterInput = paddings[i + 1]; |
| 4925 | if(paddingBeforeInput < 0 || paddingAfterInput < 0) |
| 4926 | { |
| 4927 | return Fail("%s: Operation has invalid paddings operand, invalid padding values.", __func__); |
| 4928 | } |
| 4929 | |
| 4930 | paddingList.emplace_back((unsigned int) paddingBeforeInput, (unsigned int) paddingAfterInput); |
| 4931 | } |
| 4932 | |
| 4933 | armnn::SpaceToBatchNdDescriptor descriptor; |
| 4934 | descriptor.m_DataLayout = armnn::DataLayout::NHWC; |
| 4935 | descriptor.m_BlockShape.assign(blockShape.cbegin(), blockShape.cend()); |
| 4936 | descriptor.m_PadList.assign(paddingList.cbegin(), paddingList.cend()); |
| 4937 | |
| 4938 | if(Is12OrLaterOperand(*output)) |
| 4939 | { |
| 4940 | descriptor.m_DataLayout = OptionalDataLayout(operation, 3, model, data); |
| 4941 | } |
| 4942 | |
| 4943 | bool isSupported = false; |
Cathal Corbett | 5383767 | 2022-09-01 11:34:37 +0100 | [diff] [blame] | 4944 | armnn::BackendId setBackend; |
Sadik Armagan | 8f397a1 | 2022-06-17 15:38:22 +0100 | [diff] [blame] | 4945 | auto validateFunc = [&](const armnn::TensorInfo &outputInfo, bool &isSupported) |
| 4946 | { |
| 4947 | FORWARD_LAYER_SUPPORT_FUNC(__func__, |
| 4948 | IsSpaceToBatchNdSupported, |
| 4949 | data.m_Backends, |
| 4950 | isSupported, |
Cathal Corbett | 5383767 | 2022-09-01 11:34:37 +0100 | [diff] [blame] | 4951 | setBackend, |
Sadik Armagan | 8f397a1 | 2022-06-17 15:38:22 +0100 | [diff] [blame] | 4952 | inputInfo, |
| 4953 | outputInfo, |
| 4954 | descriptor); |
| 4955 | }; |
| 4956 | |
| 4957 | if(IsDynamicTensor(outputInfo)) |
| 4958 | { |
| 4959 | isSupported = AreDynamicTensorsSupported(); |
| 4960 | } else |
| 4961 | { |
| 4962 | validateFunc(outputInfo, isSupported); |
| 4963 | } |
| 4964 | |
| 4965 | if(!isSupported) |
| 4966 | { |
| 4967 | return false; |
| 4968 | } |
| 4969 | |
| 4970 | armnn::IConnectableLayer *const layer = data.m_Network->AddSpaceToBatchNdLayer(descriptor); |
Cathal Corbett | 5383767 | 2022-09-01 11:34:37 +0100 | [diff] [blame] | 4971 | layer->SetBackendId(setBackend); |
Sadik Armagan | 8f397a1 | 2022-06-17 15:38:22 +0100 | [diff] [blame] | 4972 | assert(layer != nullptr); |
| 4973 | input.Connect(layer->GetInputSlot(0)); |
| 4974 | |
| 4975 | return SetupAndTrackLayerOutputSlot(operation, 0, *layer, model, data, nullptr, validateFunc); |
| 4976 | } |
| 4977 | |
| 4978 | bool Converter::ConvertSpaceToDepth(const Operation& operation, const Model& model, ConversionData& data) |
| 4979 | { |
| 4980 | VLOG(DRIVER) << "Converter::ConvertSpaceToDepth()"; |
| 4981 | |
| 4982 | LayerInputHandle input = ConvertToLayerInputHandle(operation, 0, model, data); |
| 4983 | if (!input.IsValid() ) |
| 4984 | { |
| 4985 | return Fail("%s: Operation has invalid inputs", __func__); |
| 4986 | } |
| 4987 | |
| 4988 | const TensorInfo& inputInfo = input.GetTensorInfo(); |
| 4989 | unsigned int rank = inputInfo.GetNumDimensions(); |
| 4990 | if (rank != 4) |
| 4991 | { |
| 4992 | return Fail("%s: Only inputs with rank 4 are supported", __func__); |
| 4993 | } |
| 4994 | |
| 4995 | const Operand* output = GetOutputOperand(operation, 0, model); |
| 4996 | if (!output) |
| 4997 | { |
| 4998 | return Fail("%s: Could not read output 0", __func__); |
| 4999 | } |
| 5000 | |
| 5001 | const TensorInfo& outputInfo = GetTensorInfoForOperand(*output); |
| 5002 | |
| 5003 | SpaceToDepthDescriptor desc; |
| 5004 | |
| 5005 | GetInputScalar(operation, 1, OperandType::INT32, desc.m_BlockSize, model, data); |
| 5006 | |
| 5007 | if (desc.m_BlockSize <= 1) |
| 5008 | { |
Kevin May | 21e3999 | 2023-07-31 12:20:03 +0100 | [diff] [blame] | 5009 | return Fail("%s: Block size must be at least 1 in all dimensions", __func__); |
Sadik Armagan | 8f397a1 | 2022-06-17 15:38:22 +0100 | [diff] [blame] | 5010 | } |
| 5011 | |
| 5012 | desc.m_DataLayout = OptionalDataLayout(operation, 2, model, data); |
| 5013 | |
| 5014 | bool isSupported = false; |
Cathal Corbett | 5383767 | 2022-09-01 11:34:37 +0100 | [diff] [blame] | 5015 | armnn::BackendId setBackend; |
Sadik Armagan | 8f397a1 | 2022-06-17 15:38:22 +0100 | [diff] [blame] | 5016 | auto validateFunc = [&](const armnn::TensorInfo& outputInfo, bool& isSupported) |
| 5017 | { |
| 5018 | FORWARD_LAYER_SUPPORT_FUNC(__func__, |
| 5019 | IsSpaceToDepthSupported, |
| 5020 | data.m_Backends, |
| 5021 | isSupported, |
