telsoa01 | 5307bc1 | 2018-03-09 13:51:08 +0000 | [diff] [blame] | 1 | // |
| 2 | // Copyright © 2017 Arm Ltd. All rights reserved. |
| 3 | // See LICENSE file in the project root for full license information. |
| 4 | // |
| 5 | |
| 6 | #define LOG_TAG "ArmnnDriver" |
| 7 | |
| 8 | #include "ModelToINetworkConverter.hpp" |
telsoa01 | ce3e84a | 2018-08-31 09:31:35 +0100 | [diff] [blame] | 9 | #include <OperationsUtils.h> |
telsoa01 | 5307bc1 | 2018-03-09 13:51:08 +0000 | [diff] [blame] | 10 | |
| 11 | #include <armnn/LayerSupport.hpp> |
| 12 | #include <Permute.hpp> |
| 13 | |
| 14 | #include <log/log.h> |
| 15 | #include <cassert> |
| 16 | |
| 17 | #include <boost/format.hpp> |
| 18 | #include <boost/core/ignore_unused.hpp> |
| 19 | #include <boost/test/tools/floating_point_comparison.hpp> |
| 20 | #include <boost/cast.hpp> |
arovir01 | 3b0a2da | 2018-08-29 10:16:58 +0100 | [diff] [blame] | 21 | #include <boost/optional.hpp> |
telsoa01 | 5307bc1 | 2018-03-09 13:51:08 +0000 | [diff] [blame] | 22 | |
telsoa01 | ce3e84a | 2018-08-31 09:31:35 +0100 | [diff] [blame] | 23 | using namespace android::hardware; |
| 24 | |
surmeh01 | 49b9e10 | 2018-05-17 14:11:25 +0100 | [diff] [blame] | 25 | namespace armnn_driver |
| 26 | { |
kevmay01 | bc5f784 | 2018-08-30 12:34:39 +0100 | [diff] [blame^] | 27 | |
surmeh01 | 49b9e10 | 2018-05-17 14:11:25 +0100 | [diff] [blame] | 28 | class LayerInputHandle |
| 29 | { |
| 30 | public: |
| 31 | LayerInputHandle() |
| 32 | : m_OutputSlot(nullptr) |
| 33 | , m_Valid(false) |
| 34 | {} |
| 35 | |
| 36 | LayerInputHandle(bool valid, armnn::IOutputSlot* outputSlot, armnn::TensorInfo tensorInfo) |
| 37 | : m_OutputSlot(outputSlot) |
| 38 | , m_Valid(valid) |
| 39 | , m_TensorInfo(tensorInfo) |
| 40 | {} |
| 41 | |
| 42 | bool IsValid() const { return m_Valid; } |
| 43 | void Connect(armnn::IInputSlot& inputSlot) |
| 44 | { |
| 45 | assert(IsValid()); |
| 46 | |
| 47 | if (m_OutputSlot) |
| 48 | { |
| 49 | m_OutputSlot->Connect(inputSlot); |
| 50 | } |
| 51 | } |
| 52 | const armnn::TensorInfo& GetTensorInfo() const { return m_TensorInfo; } |
| 53 | |
| 54 | private: |
| 55 | armnn::IOutputSlot* m_OutputSlot; |
| 56 | bool m_Valid; |
| 57 | armnn::TensorInfo m_TensorInfo; |
| 58 | }; |
kevmay01 | bc5f784 | 2018-08-30 12:34:39 +0100 | [diff] [blame^] | 59 | |
| 60 | } // namespace armnn_driver |
surmeh01 | 49b9e10 | 2018-05-17 14:11:25 +0100 | [diff] [blame] | 61 | |
telsoa01 | 5307bc1 | 2018-03-09 13:51:08 +0000 | [diff] [blame] | 62 | namespace |
| 63 | { |
kevmay01 | bc5f784 | 2018-08-30 12:34:39 +0100 | [diff] [blame^] | 64 | |
telsoa01 | 5307bc1 | 2018-03-09 13:51:08 +0000 | [diff] [blame] | 65 | using namespace armnn_driver; |
| 66 | using namespace android::nn; |
| 67 | |
| 68 | // Convenience function to log the reason for failing to convert a model. |
| 69 | // @return Always returns false (so that it can be used by callers as a quick way to signal an error and return) |
| 70 | template<class... Args> |
| 71 | static bool Fail(const char* formatStr, Args&&... args) |
| 72 | { |
| 73 | ALOGD(formatStr, std::forward<Args>(args)...); |
| 74 | return false; |
| 75 | } |
| 76 | |
| 77 | // Convenience function to call an Is*Supported function and log caller name together with reason for lack of support. |
| 78 | // Called as: IsLayerSupported(__func__, Is*Supported, a, b, c, d, e) |
| 79 | template<typename IsLayerSupportedFunc, typename ... Args> |
| 80 | bool IsLayerSupported(const char* funcName, IsLayerSupportedFunc f, Args&&... args) |
| 81 | { |
| 82 | std::vector<char> unsupportedReason(1024+1); |
| 83 | bool isSupported = f(std::forward<Args>(args)..., unsupportedReason.data(), unsupportedReason.size()-1); |
| 84 | if(isSupported) |
| 85 | { |
| 86 | return true; |
| 87 | } |
| 88 | else |
| 89 | { |
| 90 | std::string sUnsupportedReason(unsupportedReason.data()); |
| 91 | if (sUnsupportedReason.size() > 0) |
| 92 | { |
| 93 | ALOGD("%s: not supported by armnn: %s", funcName, sUnsupportedReason.c_str()); |
| 94 | } else |
| 95 | { |
| 96 | ALOGD("%s: not supported by armnn", funcName); |
| 97 | } |
| 98 | return false; |
| 99 | } |
| 100 | } |
| 101 | |
| 102 | armnn::TensorShape GetTensorShapeForOperand(const Operand& operand) |
| 103 | { |
| 104 | return armnn::TensorShape(operand.dimensions.size(), operand.dimensions.data()); |
| 105 | } |
| 106 | |
| 107 | inline bool IsOperandTypeSupportedForTensors(OperandType type) |
| 108 | { |
| 109 | return type == OperandType::TENSOR_FLOAT32 || |
| 110 | type == OperandType::TENSOR_QUANT8_ASYMM || |
| 111 | type == OperandType::TENSOR_INT32; |
| 112 | } |
| 113 | |
telsoa01 | ce3e84a | 2018-08-31 09:31:35 +0100 | [diff] [blame] | 114 | void BroadcastTensor(LayerInputHandle& input0, LayerInputHandle& input1, armnn::IConnectableLayer* startLayer, |
| 115 | armnn::INetwork& network) |
| 116 | { |
| 117 | BOOST_ASSERT(startLayer != nullptr); |
| 118 | const armnn::TensorInfo& inputTensorInfo0 = input0.GetTensorInfo(); |
| 119 | const armnn::TensorInfo& inputTensorInfo1 = input1.GetTensorInfo(); |
| 120 | |
| 121 | if (inputTensorInfo0.GetNumDimensions() != inputTensorInfo1.GetNumDimensions()) |
| 122 | { |
| 123 | // If the number of dimensions do not match then we need to add degenerate dimensions |