Cathal Corbett | 5383767 | 2022-09-01 11:34:37 +0100 | [diff] [blame] | 5022 | setBackend, |
Sadik Armagan | 8f397a1 | 2022-06-17 15:38:22 +0100 | [diff] [blame] | 5023 | inputInfo, |
| 5024 | outputInfo, |
| 5025 | desc); |
| 5026 | }; |
| 5027 | |
| 5028 | if(IsDynamicTensor(outputInfo)) |
| 5029 | { |
| 5030 | isSupported = AreDynamicTensorsSupported(); |
| 5031 | } |
| 5032 | else |
| 5033 | { |
| 5034 | validateFunc(outputInfo, isSupported); |
| 5035 | } |
| 5036 | |
| 5037 | if (!isSupported) |
| 5038 | { |
| 5039 | return false; |
| 5040 | } |
| 5041 | |
| 5042 | IConnectableLayer* const layer = data.m_Network->AddSpaceToDepthLayer(desc); |
Cathal Corbett | 5383767 | 2022-09-01 11:34:37 +0100 | [diff] [blame] | 5043 | layer->SetBackendId(setBackend); |
Sadik Armagan | 8f397a1 | 2022-06-17 15:38:22 +0100 | [diff] [blame] | 5044 | assert(layer != nullptr); |
| 5045 | input.Connect(layer->GetInputSlot(0)); |
| 5046 | |
| 5047 | return SetupAndTrackLayerOutputSlot(operation, 0, *layer, model, data, nullptr, validateFunc); |
| 5048 | } |
| 5049 | |
| 5050 | bool Converter::ConvertSoftmax(const Operation& operation, const Model& model, ConversionData& data) |
| 5051 | { |
| 5052 | VLOG(DRIVER) << "Converter::ConvertSoftmax()"; |
| 5053 | |
| 5054 | LayerInputHandle input = ConvertToLayerInputHandle(operation, 0, model, data); |
| 5055 | if (!input.IsValid()) |
| 5056 | { |
| 5057 | return Fail("%s: Operation has invalid inputs", __func__); |
| 5058 | } |
| 5059 | |
| 5060 | const Operand* outputOperand = GetOutputOperand(operation, 0, model); |
| 5061 | if (!outputOperand) |
| 5062 | { |
| 5063 | return Fail("%s: Operation has no outputs", __func__); |
| 5064 | } |
| 5065 | |
| 5066 | const TensorInfo& outputInfo = GetTensorInfoForOperand(*outputOperand); |
| 5067 | |
| 5068 | SoftmaxDescriptor desc; |
| 5069 | OperandType outputType = outputOperand->type; |
| 5070 | |
| 5071 | // Read beta value |
| 5072 | if (outputType == OperandType::TENSOR_FLOAT16) |
| 5073 | { |
| 5074 | Half value; |
| 5075 | |
| 5076 | if (!GetInputScalar(operation, 1, OperandType::FLOAT16, value, model, data)) |
| 5077 | { |
| 5078 | return Fail("%s: Operation has invalid inputs %d", __func__, outputType); |
| 5079 | } |
| 5080 | |
| 5081 | desc.m_Beta = static_cast<float>(value); |
| 5082 | } |
| 5083 | else |
| 5084 | { |
| 5085 | if (!GetInputFloat32(operation, 1, desc.m_Beta, model, data)) |
| 5086 | { |
| 5087 | return Fail("%s: Operation has invalid inputs %d", __func__, outputType); |
| 5088 | } |
| 5089 | } |
| 5090 | |
| 5091 | if (operation.inputs.size() > 2 && !GetInputScalar(operation, |
| 5092 | 2, |
| 5093 | OperandType::INT32, |
| 5094 | desc.m_Axis, |
| 5095 | model, |
| 5096 | data)) |
| 5097 | { |
| 5098 | return Fail("%s: Operation has invalid inputs", __func__); |
| 5099 | } |
| 5100 | |
| 5101 | bool isSupported = false; |
Cathal Corbett | 5383767 | 2022-09-01 11:34:37 +0100 | [diff] [blame] | 5102 | armnn::BackendId setBackend; |
Sadik Armagan | 8f397a1 | 2022-06-17 15:38:22 +0100 | [diff] [blame] | 5103 | auto validateFunc = [&](const armnn::TensorInfo& outputInfo, bool& isSupported) |
| 5104 | { |
| 5105 | FORWARD_LAYER_SUPPORT_FUNC(__func__, |
| 5106 | IsSoftmaxSupported, |
| 5107 | data.m_Backends, |
| 5108 | isSupported, |
Cathal Corbett | 5383767 | 2022-09-01 11:34:37 +0100 | [diff] [blame] | 5109 | setBackend, |
Sadik Armagan | 8f397a1 | 2022-06-17 15:38:22 +0100 | [diff] [blame] | 5110 | input.GetTensorInfo(), |
| 5111 | outputInfo, |
| 5112 | desc); |
| 5113 | }; |
| 5114 | |
| 5115 | if(IsDynamicTensor(outputInfo)) |
| 5116 | { |
| 5117 | isSupported = AreDynamicTensorsSupported(); |
| 5118 | } |
| 5119 | else |
| 5120 | { |
| 5121 | validateFunc(outputInfo, isSupported); |
| 5122 | } |
| 5123 | |
| 5124 | if (!isSupported) |
| 5125 | { |
| 5126 | return false; |
| 5127 | } |
| 5128 | |
| 5129 | IConnectableLayer* layer = data.m_Network->AddSoftmaxLayer(desc); |
Cathal Corbett | 5383767 | 2022-09-01 11:34:37 +0100 | [diff] [blame] | 5130 | layer->SetBackendId(setBackend); |
Sadik Armagan | 8f397a1 | 2022-06-17 15:38:22 +0100 | [diff] [blame] | 5131 | assert(layer != nullptr); |
| 5132 | input.Connect(layer->GetInputSlot(0)); |
| 5133 | |
| 5134 | return SetupAndTrackLayerOutputSlot(operation, 0, *layer, model, data, nullptr, validateFunc); |
| 5135 | } |
| 5136 | |
Sadik Armagan | 8f397a1 | 2022-06-17 15:38:22 +0100 | [diff] [blame] | 5137 | bool Converter::ConvertTanH(const Operation& operation, const Model& model, ConversionData& data) |
| 5138 | { |
| 5139 | VLOG(DRIVER) << "Converter::ConvertTanH()"; |
| 5140 | |
| 5141 | armnn::ActivationDescriptor desc; |
| 5142 | desc.m_Function = armnn::ActivationFunction::TanH; |
| 5143 | desc.m_A = 1.0f; // android nn does not support tanH parameters |
| 5144 | desc.m_B = 1.0f; // set to 1.0f for unity scaling |
| 5145 | |
| 5146 | return ConvertToActivation(operation, __func__, desc, model, data); |
| 5147 | } |
| 5148 | |
Teresa Charlin | a0cc6ff | 2023-07-27 14:11:54 +0100 | [diff] [blame] | 5149 | bool Converter::ConvertTile(const Operation& operation, const Model& model, ConversionData& data) |
| 5150 | { |
| 5151 | VLOG(DRIVER) << "Converter::ConvertTile()"; |
| 5152 | |
| 5153 | LayerInputHandle input = ConvertToLayerInputHandle(operation, 0, model, data); |
| 5154 | if (!input.IsValid()) |
| 5155 | { |
| 5156 | return Fail("%s: Operation has invalid inputs", __func__); |
| 5157 | } |
| 5158 | const armnn::TensorInfo& inputInfo = input.GetTensorInfo(); |
| 5159 | |
| 5160 | const Operand* outputOperand = GetOutputOperand(operation, 0, model); |
| 5161 | if (!outputOperand) |
| 5162 | { |
| 5163 | return Fail("%s: Operation has no outputs", __func__); |
| 5164 | } |
| 5165 | const TensorInfo& outputInfo = GetTensorInfoForOperand(*outputOperand); |
| 5166 | |
| 5167 | const Operand* multiplesOperand = GetInputOperand(operation, 1, model); |
| 5168 | if (!multiplesOperand) |
| 5169 | { |
| 5170 | return Fail("%s: Could not read input 1", __func__); |
| 5171 | } |
| 5172 | std::vector<int32_t> multiples; |
| 5173 | if (!GetTensorInt32Values(*multiplesOperand, multiples, model, data)) |
| 5174 | { |
| 5175 | return Fail("%s: Input 1 has invalid values", __func__); |
| 5176 | } |
| 5177 | |
| 5178 | TileDescriptor descriptor; |
| 5179 | descriptor.m_Multiples.assign(multiples.begin(), multiples.end()); |
| 5180 | |
| 5181 | bool isSupported = false; |
| 5182 | armnn::BackendId setBackend; |
| 5183 | auto validateFunc = [&](const armnn::TensorInfo& outputInfo, bool& isSupported) |
| 5184 | { |
| 5185 | FORWARD_LAYER_SUPPORT_FUNC(__func__, |
| 5186 | IsTileSupported, |
| 5187 | data.m_Backends, |
| 5188 | isSupported, |
| 5189 | setBackend, |
| 5190 | inputInfo, |
| 5191 | outputInfo, |
| 5192 | descriptor); |
| 5193 | }; |
| 5194 | |
| 5195 | if(IsDynamicTensor(outputInfo)) |
| 5196 | { |
| 5197 | isSupported = AreDynamicTensorsSupported(); |
| 5198 | } |
| 5199 | else |
| 5200 | { |
| 5201 | validateFunc(outputInfo, isSupported); |
| 5202 | } |
| 5203 | |
| 5204 | if (!isSupported) |
| 5205 | { |
| 5206 | return false; |
| 5207 | } |
| 5208 | |
| 5209 | IConnectableLayer* const layer = data.m_Network->AddTileLayer(descriptor); |
| 5210 | layer->SetBackendId(setBackend); |
| 5211 | assert(layer != nullptr); |
| 5212 | input.Connect(layer->GetInputSlot(0)); |
| 5213 | |
| 5214 | return SetupAndTrackLayerOutputSlot(operation, 0, *layer, model, data, nullptr, validateFunc); |
| 5215 | } |
| 5216 | |
Sadik Armagan | 8f397a1 | 2022-06-17 15:38:22 +0100 | [diff] [blame] | 5217 | bool Converter::ConvertTransposeConv2d(const Operation& operation, const Model& model, ConversionData& data) |
| 5218 | { |
| 5219 | VLOG(DRIVER) << "Converter::ConvertTransposeConv2d()"; |
| 5220 | |
| 5221 | LayerInputHandle input = ConvertToLayerInputHandle(operation, 0, model, data); |
| 5222 | |
| 5223 | if (!input.IsValid()) |
| 5224 | { |
| 5225 | return Fail("%s: Operation has invalid inputs", __func__); |
| 5226 | } |
| 5227 | |
| 5228 | const Operand* output = GetOutputOperand(operation, 0, model); |
| 5229 | |
| 5230 | if (!output) |
| 5231 | { |
| 5232 | return Fail("%s: Could not read output 0", __func__); |
| 5233 | } |
| 5234 | |
| 5235 | const TensorInfo& inputInfo = input.GetTensorInfo(); |
| 5236 | const TensorInfo& outputInfo = GetTensorInfoForOperand(*output); |
| 5237 | |
| 5238 | // ArmNN does not currently support non-fixed weights or bias |
| 5239 | // Find the shape of the weights tensor. In AndroidNN this will be [ 1, H, W, I * M ] |
| 5240 | const Operand* weightsOperand = GetInputOperand(operation, 1, model); |
| 5241 | |
| 5242 | if (weightsOperand == nullptr) |