| 124 | // to the "smaller" tensor using a reshape: |
| 125 | // Small Big |
| 126 | // | | |
| 127 | // Reshape | |
| 128 | // \ / |
| 129 | // Add |
| 130 | bool input0IsBigger = inputTensorInfo0.GetNumDimensions() > inputTensorInfo1.GetNumDimensions(); |
| 131 | |
| 132 | LayerInputHandle& smallTensorHandle = input0IsBigger ? input1 : input0; |
| 133 | const armnn::TensorInfo& smallTensorDims = smallTensorHandle.GetTensorInfo(); |
| 134 | |
| 135 | LayerInputHandle& bigTensorHandle = input0IsBigger ? input0 : input1; |
| 136 | const armnn::TensorInfo& bigTensorDims = bigTensorHandle.GetTensorInfo(); |
| 137 | |
| 138 | const unsigned int bigTensorDimsNumber = bigTensorDims.GetNumDimensions(); |
| 139 | std::vector<unsigned int> reshapedDims(bigTensorDimsNumber, 1); |
| 140 | unsigned int sizeDifference = bigTensorDimsNumber - smallTensorDims.GetNumDimensions(); |
| 141 | for (unsigned i = sizeDifference; i < bigTensorDimsNumber; ++i) |
| 142 | { |
| 143 | reshapedDims[i] = smallTensorDims.GetShape()[i-sizeDifference]; |
| 144 | } |
| 145 | armnn::TensorInfo reshapedInfo = smallTensorDims; |
| 146 | reshapedInfo.SetShape(armnn::TensorShape{ static_cast<unsigned int>(reshapedDims.size()), |
| 147 | reshapedDims.data() }); |
| 148 | |
| 149 | armnn::ReshapeDescriptor reshapeDesc; |
| 150 | reshapeDesc.m_TargetShape = reshapedInfo.GetShape(); |
| 151 | armnn::IConnectableLayer* const reshapeLayer = network.AddReshapeLayer(reshapeDesc); |
| 152 | smallTensorHandle.Connect(reshapeLayer->GetInputSlot(0)); |
| 153 | reshapeLayer->GetOutputSlot(0).SetTensorInfo(reshapedInfo); |
| 154 | |
| 155 | // Connect the outputs from new reshape and original input layer |
| 156 | reshapeLayer->GetOutputSlot(0).Connect(startLayer->GetInputSlot(0)); |
| 157 | bigTensorHandle.Connect(startLayer->GetInputSlot(1)); |
| 158 | } |
| 159 | else |
| 160 | { |
| 161 | input0.Connect(startLayer->GetInputSlot(0)); |
| 162 | input1.Connect(startLayer->GetInputSlot(1)); |
| 163 | } |
| 164 | } |
| 165 | |
telsoa01 | 5307bc1 | 2018-03-09 13:51:08 +0000 | [diff] [blame] | 166 | void CalcPadding(uint32_t input, uint32_t kernel, uint32_t stride, uint32_t& outPadHead, uint32_t& outPadTail, |
| 167 | android::nn::PaddingScheme scheme) |
| 168 | { |
| 169 | int32_t padHead; |
| 170 | int32_t padTail; |
| 171 | calculateExplicitPadding(input, stride, kernel, scheme, &padHead, &padTail); |
| 172 | outPadHead = boost::numeric_cast<uint32_t>(padHead); |
| 173 | outPadTail = boost::numeric_cast<uint32_t>(padTail); |
| 174 | } |
| 175 | |
telsoa01 | 5307bc1 | 2018-03-09 13:51:08 +0000 | [diff] [blame] | 176 | Shape GetOperandShape(const Operand& operand) |
| 177 | { |
| 178 | Shape shape; |
| 179 | shape.type = operand.type; |
| 180 | shape.dimensions = operand.dimensions; |
| 181 | shape.scale = operand.scale; |
| 182 | shape.offset = operand.zeroPoint; |
| 183 | return shape; |
| 184 | } |
| 185 | |
| 186 | // ArmNN requires the bias scale to be equal to the product of the weight and input scales, which is also |
| 187 | // what AndroidNN requires. However for some of the AndroidNN tests the values don't exactly match so |
| 188 | // we accept some tolerance. We don't want to ArmNN itself to accept these inconsistencies as it is up to the user |
| 189 | // (us, in this case) to ensure they match. |
| 190 | void SanitizeBiasQuantizationScale(armnn::TensorInfo& biasInfo, |
| 191 | const armnn::TensorInfo& weightInfo, const armnn::TensorInfo& inputInfo) |
| 192 | { |
| 193 | const float expectedBiasScale = weightInfo.GetQuantizationScale() * inputInfo.GetQuantizationScale(); |
| 194 | if (biasInfo.GetQuantizationScale() != expectedBiasScale) |
| 195 | { |
| 196 | boost::math::fpc::close_at_tolerance<float> comparer(boost::math::fpc::percent_tolerance(1.0f)); |
| 197 | if (comparer(biasInfo.GetQuantizationScale(), expectedBiasScale)) |
| 198 | { |
| 199 | ALOGW("Bias quantization scale has been modified to match input*weights"); |
| 200 | biasInfo.SetQuantizationScale(expectedBiasScale); |
| 201 | } |
| 202 | } |
| 203 | } |
| 204 | |
telsoa01 | ce3e84a | 2018-08-31 09:31:35 +0100 | [diff] [blame] | 205 | // 4D Tensor Permutations |
| 206 | const armnn::PermutationVector IdentityPermutation4D({ 0U, 1U, 2U, 3U }); |
telsoa01 | 5307bc1 | 2018-03-09 13:51:08 +0000 | [diff] [blame] | 207 | const armnn::PermutationVector NHWCToArmNN({ 0U, 2U, 3U, 1U }); |
surmeh01 | 49b9e10 | 2018-05-17 14:11:25 +0100 | [diff] [blame] | 208 | const armnn::PermutationVector ArmNNToNHWC({ 0U, 3U, 1U, 2U }); |
| 209 | const armnn::PermutationVector SwapDim1And2({ 0U, 2U, 1U, 3U }); |
telsoa01 | 5307bc1 | 2018-03-09 13:51:08 +0000 | [diff] [blame] | 210 | |
telsoa01 | ce3e84a | 2018-08-31 09:31:35 +0100 | [diff] [blame] | 211 | // 3D Permutation Vectors |
| 212 | const armnn::PermutationVector IdentityPermutation3D({ 0U, 1U, 2U }); |
| 213 | const armnn::PermutationVector RotateTensorLeft({ 2U, 0U, 1U }); |
| 214 | const armnn::PermutationVector RotateTensorRight({ 1U, 2U, 0U }); |
| 215 | |
kevmay01 | bc5f784 | 2018-08-30 12:34:39 +0100 | [diff] [blame^] | 216 | template<typename OSlot> |
telsoa01 | 5307bc1 | 2018-03-09 13:51:08 +0000 | [diff] [blame] | 217 | armnn::IConnectableLayer& AddPermuteLayer(armnn::INetwork& network, OSlot& input, |