| 5243 | { |
| 5244 | return Fail("%s: Operand is invalid", __func__); |
| 5245 | } |
| 5246 | TransposeConvolution2dDescriptor desc; |
| 5247 | desc.m_DataLayout = DataLayout::NHWC; |
| 5248 | |
| 5249 | // Determine whether padding is implicit or explicit |
| 5250 | bool implicitPadding = operation.inputs.size() == 9; |
| 5251 | |
| 5252 | if (implicitPadding ) |
| 5253 | { |
| 5254 | desc.m_DataLayout = OptionalDataLayout(operation, 8, model, data); |
| 5255 | } |
| 5256 | else |
| 5257 | { |
| 5258 | desc.m_DataLayout = OptionalDataLayout(operation, 10, model, data); |
| 5259 | } |
| 5260 | |
| 5261 | armnnUtils::DataLayoutIndexed dataLayoutIndexed(desc.m_DataLayout); |
| 5262 | unsigned int widthIndex = dataLayoutIndexed.GetWidthIndex(); |
| 5263 | unsigned int heightIndex = dataLayoutIndexed.GetHeightIndex(); |
| 5264 | |
| 5265 | const PermutationVector OHWIToOIHW = {0, 2, 3, 1}; |
| 5266 | |
| 5267 | // The shape of the weight is [depth_out, filter_height, filter_width, depth_in]. |
| 5268 | // We have to permute it to OIHW if the data layout is NCHW. |
| 5269 | const ConstTensorPin weightsPin = (desc.m_DataLayout == DataLayout::NCHW) ? |
| 5270 | ConvertOperationInputToConstTensorPin(operation, 1, |
| 5271 | model, data, OHWIToOIHW) : |
| 5272 | ConvertOperationInputToConstTensorPin(operation, 1, model, data); |
| 5273 | |
| 5274 | // Bias is a 1D tensor |
| 5275 | const ConstTensorPin biasPin = |
| 5276 | ConvertOperationInputToConstTensorPin(operation, 2, model, data); |
| 5277 | |
| 5278 | if (!weightsPin.IsValid()) |
| 5279 | { |
| 5280 | return Fail("%s: Operation has invalid weights", __func__); |
| 5281 | } |
| 5282 | |
| 5283 | if (!biasPin.IsValid()) |
| 5284 | { |
| 5285 | return Fail("%s: Operation has invalid biases", __func__); |
| 5286 | } |
| 5287 | |
| 5288 | ConstTensor weights = weightsPin.GetConstTensor(); |
| 5289 | ConstTensor bias = biasPin.GetConstTensor(); |
| 5290 | SanitizeBiasQuantizationScale(bias.GetInfo(), weights.GetInfo(), inputInfo); |
| 5291 | |
| 5292 | ActivationFn activation; |
| 5293 | |
| 5294 | if (implicitPadding) |
| 5295 | { |
| 5296 | int32_t strideX{0}; |
| 5297 | int32_t strideY{0}; |
| 5298 | int32_t padLeft{0}; |
| 5299 | int32_t padRight{0}; |
| 5300 | int32_t padTop{0}; |
| 5301 | int32_t padBottom{0}; |
| 5302 | |
| 5303 | ::android::nn::PaddingScheme paddingScheme; |
| 5304 | if (!GetInputPaddingScheme(operation, 4, paddingScheme, model, data) || |
| 5305 | !GetInputScalar(operation, 5, OperandType::INT32, strideX, model, data) || |
| 5306 | !GetInputScalar(operation, 6, OperandType::INT32, strideY, model, data) || |
| 5307 | !GetInputActivationFunction(operation, 7, activation, model, data)) |
| 5308 | { |
| 5309 | return Fail("%s: Operation has invalid inputs (implicit padding)", __func__); |
| 5310 | } |
| 5311 | |
| 5312 | const uint32_t kernelX = weights.GetShape()[widthIndex]; |
| 5313 | const uint32_t kernelY = weights.GetShape()[heightIndex]; |
| 5314 | |
| 5315 | // If output shape has been specified as a parameter then extract it and make it available. |
| 5316 | const Operand* outputShapeOperand = GetInputOperand(operation, 3, model, false); |
| 5317 | std::vector<int32_t> outputShape; |
| 5318 | if ((outputShapeOperand) && (GetTensorInt32Values(*outputShapeOperand, outputShape, model, data))) |
| 5319 | { |
| 5320 | // Change from signed to unsigned int to store in TransposeConvolution2dDescriptor. |
| 5321 | for (int dimension : outputShape) |
| 5322 | { |
| 5323 | desc.m_OutputShape.push_back(static_cast<unsigned int>(dimension)); |
| 5324 | } |
| 5325 | desc.m_OutputShapeEnabled = true; |
| 5326 | } |
| 5327 | |
| 5328 | uint32_t outputX; |
| 5329 | uint32_t outputY; |
| 5330 | |
| 5331 | if (IsDynamicTensor(outputInfo)) |
| 5332 | { |
| 5333 | if (outputShape.size() == 0) |
| 5334 | { |
| 5335 | return Fail("%s: Padding sizes cannot be inferred", __func__); |
| 5336 | } |
| 5337 | |
| 5338 | outputX = outputShape[widthIndex]; |
| 5339 | outputY = outputShape[heightIndex]; |
| 5340 | } |
| 5341 | else |
| 5342 | { |
| 5343 | outputX = outputInfo.GetShape()[widthIndex]; |
| 5344 | outputY = outputInfo.GetShape()[heightIndex]; |
| 5345 | } |
| 5346 | |
| 5347 | CalcPaddingTransposeConv(outputX, kernelX, strideX, padLeft, padRight, paddingScheme); |
| 5348 | CalcPaddingTransposeConv(outputY, kernelY, strideY, padTop, padBottom, paddingScheme); |
| 5349 | |
| 5350 | // NOTE: The Android NN API allows for negative padding values in TransposeConv2d, |
| 5351 | // but Arm NN only supports values >= 0 |
| 5352 | if (padLeft < 0 || padRight < 0 || padTop < 0 || padBottom < 0) |
| 5353 | { |
| 5354 | return Fail("%s: Negative padding values are not supported", __func__); |
| 5355 | } |
| 5356 | |
| 5357 | desc.m_StrideX = armnn::numeric_cast<uint32_t>(strideX); |
| 5358 | desc.m_StrideY = armnn::numeric_cast<uint32_t>(strideY); |
| 5359 | desc.m_PadLeft = armnn::numeric_cast<uint32_t>(padLeft); |
| 5360 | desc.m_PadRight = armnn::numeric_cast<uint32_t>(padRight); |
| 5361 | desc.m_PadTop = armnn::numeric_cast<uint32_t>(padTop); |
| 5362 | desc.m_PadBottom = armnn::numeric_cast<uint32_t>(padBottom); |
| 5363 | } |
| 5364 | else if (operation.inputs.size() == 11) |
| 5365 | { |
| 5366 | // explicit padding |
| 5367 | if (!GetInputScalar(operation, 3, OperandType::INT32, desc.m_PadLeft, model, data) || |
| 5368 | !GetInputScalar(operation, 4, OperandType::INT32, desc.m_PadRight, model, data) || |
| 5369 | !GetInputScalar(operation, 5, OperandType::INT32, desc.m_PadTop, model, data) || |
| 5370 | !GetInputScalar(operation, 6, OperandType::INT32, desc.m_PadBottom, model, data) || |
| 5371 | !GetInputScalar(operation, 7, OperandType::INT32, desc.m_StrideX, model, data) || |
| 5372 | !GetInputScalar(operation, 8, OperandType::INT32, desc.m_StrideY, model, data) || |
| 5373 | !GetInputActivationFunction(operation, 9, activation, model, data)) |
| 5374 | { |
| 5375 | return Fail("%s: Operation has invalid inputs (explicit padding)", __func__); |
| 5376 | } |
| 5377 | } |
| 5378 | else |
| 5379 | { |
| 5380 | return Fail("%s: Unsupported number of operation inputs", __func__); |
| 5381 | } |
| 5382 | |
| 5383 | desc.m_BiasEnabled = true; |
| 5384 | Optional<TensorInfo> biases(bias.GetInfo()); |
| 5385 | |
| 5386 | bool isSupported = false; |
Cathal Corbett | 5383767 | 2022-09-01 11:34:37 +0100 | [diff] [blame] | 5387 | armnn::BackendId setBackend; |
Sadik Armagan | 8f397a1 | 2022-06-17 15:38:22 +0100 | [diff] [blame] | 5388 | auto validateFunc = [&](const armnn::TensorInfo& outputInfo, bool& isSupported) |
| 5389 | { |
| 5390 | FORWARD_LAYER_SUPPORT_FUNC(__func__, |
| 5391 | IsTransposeConvolution2dSupported, |
| 5392 | data.m_Backends, |
| 5393 | isSupported, |
Cathal Corbett | 5383767 | 2022-09-01 11:34:37 +0100 | [diff] [blame] | 5394 | setBackend, |
Sadik Armagan | 8f397a1 | 2022-06-17 15:38:22 +0100 | [diff] [blame] | 5395 | inputInfo, |
| 5396 | outputInfo, |
| 5397 | desc, |
| 5398 | weights.GetInfo(), |
| 5399 | biases); |
| 5400 | }; |
| 5401 | |
| 5402 | if(IsDynamicTensor(outputInfo)) |
| 5403 | { |
| 5404 | isSupported = AreDynamicTensorsSupported(); |
| 5405 | } |
| 5406 | else |
| 5407 | { |
| 5408 | validateFunc(outputInfo, isSupported); |
| 5409 | } |
| 5410 | if (!isSupported) |
| 5411 | { |
| 5412 | return false; |
| 5413 | } |
| 5414 | |
| 5415 | IConnectableLayer* startLayer = |
| 5416 | data.m_Network->AddTransposeConvolution2dLayer(desc, weights, Optional<ConstTensor>(bias)); |
Cathal Corbett | 5383767 | 2022-09-01 11:34:37 +0100 | [diff] [blame] | 5417 | startLayer->SetBackendId(setBackend); |
Sadik Armagan | 8f397a1 | 2022-06-17 15:38:22 +0100 | [diff] [blame] | 5418 | if (!startLayer) |
| 5419 | { |
| 5420 | return Fail("%s: AddTransposeConvolution2dLayer failed", __func__); |
| 5421 | } |
| 5422 | |
| 5423 | input.Connect(startLayer->GetInputSlot(0)); |
| 5424 | |
| 5425 | return SetupAndTrackLayerOutputSlot(operation, 0, *startLayer, model, |
| 5426 | data, nullptr, validateFunc, activation); |
| 5427 | } |
| 5428 | |
| 5429 | bool Converter::ConvertSqrt(const Operation& operation, const Model& model, ConversionData& data) |
| 5430 | { |
| 5431 | VLOG(DRIVER) << "Converter::ConvertSqrt()"; |
| 5432 | ActivationDescriptor desc; |
| 5433 | desc.m_Function = ActivationFunction::Sqrt; |
| 5434 | |
| 5435 | return ::ConvertToActivation(operation, __func__, desc, model, data); |
| 5436 | } |
| 5437 | |
| 5438 | bool Converter::ConvertSqueeze(const Operation& operation, const Model& model, ConversionData& data) |
| 5439 | { |
| 5440 | VLOG(DRIVER) << "Converter::ConvertSqueeze()"; |
| 5441 | |