| 218 | const armnn::PermutationVector& mappings) |
| 219 | { |
| 220 | // Add swizzle layer |
| 221 | armnn::IConnectableLayer* const layer = network.AddPermuteLayer(mappings); |
| 222 | |
| 223 | assert(layer != nullptr); |
| 224 | |
telsoa01 | ce3e84a | 2018-08-31 09:31:35 +0100 | [diff] [blame] | 225 | // Connect input to swizzle layer |
telsoa01 | 5307bc1 | 2018-03-09 13:51:08 +0000 | [diff] [blame] | 226 | input.Connect(layer->GetInputSlot(0)); |
| 227 | |
| 228 | // Setup swizzled output |
| 229 | const armnn::TensorInfo outInfo = armnnUtils::Permuted(input.GetTensorInfo(), mappings); |
| 230 | layer->GetOutputSlot(0).SetTensorInfo(outInfo); |
| 231 | |
| 232 | return *layer; |
| 233 | } |
| 234 | |
telsoa01 | ce3e84a | 2018-08-31 09:31:35 +0100 | [diff] [blame] | 235 | void SwizzleIn(armnn::INetwork& network, LayerInputHandle& input, armnn::IConnectableLayer& layer, unsigned int index) |
telsoa01 | 5307bc1 | 2018-03-09 13:51:08 +0000 | [diff] [blame] | 236 | { |
telsoa01 | 5307bc1 | 2018-03-09 13:51:08 +0000 | [diff] [blame] | 237 | // Add swizzle layer |
| 238 | armnn::IConnectableLayer& swizzleLayer = AddPermuteLayer(network, input, NHWCToArmNN); |
telsoa01 | 5307bc1 | 2018-03-09 13:51:08 +0000 | [diff] [blame] | 239 | // Connect swizzled input to layer |
telsoa01 | ce3e84a | 2018-08-31 09:31:35 +0100 | [diff] [blame] | 240 | swizzleLayer.GetOutputSlot(0).Connect(layer.GetInputSlot(index)); |
| 241 | } |
telsoa01 | 5307bc1 | 2018-03-09 13:51:08 +0000 | [diff] [blame] | 242 | |
telsoa01 | ce3e84a | 2018-08-31 09:31:35 +0100 | [diff] [blame] | 243 | armnn::IConnectableLayer& DeswizzleOut(armnn::INetwork& network, armnn::IConnectableLayer& layer, unsigned int index) |
| 244 | { |
telsoa01 | 5307bc1 | 2018-03-09 13:51:08 +0000 | [diff] [blame] | 245 | // Add deswizzle layer |
telsoa01 | ce3e84a | 2018-08-31 09:31:35 +0100 | [diff] [blame] | 246 | armnn::IConnectableLayer& deswizzleLayer = AddPermuteLayer(network, layer.GetOutputSlot(index), ArmNNToNHWC); |
telsoa01 | 5307bc1 | 2018-03-09 13:51:08 +0000 | [diff] [blame] | 247 | return deswizzleLayer; |
| 248 | } |
| 249 | |
telsoa01 | ce3e84a | 2018-08-31 09:31:35 +0100 | [diff] [blame] | 250 | // only suitable for input/output slot index 0, for other slots, use SwizzleIn and DeswizzleOut directly |
| 251 | armnn::IConnectableLayer& SwizzleInDeswizzleOut(armnn::INetwork& network, |
| 252 | LayerInputHandle& input, |
| 253 | armnn::IConnectableLayer& firstLayer, |
| 254 | armnn::IConnectableLayer& lastLayer) |
| 255 | { |
| 256 | SwizzleIn(network, input, firstLayer, 0); |
| 257 | return DeswizzleOut(network, lastLayer, 0); |
| 258 | } |
| 259 | |
| 260 | // only suitable for input/output slot index 0, for other slots, use SwizzleIn and DeswizzleOut directly |
telsoa01 | 5307bc1 | 2018-03-09 13:51:08 +0000 | [diff] [blame] | 261 | armnn::IConnectableLayer& SwizzleInDeswizzleOut(armnn::INetwork& network, LayerInputHandle& input, |
| 262 | armnn::IConnectableLayer& layer) |
| 263 | { |
| 264 | return SwizzleInDeswizzleOut(network, input, layer, layer); |
| 265 | } |
surmeh01 | 49b9e10 | 2018-05-17 14:11:25 +0100 | [diff] [blame] | 266 | |
| 267 | bool ValidateConcatOutputShape(const std::vector<armnn::TensorShape> & inputShapes, |
| 268 | const armnn::TensorShape & outputShape, |
| 269 | uint32_t concatDim) |
| 270 | { |
| 271 | // Validate the output shape is correct given the input shapes (which have just been validated) |
| 272 | unsigned int numDimensions = inputShapes[0].GetNumDimensions(); |
| 273 | if (outputShape.GetNumDimensions() != numDimensions) |
| 274 | { |
| 275 | return Fail("%s: Output shape has wrong number of dimensions", __func__); |
| 276 | } |
| 277 | |
| 278 | unsigned int outputSizeAlongConcatenatedDimension = 0; |
| 279 | for (unsigned int i = 0; i < inputShapes.size(); i++) |
| 280 | { |
| 281 | outputSizeAlongConcatenatedDimension += inputShapes[i][concatDim]; |
| 282 | } |
| 283 | |
| 284 | for (unsigned int i = 0; i < numDimensions; ++i) |
| 285 | { |
| 286 | if (i == concatDim) |
| 287 | { |
| 288 | if (outputShape[i] != outputSizeAlongConcatenatedDimension) |
| 289 | { |
| 290 | return Fail( |
| 291 | "%s: Invalid output shape for dimension %d (%d != %d)", |
| 292 | __func__, |
| 293 | i, |
| 294 | outputShape[i], |
| 295 | outputSizeAlongConcatenatedDimension); |
| 296 | } |
| 297 | } |
| 298 | else |
| 299 | { |
| 300 | if (outputShape[i] != inputShapes[0][i]) |
| 301 | { |
| 302 | return Fail("%s: Invalid output shape", __func__); |
| 303 | } |
| 304 | } |
| 305 | } |
| 306 | |
| 307 | return true; |
| 308 | } |
| 309 | |
telsoa01 | ce3e84a | 2018-08-31 09:31:35 +0100 | [diff] [blame] | 310 | bool RequiresReshape(armnn::TensorShape & inputShape) |
| 311 | { |
| 312 | return inputShape.GetNumDimensions() < 3; |
| 313 | } |
| 314 | |
kevmay01 | bc5f784 | 2018-08-30 12:34:39 +0100 | [diff] [blame^] | 315 | template<typename OSlot> |
telsoa01 | ce3e84a | 2018-08-31 09:31:35 +0100 | [diff] [blame] | 316 | armnn::IConnectableLayer& AddReshapeLayer(armnn::INetwork& network, OSlot& inputLayer, |
| 317 | armnn::TensorInfo reshapeInfo) |
| 318 | { |
| 319 | armnn::ReshapeDescriptor reshapeDescriptor; |
| 320 | reshapeDescriptor.m_TargetShape = reshapeInfo.GetShape(); |