| 5442 | LayerInputHandle input = ConvertToLayerInputHandle(operation, 0, model, data); |
| 5443 | if (!input.IsValid()) |
| 5444 | { |
| 5445 | return Fail("%s: Operation has invalid inputs", __func__); |
| 5446 | } |
| 5447 | |
| 5448 | const armnn::TensorInfo& inputInfo = input.GetTensorInfo(); |
| 5449 | unsigned int rank = inputInfo.GetNumDimensions(); |
| 5450 | if (rank > 4) |
| 5451 | { |
| 5452 | Fail("%s: Inputs with rank greater than 4 are not supported", __func__); |
| 5453 | } |
| 5454 | |
| 5455 | const Operand* output = GetOutputOperand(operation, 0, model); |
| 5456 | if (!output) |
| 5457 | { |
| 5458 | return Fail("%s: Could not read output 0", __func__); |
| 5459 | } |
| 5460 | |
| 5461 | if (IsDynamicTensor(GetTensorInfoForOperand(*output)) && !(AreDynamicTensorsSupported())) |
| 5462 | { |
| 5463 | return Fail("%s: Dynamic output tensors are not supported", __func__); |
| 5464 | } |
| 5465 | |
| 5466 | // NOTE: Axis is an optional parameter to SQUEEZE, therefore we do not want to generate a failure |
| 5467 | // if the operand index is out of bounds. |
| 5468 | const Operand* axisOperand = GetInputOperand(operation, 1, model, false); |
| 5469 | |
| 5470 | const uint32_t dimensionSequence[] = { 0, 1, 2, 3 }; |
| 5471 | |
| 5472 | std::vector<int32_t> axis; |
| 5473 | if (!axisOperand) |
| 5474 | { |
| 5475 | axis.assign(dimensionSequence, |
| 5476 | dimensionSequence + rank); |
| 5477 | } |
| 5478 | else if (!GetTensorInt32Values(*axisOperand, axis, model, data)) |
| 5479 | { |
| 5480 | return Fail("%s: Operation has an invalid or unsupported axis operand", __func__); |
| 5481 | } |
| 5482 | |
| 5483 | std::vector<uint32_t> outputDims; |
| 5484 | for (unsigned int i = 0; i < rank; i++) |
| 5485 | { |
| 5486 | bool skipSqueeze = (std::find(axis.begin(), axis.end(), i) == axis.end()); |
| 5487 | auto currentDimension = inputInfo.GetShape()[i]; |
| 5488 | if (skipSqueeze || currentDimension != 1) |
| 5489 | { |
| 5490 | outputDims.push_back(currentDimension); |
| 5491 | } |
| 5492 | } |
| 5493 | |
| 5494 | armnn::TensorShape outShape = armnn::TensorShape(outputDims.size(), outputDims.data()); |
| 5495 | |
| 5496 | armnn::TensorInfo outputInfo = inputInfo; |
| 5497 | outputInfo.SetShape(outShape); |
| 5498 | |
| 5499 | armnn::ReshapeDescriptor reshapeDesc; |
| 5500 | reshapeDesc.m_TargetShape = outputInfo.GetShape(); |
| 5501 | |
| 5502 | bool isSupported = false; |
Cathal Corbett | 5383767 | 2022-09-01 11:34:37 +0100 | [diff] [blame] | 5503 | armnn::BackendId setBackend; |
Sadik Armagan | 8f397a1 | 2022-06-17 15:38:22 +0100 | [diff] [blame] | 5504 | FORWARD_LAYER_SUPPORT_FUNC(__func__, |
| 5505 | IsReshapeSupported, |
| 5506 | data.m_Backends, |
| 5507 | isSupported, |
Cathal Corbett | 5383767 | 2022-09-01 11:34:37 +0100 | [diff] [blame] | 5508 | setBackend, |
Sadik Armagan | 8f397a1 | 2022-06-17 15:38:22 +0100 | [diff] [blame] | 5509 | inputInfo, |
| 5510 | outputInfo, |
| 5511 | reshapeDesc); |
| 5512 | |
| 5513 | if (!isSupported) |
| 5514 | { |
| 5515 | return false; |
| 5516 | } |
| 5517 | |
| 5518 | armnn::IConnectableLayer* const layer = data.m_Network->AddReshapeLayer(reshapeDesc); |
Cathal Corbett | 5383767 | 2022-09-01 11:34:37 +0100 | [diff] [blame] | 5519 | layer->SetBackendId(setBackend); |
Sadik Armagan | 8f397a1 | 2022-06-17 15:38:22 +0100 | [diff] [blame] | 5520 | assert(layer != nullptr); |
| 5521 | input.Connect(layer->GetInputSlot(0)); |
| 5522 | |
| 5523 | return SetupAndTrackLayerOutputSlot(operation, 0, *layer, model, data); |
| 5524 | } |
| 5525 | |
| 5526 | bool Converter::ConvertStridedSlice(const Operation& operation, const Model& model, ConversionData& data) |
| 5527 | { |
| 5528 | VLOG(DRIVER) << "Converter::ConvertStridedSlice()"; |
| 5529 | |
| 5530 | LayerInputHandle input = ConvertToLayerInputHandle(operation, 0, model, data); |
| 5531 | if (!input.IsValid()) |
| 5532 | { |
| 5533 | return Fail("%s: Operation has invalid inputs", __func__); |
| 5534 | } |
| 5535 | |
| 5536 | const armnn::TensorInfo& inputInfo = input.GetTensorInfo(); |
| 5537 | unsigned int rank = inputInfo.GetNumDimensions(); |
| 5538 | if (rank > 4) |
| 5539 | { |
| 5540 | Fail("%s: Inputs with rank greater than 4 are not supported", __func__); |
| 5541 | } |
| 5542 | |
| 5543 | const Operand* output = GetOutputOperand(operation, 0, model); |