| 321 | |
| 322 | armnn::IConnectableLayer* reshapeLayer = network.AddReshapeLayer(reshapeDescriptor); |
| 323 | assert(reshapeLayer != nullptr); |
| 324 | |
| 325 | // Attach the input layer to the reshape layer |
| 326 | inputLayer.Connect(reshapeLayer->GetInputSlot(0)); |
| 327 | reshapeLayer->GetOutputSlot(0).SetTensorInfo(reshapeInfo); |
| 328 | |
| 329 | return *reshapeLayer; |
| 330 | } |
| 331 | |
surmeh01 | 49b9e10 | 2018-05-17 14:11:25 +0100 | [diff] [blame] | 332 | void SwizzleInputs(armnn::INetwork& network, |
| 333 | std::vector<LayerInputHandle>& inputs, |
| 334 | std::vector<armnn::TensorShape>& inputShapes, |
| 335 | const armnn::PermutationVector& mapping) |
| 336 | { |
telsoa01 | ce3e84a | 2018-08-31 09:31:35 +0100 | [diff] [blame] | 337 | if (!mapping.IsEqual(IdentityPermutation4D)) |
surmeh01 | 49b9e10 | 2018-05-17 14:11:25 +0100 | [diff] [blame] | 338 | { |
| 339 | size_t nInputs = inputs.size(); |
| 340 | for (size_t i=0; i<nInputs; ++i) |
| 341 | { |
| 342 | // add swizzle layer |
| 343 | armnn::IConnectableLayer& swizzleLayer = AddPermuteLayer(network, inputs[i], mapping); |
| 344 | auto& outputSlot = swizzleLayer.GetOutputSlot(0); |
| 345 | auto& outputInfo = outputSlot.GetTensorInfo(); |
| 346 | // replace inputs with the swizzled ones |
| 347 | inputs[i] = LayerInputHandle(true, &outputSlot, outputInfo); |
| 348 | inputShapes[i] = inputs[i].GetTensorInfo().GetShape(); |
| 349 | } |
| 350 | } |
| 351 | } |
| 352 | |
telsoa01 | ce3e84a | 2018-08-31 09:31:35 +0100 | [diff] [blame] | 353 | void CreatePermutationParameters(const unsigned int numberOfDimensions, |
| 354 | int32_t & concatDimension, |
| 355 | std::pair<armnn::PermutationVector, armnn::PermutationVector> & permutationPair) |
| 356 | { |
| 357 | assert(numberOfDimensions >= 3); |
| 358 | |
| 359 | // ArmNN uses Compute Library subtensors to perform concatenation |
| 360 | // This only works when concatenating along dimension 0 or 1 for a 4-D tensor, |
| 361 | // or along dimension 0 for a 3-D tensor. |
| 362 | if (numberOfDimensions == 4) |
| 363 | { |
| 364 | if (concatDimension == 3) |
| 365 | { |
| 366 | concatDimension = 1; |
| 367 | permutationPair = std::make_pair(NHWCToArmNN, ArmNNToNHWC); |
| 368 | } |
| 369 | else if (concatDimension == 2) |
| 370 | { |
| 371 | concatDimension = 1; |
| 372 | permutationPair = std::make_pair(SwapDim1And2, SwapDim1And2); |
| 373 | } |
| 374 | else |
| 375 | { |
| 376 | permutationPair = std::make_pair(IdentityPermutation4D, IdentityPermutation4D); |
| 377 | } |
| 378 | |
| 379 | } |
| 380 | else if (numberOfDimensions == 3) |
| 381 | { |
| 382 | if (concatDimension == 2) |
| 383 | { |
| 384 | concatDimension = 0; |
| 385 | permutationPair = std::make_pair(RotateTensorRight, RotateTensorLeft); |
| 386 | } |
| 387 | else if (concatDimension == 1) |
| 388 | { |
| 389 | concatDimension = 0; |
| 390 | permutationPair = std::make_pair(RotateTensorLeft, RotateTensorRight); |
| 391 | } |
| 392 | else |
| 393 | { |
| 394 | permutationPair = std::make_pair(IdentityPermutation3D, IdentityPermutation3D); |
| 395 | } |
| 396 | } |
| 397 | } |
| 398 | |
kevmay01 | bc5f784 | 2018-08-30 12:34:39 +0100 | [diff] [blame^] | 399 | } // anonymous namespace |
telsoa01 | 5307bc1 | 2018-03-09 13:51:08 +0000 | [diff] [blame] | 400 | |
| 401 | namespace armnn_driver |
| 402 | { |
| 403 | |
| 404 | class ConstTensorPin |
| 405 | { |
| 406 | public: |
| 407 | // Creates an invalid tensor pin (can be used to signal errors) |
telsoa01 | ce3e84a | 2018-08-31 09:31:35 +0100 | [diff] [blame] | 408 | // The optional flag can be set to indicate the tensor values were missing, but it was otherwise valid |
| 409 | ConstTensorPin(bool optional = false) : m_Optional(optional) {} |
telsoa01 | 5307bc1 | 2018-03-09 13:51:08 +0000 | [diff] [blame] | 410 | |
| 411 | // @param tensorInfo TensorInfo associated with the tensor. |
| 412 | // @param valueStart Start address of tensor data. Belongs to one of the memory pools associated with |
| 413 | // the model being converted. |
| 414 | // @param numBytes Number of bytes for the tensor data. |
| 415 | ConstTensorPin(const armnn::TensorInfo& tensorInfo, const void* valueStart, uint32_t numBytes, |
| 416 | const armnn::PermutationVector& mappings) |
| 417 | { |
| 418 | boost::ignore_unused(numBytes); |
| 419 | assert(tensorInfo.GetNumBytes() == numBytes); |
| 420 | |
| 421 | const bool needsSwizzling = (mappings.GetSize() > 0); |
| 422 | if (needsSwizzling) |
| 423 | { |
| 424 | m_SwizzledTensorData.resize(tensorInfo.GetNumBytes()); |
| 425 | SwizzleAndroidNn4dTensorToArmNn(tensorInfo, valueStart, m_SwizzledTensorData.data(), mappings); |
| 426 | |
| 427 | m_ConstTensor = armnn::ConstTensor(armnnUtils::Permuted(tensorInfo, mappings), m_SwizzledTensorData.data()); |
| 428 | } |
| 429 | else |
| 430 | { |
| 431 | m_ConstTensor = armnn::ConstTensor(tensorInfo, valueStart); |
| 432 | } |
| 433 | } |
| 434 | |
| 435 | ConstTensorPin(const ConstTensorPin& other) = delete; |
| 436 | ConstTensorPin(ConstTensorPin&& other) = default; |
| 437 | |
| 438 | bool IsValid() const { return m_ConstTensor.GetMemoryArea() != nullptr; } |
telsoa01 | ce3e84a | 2018-08-31 09:31:35 +0100 | [diff] [blame] | 439 | bool IsOptional() const { return m_Optional; } |
telsoa01 | 5307bc1 | 2018-03-09 13:51:08 +0000 | [diff] [blame] | 440 | const armnn::ConstTensor& GetConstTensor() const { return m_ConstTensor; } |
telsoa01 | ce3e84a | 2018-08-31 09:31:35 +0100 | [diff] [blame] | 441 | const armnn::ConstTensor* GetConstTensorPtr() const |
| 442 | { |
| 443 | if (IsValid() && m_ConstTensor.GetNumElements() > 0) |
| 444 | { |
| 445 | return &m_ConstTensor; |
| 446 | } |
| 447 | // tensor is either invalid, or has no elements (indicating an optional tensor that was not provided) |
| 448 | return nullptr; |
| 449 | } |
telsoa01 | 5307bc1 | 2018-03-09 13:51:08 +0000 | [diff] [blame] | 450 | |
| 451 | private: |
| 452 | armnn::ConstTensor m_ConstTensor; |
| 453 | // Owned memory for swizzled tensor data, only required if the tensor needed |
| 454 | // swizzling. Otherwise, @ref m_ConstTensor will reference memory from one of |
| 455 | // the pools associated with the model being converted. |
| 456 | std::vector<uint8_t> m_SwizzledTensorData; |
telsoa01 | ce3e84a | 2018-08-31 09:31:35 +0100 | [diff] [blame] | 457 | // optional flag to indicate that an invalid tensor pin is not an error, but the optional values were not given |
| 458 | bool m_Optional; |
telsoa01 | 5307bc1 | 2018-03-09 13:51:08 +0000 | [diff] [blame] | 459 | }; |
| 460 | |
kevmay01 | bc5f784 | 2018-08-30 12:34:39 +0100 | [diff] [blame^] | 461 | template<typename HalVersion> |
| 462 | ModelToINetworkConverter<HalVersion>::ModelToINetworkConverter(armnn::Compute compute, |
| 463 | const HalModel& model, |
telsoa01 | 5307bc1 | 2018-03-09 13:51:08 +0000 | [diff] [blame] | 464 | const std::set<unsigned int>& forcedUnsupportedOperations) |
| 465 | : m_Compute(compute) |
| 466 | , m_Model(model) |
| 467 | , m_ForcedUnsupportedOperations(forcedUnsupportedOperations) |
| 468 | , m_Network(nullptr, nullptr) |
| 469 | , m_ConversionResult(ConversionResult::Success) |
| 470 | { |
| 471 | try |
| 472 | { |
| 473 | Convert(); |
| 474 | } |
| 475 | catch (armnn::Exception& e) |
| 476 | { |
| 477 | m_ConversionResult = ConversionResult::UnsupportedFeature; |
| 478 | ALOGE("%s: Unexpected exception: %s", __func__, e.what()); |
| 479 | assert(false); |
| 480 | } |
| 481 | } |
| 482 | |
kevmay01 | bc5f784 | 2018-08-30 12:34:39 +0100 | [diff] [blame^] | 483 | template<typename HalVersion> |
| 484 | void ModelToINetworkConverter<HalVersion>::Convert() |
telsoa01 | 5307bc1 | 2018-03-09 13:51:08 +0000 | [diff] [blame] | 485 | { |
kevmay01 | bc5f784 | 2018-08-30 12:34:39 +0100 | [diff] [blame^] | 486 | using Model = typename HalVersion::Model; |
| 487 | ALOGV("ModelToINetworkConverter::Convert(): %s", GetModelSummary<Model>(m_Model).c_str()); |
telsoa01 | 5307bc1 | 2018-03-09 13:51:08 +0000 | [diff] [blame] | 488 | |
| 489 | // map the memory pool into shared pointers |
| 490 | m_MemPools.clear(); |
| 491 | if (!setRunTimePoolInfosFromHidlMemories(&m_MemPools, m_Model.pools)) |
| 492 | { |
| 493 | Fail("%s: Setting of run time pool infos from Hidl Memories has failed.", __func__); |
| 494 | m_ConversionResult = ConversionResult::ErrorMappingPools; |
| 495 | return; |
| 496 | } |
| 497 | |
| 498 | uint32_t totalPoolSize = 0; |
| 499 | for (auto&& pool : m_Model.pools) |
| 500 | { |
| 501 | totalPoolSize += pool.size(); |
| 502 | } |
| 503 | |
| 504 | // Create armnn::INetwork |
| 505 | m_Network = armnn::INetwork::Create(); |
| 506 | |
| 507 | // add operations to it |
| 508 | // track which layer outputs each operand |
| 509 | m_OutputSlotForOperand = std::vector<armnn::IOutputSlot*>(m_Model.operands.size(), nullptr); |
| 510 | |
| 511 | try |
| 512 | { |
| 513 | for (uint32_t i = 0; i < m_Model.inputIndexes.size(); i++) |
| 514 | { |
| 515 | // inputs in android nn are represented by operands |
| 516 | uint32_t inputIndex = m_Model.inputIndexes[i]; |
| 517 | const Operand& operand = m_Model.operands[inputIndex]; |
| 518 | const armnn::TensorInfo& tensor = GetTensorInfoForOperand(operand); |
| 519 | armnn::IConnectableLayer* layer = m_Network->AddInputLayer(i); |
| 520 | |
| 521 | armnn::IOutputSlot& outputSlot = layer->GetOutputSlot(0); |
| 522 | outputSlot.SetTensorInfo(GetTensorInfoForOperand(operand)); |
| 523 | |
| 524 | // store for later layers |
| 525 | m_OutputSlotForOperand[inputIndex] = &outputSlot; |
| 526 | } |
| 527 | } |
| 528 | catch (UnsupportedOperand& e) |
| 529 | { |
| 530 | Fail("%s: Operand type %s not supported in ArmnnDriver", __func__, toString(e.m_type).c_str()); |
| 531 | m_ConversionResult = ConversionResult::UnsupportedFeature; |
| 532 | } |
| 533 | catch (const armnn::InvalidArgumentException& e) |
| 534 | { |
| 535 | Fail("%s: Failed to convert input operand to TensorShape: %s", __func__, e.what()); |
| 536 | m_ConversionResult = ConversionResult::UnsupportedFeature; |
| 537 | } |
| 538 | |
| 539 | for (uint32_t operationIdx = 0; operationIdx < m_Model.operations.size(); operationIdx++) |
| 540 | { |
| 541 | const auto& operation = m_Model.operations[operationIdx]; |
| 542 | |
| 543 | bool ok = true; |
| 544 | if (m_ForcedUnsupportedOperations.find(operationIdx) != m_ForcedUnsupportedOperations.end()) |
| 545 | { |
| 546 | Fail("%s: Operation at index %i has been forced to be unsupported.", __func__, operationIdx); |
| 547 | ok = false; |
| 548 | } |