| 5544 | if (!output) |
| 5545 | { |
| 5546 | return Fail("%s: Could not read output 0", __func__); |
| 5547 | } |
| 5548 | |
| 5549 | const armnn::TensorInfo& outputInfo = GetTensorInfoForOperand(*output); |
| 5550 | |
| 5551 | const Operand* beginOperand = GetInputOperand(operation, 1, model); |
| 5552 | const Operand* endOperand = GetInputOperand(operation, 2, model); |
| 5553 | const Operand* stridesOperand = GetInputOperand(operation, 3, model); |
| 5554 | |
| 5555 | std::vector<int32_t> beginValues; |
| 5556 | std::vector<int32_t> endValues; |
| 5557 | std::vector<int32_t> stridesValues; |
| 5558 | |
| 5559 | // The length of the beginOperand, endOperand and stridesOperand must be of a rank(input) |
| 5560 | auto ValidateInputOperands = [&] (const Operand& operand, std::vector<int32_t>& operandValues) |
| 5561 | { |
| 5562 | if (!GetTensorInt32Values(operand, operandValues, model, data)) |
| 5563 | { |
| 5564 | return false; |
| 5565 | } |
| 5566 | |
| 5567 | if (operandValues.size() != rank) |
| 5568 | { |
| 5569 | return false; |
| 5570 | } |
| 5571 | |
| 5572 | return true; |
| 5573 | }; |
| 5574 | |
| 5575 | if (!ValidateInputOperands(*beginOperand, beginValues) |
| 5576 | || !ValidateInputOperands(*endOperand, endValues) |
| 5577 | || !ValidateInputOperands(*stridesOperand, stridesValues)) |
| 5578 | { |
| 5579 | return Fail("%s: Operation has invalid input operand", __func__); |
| 5580 | } |
| 5581 | |
| 5582 | // Stride cannot have value '0' |
| 5583 | if (std::any_of(stridesValues.cbegin(), stridesValues.cend(), [](int32_t i){ return i == 0; })) |
| 5584 | { |
| 5585 | return Fail("%s: Stride must be non-zero value.", __func__); |
| 5586 | } |
| 5587 | |
| 5588 | armnn::StridedSliceDescriptor descriptor; |
| 5589 | descriptor.m_Begin.assign(beginValues.cbegin(), beginValues.cend()); |
| 5590 | descriptor.m_End.assign(endValues.cbegin(), endValues.cend()); |
| 5591 | descriptor.m_Stride.assign(stridesValues.cbegin(), stridesValues.cend()); |
| 5592 | descriptor.m_DataLayout = armnn::DataLayout::NHWC; |
| 5593 | |
| 5594 | // Get the "begin_mask", "end_mask", and "shrink_axis_mask" flags |
| 5595 | if (!GetInputInt32(operation, 4, descriptor.m_BeginMask, model, data) || |
| 5596 | !GetInputInt32(operation, 5, descriptor.m_EndMask, model, data) || |
| 5597 | !GetInputInt32(operation, 6, descriptor.m_ShrinkAxisMask, model, data)) |
| 5598 | { |
| 5599 | return Fail("%s: Operation has invalid inputs", __func__); |
| 5600 | } |
| 5601 | |
| 5602 | bool isSupported = false; |
Cathal Corbett | 5383767 | 2022-09-01 11:34:37 +0100 | [diff] [blame] | 5603 | armnn::BackendId setBackend; |
Sadik Armagan | 8f397a1 | 2022-06-17 15:38:22 +0100 | [diff] [blame] | 5604 | auto validateFunc = [&](const armnn::TensorInfo& outputInfo, bool& isSupported) |
| 5605 | { |
| 5606 | FORWARD_LAYER_SUPPORT_FUNC(__func__, |
| 5607 | IsStridedSliceSupported, |
| 5608 | data.m_Backends, |
| 5609 | isSupported, |
Cathal Corbett | 5383767 | 2022-09-01 11:34:37 +0100 | [diff] [blame] | 5610 | setBackend, |
Sadik Armagan | 8f397a1 | 2022-06-17 15:38:22 +0100 | [diff] [blame] | 5611 | inputInfo, |
| 5612 | outputInfo, |
| 5613 | descriptor); |
| 5614 | }; |
| 5615 | |
| 5616 | if(IsDynamicTensor(outputInfo)) |
| 5617 | { |
| 5618 | isSupported = AreDynamicTensorsSupported(); |
| 5619 | } |
| 5620 | else |
| 5621 | { |
| 5622 | validateFunc(outputInfo, isSupported); |
| 5623 | } |
| 5624 | |
| 5625 | if (!isSupported) |
| 5626 | { |
| 5627 | return false; |
| 5628 | } |
| 5629 | |
| 5630 | // Check if slice can fit in a inferred output |
| 5631 | armnn::TensorShape inputShape = inputInfo.GetShape(); |
| 5632 | for (unsigned int i = 0; i < inputShape.GetNumDimensions(); i++) |
| 5633 | { |
| 5634 | int stride = descriptor.m_Stride[i]; |
| 5635 | |
| 5636 | if (descriptor.m_ShrinkAxisMask & (1 << i)) |
| 5637 | { |
| 5638 | // If the difference between the start point and the end point of the slice on an axis being shrunk |
| 5639 | // is greater than 1 then throw an error as the output will not be large enough to hold the slice |
| 5640 | if (((descriptor.m_Begin[i] - descriptor.m_End[i]) > 1) |
| 5641 | || ((descriptor.m_Begin[i] - descriptor.m_End[i]) < -1)) |
| 5642 | { |
| 5643 | return Fail("%s: StridedSlice: Output will not be large enough to hold the slice", __func__); |
| 5644 | } |