| 549 | |
| 550 | if (ok) |
| 551 | { |
| 552 | try |
| 553 | { |
| 554 | ok = ConvertOperation(operation); |
| 555 | } |
| 556 | catch (UnsupportedOperand& e) |
| 557 | { |
| 558 | Fail("%s: Operand type %s not supported in ArmnnDriver", __func__, toString(e.m_type).c_str()); |
| 559 | ok = false; |
| 560 | } |
| 561 | catch (const armnn::InvalidArgumentException& e) |
| 562 | { |
| 563 | Fail("%s: Failed to convert operation in %s", __func__, e.what()); |
| 564 | ok = false; |
| 565 | } |
| 566 | } |
| 567 | |
| 568 | // Store whether this operation was successfully converted. |
| 569 | m_OperationSupported.emplace(operationIdx, ok); |
| 570 | |
| 571 | // Any single operation failing will fail the entire conversion. |
| 572 | // We still need to continue and check the other ones. |
| 573 | if (!ok) |
| 574 | { |
| 575 | m_ConversionResult = ConversionResult::UnsupportedFeature; |
| 576 | } |
| 577 | } |
| 578 | try |
| 579 | { |
| 580 | if (m_ConversionResult == ConversionResult::Success) |
| 581 | { |
| 582 | for (uint32_t i = 0; i < m_Model.outputIndexes.size(); i++) |
| 583 | { |
| 584 | // outputs in android nn are represented by operands |
| 585 | uint32_t outputIndex = m_Model.outputIndexes[i]; |
| 586 | const Operand& operand = m_Model.operands[outputIndex]; |
| 587 | const armnn::TensorInfo& tensor = GetTensorInfoForOperand(operand); |
| 588 | armnn::IConnectableLayer* layer = m_Network->AddOutputLayer(i); |
| 589 | |
| 590 | assert(m_OutputSlotForOperand[outputIndex]); |
| 591 | m_OutputSlotForOperand[outputIndex]->Connect(layer->GetInputSlot(0)); |
| 592 | } |
| 593 | } |
| 594 | } |
| 595 | catch (const armnn::InvalidArgumentException& e) |
| 596 | { |
| 597 | Fail("%s: Failed to convert output operand to TensorShape: %s", __func__, e.what()); |
| 598 | m_ConversionResult = ConversionResult::UnsupportedFeature; |
| 599 | } |
| 600 | } |
| 601 | |
kevmay01 | bc5f784 | 2018-08-30 12:34:39 +0100 | [diff] [blame^] | 602 | template<typename HalVersion> |
| 603 | bool ModelToINetworkConverter<HalVersion>::ConvertOperation(const neuralnetworks::V1_0::Operation& operation) |
telsoa01 | 5307bc1 | 2018-03-09 13:51:08 +0000 | [diff] [blame] | 604 | { |
| 605 | switch (operation.type) |
| 606 | { |
telsoa01 | ce3e84a | 2018-08-31 09:31:35 +0100 | [diff] [blame] | 607 | case neuralnetworks::V1_0::OperationType::ADD: |
| 608 | return ConvertAdd(operation); |
| 609 | case neuralnetworks::V1_0::OperationType::AVERAGE_POOL_2D: |
| 610 | return ConvertAveragePool2d(operation); |
| 611 | case neuralnetworks::V1_0::OperationType::CONCATENATION: |
| 612 | return ConvertConcatenation(operation); |
| 613 | case neuralnetworks::V1_0::OperationType::CONV_2D: |
| 614 | return ConvertConv2d(operation); |
| 615 | case neuralnetworks::V1_0::OperationType::DEPTHWISE_CONV_2D: |
| 616 | return ConvertDepthwiseConv2d(operation); |
| 617 | case neuralnetworks::V1_0::OperationType::FLOOR: |
| 618 | return ConvertFloor(operation); |
| 619 | case neuralnetworks::V1_0::OperationType::FULLY_CONNECTED: |
| 620 | return ConvertFullyConnected(operation); |
| 621 | case neuralnetworks::V1_0::OperationType::LOCAL_RESPONSE_NORMALIZATION: |
| 622 | return ConvertLocalResponseNormalization(operation); |
| 623 | case neuralnetworks::V1_0::OperationType::LOGISTIC: |
| 624 | return ConvertLogistic(operation); |
| 625 | case neuralnetworks::V1_0::OperationType::LSTM: |
| 626 | return ConvertLstm(operation); |
| 627 | case neuralnetworks::V1_0::OperationType::L2_NORMALIZATION: |
| 628 | return ConvertL2Normalization(operation); |
| 629 | case neuralnetworks::V1_0::OperationType::L2_POOL_2D: |
| 630 | return ConvertL2Pool2d(operation); |
| 631 | case neuralnetworks::V1_0::OperationType::MAX_POOL_2D: |
| 632 | return ConvertMaxPool2d(operation); |
| 633 | case neuralnetworks::V1_0::OperationType::MUL: |
| 634 | return ConvertMul(operation); |
| 635 | case neuralnetworks::V1_0::OperationType::RELU: |
| 636 | return ConvertReLu(operation); |
| 637 | case neuralnetworks::V1_0::OperationType::RELU1: |
| 638 | return ConvertReLu1(operation); |
| 639 | case neuralnetworks::V1_0::OperationType::RELU6: |
| 640 | return ConvertReLu6(operation); |
| 641 | case neuralnetworks::V1_0::OperationType::SOFTMAX: |
| 642 | return ConvertSoftmax(operation); |
| 643 | case neuralnetworks::V1_0::OperationType::TANH: |
| 644 | return ConvertTanH(operation); |
| 645 | case neuralnetworks::V1_0::OperationType::RESHAPE: |
| 646 | return ConvertReshape(operation); |
| 647 | case neuralnetworks::V1_0::OperationType::RESIZE_BILINEAR: |
| 648 | return ConvertResizeBilinear(operation); |
| 649 | default: |
| 650 | return Fail("%s: Operation type %s not supported in ArmnnDriver", |
| 651 | __func__, toString(operation.type).c_str()); |
telsoa01 | 5307bc1 | 2018-03-09 13:51:08 +0000 | [diff] [blame] | 652 | } |
| 653 | } |
| 654 | |
kevmay01 | bc5f784 | 2018-08-30 12:34:39 +0100 | [diff] [blame^] | 655 | #if defined(ARMNN_ANDROID_NN_V1_1) |
| 656 | template<typename HalVersion> |
| 657 | bool ModelToINetworkConverter<HalVersion>::ConvertOperation(const neuralnetworks::V1_1::Operation& operation) |
| 658 | { |
| 659 | if (compliantWithV1_0(operation)) |
| 660 | { |
| 661 | neuralnetworks::V1_0::Operation v1Operation = convertToV1_0(operation); |