| 5645 | |
| 5646 | if(stride < 0) |
| 5647 | { |
| 5648 | return Fail("%s: StridedSlice: Stride can not be negative while ShrinkAxisMask is set.", __func__); |
| 5649 | } |
| 5650 | } |
| 5651 | } |
| 5652 | |
| 5653 | armnn::IConnectableLayer* const layer = data.m_Network->AddStridedSliceLayer(descriptor); |
Cathal Corbett | 5383767 | 2022-09-01 11:34:37 +0100 | [diff] [blame] | 5654 | layer->SetBackendId(setBackend); |
Sadik Armagan | 8f397a1 | 2022-06-17 15:38:22 +0100 | [diff] [blame] | 5655 | assert(layer != nullptr); |
| 5656 | input.Connect(layer->GetInputSlot(0)); |
| 5657 | |
| 5658 | return SetupAndTrackLayerOutputSlot(operation, 0, *layer, model, data, nullptr, validateFunc); |
| 5659 | } |
| 5660 | |
| 5661 | bool Converter::ConvertTranspose(const Operation& operation, const Model& model, ConversionData& data) |
| 5662 | { |
| 5663 | VLOG(DRIVER) << "Converter::ConvertTranspose()"; |
| 5664 | |
| 5665 | LayerInputHandle input = ConvertToLayerInputHandle(operation, 0, model, data); |
| 5666 | if (!input.IsValid()) |
| 5667 | { |
| 5668 | return Fail("%s: Operation has invalid inputs", __func__); |
| 5669 | } |
| 5670 | |
| 5671 | const armnn::TensorInfo& inputInfo = input.GetTensorInfo(); |
| 5672 | unsigned int rank = inputInfo.GetNumDimensions(); |
| 5673 | if (rank > 4) |
| 5674 | { |
| 5675 | Fail("%s: Inputs with rank greater than 4 are not supported", __func__); |
| 5676 | } |
| 5677 | |
| 5678 | // NOTE: Axis is an optional parameter to TRANSPOSE, therefore we do not want to generate a failure |
| 5679 | // if the operand index is out of bounds. |
| 5680 | const Operand* permOperand = GetInputOperand(operation, 1, model, false); |
| 5681 | |
| 5682 | std::vector<int32_t> perm(rank); |
| 5683 | if (!permOperand || (permOperand->lifetime == OperandLifeTime::NO_VALUE)) |
| 5684 | { |
| 5685 | for (unsigned int i = rank; i > 0; i--) |
| 5686 | { |
| 5687 | perm[rank - i] = armnn::numeric_cast<int> (i - 1); |
| 5688 | } |
| 5689 | } |
| 5690 | else if (!GetTensorInt32Values(*permOperand, perm, model, data)) |
| 5691 | { |
| 5692 | return Fail("%s: Operation has an invalid or unsupported permutation operand", __func__); |
| 5693 | } |
| 5694 | |
| 5695 | std::vector<uint32_t> outputDims(perm.begin(), perm.begin() + rank); |
| 5696 | |
| 5697 | armnn::TransposeDescriptor transposeDesc; |
| 5698 | transposeDesc.m_DimMappings = armnn::PermutationVector(outputDims.data(), outputDims.size()); |
| 5699 | |
| 5700 | const Operand* output = GetOutputOperand(operation, 0, model); |
| 5701 | if (!output) |
| 5702 | { |
| 5703 | return Fail("%s: Could not read output 0", __func__); |
| 5704 | } |
| 5705 | |
| 5706 | const armnn::TensorInfo& outputInfo = GetTensorInfoForOperand(*output); |
| 5707 | |
| 5708 | bool isSupported = false; |
Cathal Corbett | 5383767 | 2022-09-01 11:34:37 +0100 | [diff] [blame] | 5709 | armnn::BackendId setBackend; |
Sadik Armagan | 8f397a1 | 2022-06-17 15:38:22 +0100 | [diff] [blame] | 5710 | auto validateFunc = [&](const armnn::TensorInfo& outputInfo, bool& isSupported) |
| 5711 | { |
| 5712 | FORWARD_LAYER_SUPPORT_FUNC(__func__, |
| 5713 | IsTransposeSupported, |
| 5714 | data.m_Backends, |
| 5715 | isSupported, |
Cathal Corbett | 5383767 | 2022-09-01 11:34:37 +0100 | [diff] [blame] | 5716 | setBackend, |
Sadik Armagan | 8f397a1 | 2022-06-17 15:38:22 +0100 | [diff] [blame] | 5717 | inputInfo, |
| 5718 | outputInfo, |
| 5719 | transposeDesc); |
| 5720 | }; |
| 5721 | |
| 5722 | if(IsDynamicTensor(outputInfo)) |
| 5723 | { |
| 5724 | isSupported = AreDynamicTensorsSupported(); |
| 5725 | } |
| 5726 | else |
| 5727 | { |
| 5728 | validateFunc(outputInfo, isSupported); |
| 5729 | } |
| 5730 | |
| 5731 | if (!isSupported) |
| 5732 | { |
| 5733 | return false; |
| 5734 | } |
| 5735 | |
| 5736 | armnn::IConnectableLayer* const layer = data.m_Network->AddTransposeLayer(transposeDesc); |
Cathal Corbett | 5383767 | 2022-09-01 11:34:37 +0100 | [diff] [blame] | 5737 | layer->SetBackendId(setBackend); |
Sadik Armagan | 8f397a1 | 2022-06-17 15:38:22 +0100 | [diff] [blame] | 5738 | assert(layer != nullptr); |
| 5739 | input.Connect(layer->GetInputSlot(0)); |
| 5740 | |
| 5741 | return SetupAndTrackLayerOutputSlot(operation, 0, *layer, model, data, nullptr, validateFunc); |
| 5742 | } |
| 5743 | |
| 5744 | } // namespace armnn_driver |