| 662 | return ConvertOperation(v1Operation); |
| 663 | } |
| 664 | else |
| 665 | { |
| 666 | switch (operation.type) |
| 667 | { |
| 668 | // TODO: provide cases for converting the new v1.1 operations |
| 669 | default: |
| 670 | return Fail("%s: Operation type %s not supported in ArmnnDriver", |
| 671 | __func__, toString(operation.type).c_str()); |
| 672 | } |
| 673 | } |
| 674 | } |
| 675 | #endif |
| 676 | |
| 677 | template<typename HalVersion> |
| 678 | bool ModelToINetworkConverter<HalVersion>::ConvertAdd(const neuralnetworks::V1_0::Operation& operation) |
telsoa01 | 5307bc1 | 2018-03-09 13:51:08 +0000 | [diff] [blame] | 679 | { |
| 680 | LayerInputHandle input0 = ConvertToLayerInputHandle(operation, 0); |
| 681 | LayerInputHandle input1 = ConvertToLayerInputHandle(operation, 1); |
| 682 | |
| 683 | if (!input0.IsValid() || !input1.IsValid()) |
| 684 | { |
| 685 | return Fail("%s: Operation has invalid inputs", __func__); |
| 686 | } |
| 687 | |
telsoa01 | ce3e84a | 2018-08-31 09:31:35 +0100 | [diff] [blame] | 688 | // The FuseActivation parameter is always the input index 2 |
| 689 | // and it should be optional |
telsoa01 | 5307bc1 | 2018-03-09 13:51:08 +0000 | [diff] [blame] | 690 | ActivationFn activationFunction; |
telsoa01 | ce3e84a | 2018-08-31 09:31:35 +0100 | [diff] [blame] | 691 | if (!GetOptionalInputActivation(operation, 2, activationFunction)) |
telsoa01 | 5307bc1 | 2018-03-09 13:51:08 +0000 | [diff] [blame] | 692 | { |
| 693 | return Fail("%s: Operation has invalid inputs", __func__); |
| 694 | } |
| 695 | |
| 696 | const Operand* outputOperand = GetOutputOperand(operation, 0); |
| 697 | if (!outputOperand) |
| 698 | { |
| 699 | return false; |
| 700 | } |
| 701 | |
| 702 | const armnn::TensorInfo outInfo = GetTensorInfoForOperand(*outputOperand); |
| 703 | |
| 704 | if (!IsLayerSupported(__func__, |
| 705 | armnn::IsAdditionSupported, |
| 706 | m_Compute, |
| 707 | input0.GetTensorInfo(), |
| 708 | input1.GetTensorInfo(), |
| 709 | outInfo)) |
| 710 | { |
| 711 | return false; |
| 712 | } |
| 713 | |
| 714 | armnn::IConnectableLayer* const startLayer = m_Network->AddAdditionLayer(); |
| 715 | armnn::IConnectableLayer* const endLayer = ProcessActivation(outInfo, activationFunction, startLayer); |
| 716 | |
| 717 | const armnn::TensorInfo& inputTensorInfo0 = input0.GetTensorInfo(); |
| 718 | const armnn::TensorInfo& inputTensorInfo1 = input1.GetTensorInfo(); |
| 719 | |
| 720 | if (endLayer != nullptr) |
| 721 | { |
telsoa01 | ce3e84a | 2018-08-31 09:31:35 +0100 | [diff] [blame] | 722 | BroadcastTensor(input0, input1, startLayer, *m_Network); |
telsoa01 | 5307bc1 | 2018-03-09 13:51:08 +0000 | [diff] [blame] | 723 | return SetupAndTrackLayerOutputSlot(operation, 0, *endLayer); |
| 724 | } |
| 725 | else |
| 726 | { |
| 727 | return Fail("%s: ProcessActivation failed", __func__); |
| 728 | } |
| 729 | } |
| 730 | |
kevmay01 | bc5f784 | 2018-08-30 12:34:39 +0100 | [diff] [blame^] | 731 | template<typename HalVersion> |
| 732 | bool ModelToINetworkConverter<HalVersion>::ConvertAveragePool2d(const neuralnetworks::V1_0::Operation& operation) |
telsoa01 | 5307bc1 | 2018-03-09 13:51:08 +0000 | [diff] [blame] | 733 | { |
| 734 | return ConvertPooling2d(operation, __func__, armnn::PoolingAlgorithm::Average); |
| 735 | } |
| 736 | |
kevmay01 | bc5f784 | 2018-08-30 12:34:39 +0100 | [diff] [blame^] | 737 | template<typename HalVersion> |
| 738 | bool ModelToINetworkConverter<HalVersion>::ConvertConcatenation(const neuralnetworks::V1_0::Operation& operation) |
telsoa01 | 5307bc1 | 2018-03-09 13:51:08 +0000 | [diff] [blame] | 739 | { |
| 740 | // The first N (0..N-1) inputs are tensors. The Nth input is the concatenation axis. |
| 741 | if (operation.inputs.size() <= 1) |
| 742 | { |
| 743 | return Fail("%s: Operation has insufficient arguments", __func__); |
| 744 | } |
| 745 | |
| 746 | // Get inputs and outputs |
| 747 | const std::size_t numInputTensors = operation.inputs.size() - 1; |
| 748 | |
surmeh01 | 49b9e10 | 2018-05-17 14:11:25 +0100 | [diff] [blame] | 749 | int32_t concatDim; |
| 750 | if (!GetInputScalar(operation, numInputTensors, OperandType::INT32, concatDim)) |
| 751 | { |
| 752 | return Fail("%s: Operation has invalid inputs", __func__); |
| 753 | } |
| 754 | |
| 755 | const Operand* const outputOperand = GetOutputOperand(operation, 0); |
| 756 | if (!outputOperand) |
| 757 | { |
| 758 | return Fail("%s: Operation has no outputs", __func__); |
| 759 | } |
| 760 | |
telsoa01 | ce3e84a | 2018-08-31 09:31:35 +0100 | [diff] [blame] | 761 | |
surmeh01 | 49b9e10 | 2018-05-17 14:11:25 +0100 | [diff] [blame] | 762 | armnn::TensorInfo outputInfo = GetTensorInfoForOperand(*outputOperand); |
| 763 | armnn::TensorShape outputShape = outputInfo.GetShape(); |
| 764 | |
| 765 | // |
| 766 | // handle negative concat dims along the lines of tensorflow as described here: |
| 767 | // https://www.tensorflow.org/api_docs/python/tf/concat |
| 768 | // "negative axis refers to axis + rank(values)-th dimension" |
| 769 | // |
| 770 | if (concatDim < 0) |
| 771 | { |
| 772 | concatDim += outputShape.GetNumDimensions(); |
| 773 | } |
| 774 | |
| 775 | if (concatDim >= static_cast<int32_t>(outputShape.GetNumDimensions()) || concatDim < 0) |
| 776 | { |
| 777 | return Fail("%s: Operation has invalid concat axis: %d", __func__, concatDim); |
| 778 | } |
| 779 | |
telsoa01 | 5307bc1 | 2018-03-09 13:51:08 +0000 | [diff] [blame] | 780 | std::vector<LayerInputHandle> inputHandles; |
| 781 | std::vector<armnn::TensorShape> inputShapes; |
| 782 | |
| 783 | inputHandles.reserve(numInputTensors); |
| 784 | inputShapes.reserve(numInputTensors); |
| 785 | |
telsoa01 | ce3e84a | 2018-08-31 09:31:35 +0100 | [diff] [blame] | 786 | bool inputsHaveBeenReshaped = false; |
| 787 | unsigned int tensorDimensionsAdded = 0; |
| 788 | |
telsoa01 | 5307bc1 | 2018-03-09 13:51:08 +0000 | [diff] [blame] | 789 | for (uint32_t i = 0; i < numInputTensors; ++i) |
| 790 | { |
| 791 | const Operand* const operand = GetInputOperand(operation, i); |
| 792 | if (!operand) |
| 793 | { |
| 794 | return Fail("%s: Operation has invalid inputs", __func__); |
| 795 | } |
| 796 | |
telsoa01 | ce3e84a | 2018-08-31 09:31:35 +0100 | [diff] [blame] | 797 | armnn::TensorShape operandShape = GetTensorShapeForOperand(*operand); |
| 798 | LayerInputHandle operandInputHandle = ConvertToLayerInputHandle(operation, i); |
surmeh01 | 49b9e10 | 2018-05-17 14:11:25 +0100 | [diff] [blame] | 799 | |
telsoa01 | ce3e84a | 2018-08-31 09:31:35 +0100 | [diff] [blame] | 800 | if (operandShape.GetNumDimensions() == 0) |
| 801 | { |
| 802 | return Fail("%s: Operands with rank 0 are not supported", __func__); |
| 803 | } |
| 804 | |
| 805 | if (RequiresReshape(operandShape)) |
| 806 | { |
| 807 | inputsHaveBeenReshaped = true; |
| 808 | |
| 809 | armnn::TensorInfo reshapeInfo = operandInputHandle.GetTensorInfo(); |
| 810 | |
| 811 | // Expand the tensor to three dimensions |
| 812 | if (operandShape.GetNumDimensions() == 2) |
| 813 | { |
| 814 | reshapeInfo.SetShape(armnn::TensorShape({1, operandShape[0], operandShape[1]})); |
| 815 | tensorDimensionsAdded = 1; |
| 816 | } |
| 817 | else |
| 818 | { |
| 819 | reshapeInfo.SetShape(armnn::TensorShape({1, 1, operandShape[0]})); |
| 820 | tensorDimensionsAdded = 2; |
| 821 | } |
| 822 | |
| 823 | armnn::IConnectableLayer& newReshape = AddReshapeLayer( |
| 824 | *m_Network, |
| 825 | operandInputHandle, |
| 826 | reshapeInfo |
| 827 | ); |
| 828 | |
| 829 | // Point to the reshape operation rather then the input operation |
| 830 | operandShape = reshapeInfo.GetShape(); |
| 831 | operandInputHandle = LayerInputHandle(true, &newReshape.GetOutputSlot(0), reshapeInfo); |
| 832 | } |
| 833 | |
| 834 | inputShapes.emplace_back(operandShape); |
| 835 | inputHandles.emplace_back(operandInputHandle); |
surmeh01 | 49b9e10 | 2018-05-17 14:11:25 +0100 | [diff] [blame] | 836 | |
telsoa01 | 5307bc1 | 2018-03-09 13:51:08 +0000 | [diff] [blame] | 837 | if (!inputHandles.back().IsValid()) |
| 838 | { |
| 839 | return Fail("%s: Operation has invalid inputs", __func__); |
| 840 | } |
| 841 | } |
| 842 | |
| 843 | assert(inputShapes.size() == inputHandles.size()); |
| 844 | |
telsoa01 | ce3e84a | 2018-08-31 09:31:35 +0100 | [diff] [blame] | 845 | if (inputsHaveBeenReshaped) |
| 846 | { |
| 847 | // Adjust the concatenation dimension by the amount of dimensions added (if any) |
| 848 | concatDim += tensorDimensionsAdded; |
| 849 | |
| 850 | // Add extra dimensions to the output shape to reflect the addition of the reshape layers |
| 851 | if (tensorDimensionsAdded == 1) |
| 852 | { |
| 853 | outputShape = armnn::TensorShape({1, outputShape[0], outputShape[1]}); |
| 854 | } |
| 855 | else if (tensorDimensionsAdded == 2) |
| 856 | { |
| 857 | outputShape = armnn::TensorShape({1, 1, outputShape[0], outputShape[1]}); |
| 858 | } |
| 859 | } |
| 860 | |
| 861 | // Get the pair of permutations required for the concatenation |
| 862 | std::pair<armnn::PermutationVector, armnn::PermutationVector> permutationPair = |
| 863 | std::make_pair(IdentityPermutation4D, IdentityPermutation4D); |
| 864 | |
| 865 | CreatePermutationParameters(inputShapes[0].GetNumDimensions(), concatDim, permutationPair); |
| 866 | |
| 867 | outputShape = armnnUtils::Permuted(outputShape, permutationPair.first); |
| 868 | outputInfo.SetShape(outputShape); |
| 869 | |
surmeh01 | 49b9e10 | 2018-05-17 14:11:25 +0100 | [diff] [blame] | 870 | // this is no-op for identity swizzles, otherwise it replaces both |
| 871 | // the handles and shapes with the swizzled layer output handles and shapes |
telsoa01 | ce3e84a | 2018-08-31 09:31:35 +0100 | [diff] [blame] | 872 | SwizzleInputs(*m_Network, inputHandles, inputShapes, permutationPair.first); |
telsoa01 | 5307bc1 | 2018-03-09 13:51:08 +0000 | [diff] [blame] | 873 | |
| 874 | // Create an armnn merger layer descriptor - this will also perform validation on the input shapes |
| 875 | armnn::OriginsDescriptor mergerDescriptor; |
| 876 | try |
| 877 | { |
surmeh01 | 49b9e10 | 2018-05-17 14:11:25 +0100 | [diff] [blame] | 878 | // The merger descriptor is always created across the only supported concat |
| 879 | // dimension, which is 0 or 1 |
| 880 | mergerDescriptor = |
| 881 | armnn::CreateMergerDescriptorForConcatenation( |
| 882 | inputShapes.begin(), inputShapes.end(), concatDim); |
telsoa01 | 5307bc1 | 2018-03-09 13:51:08 +0000 | [diff] [blame] | 883 | } |
| 884 | catch (const armnn::Exception& error) |
| 885 | { |
| 886 | return Fail("%s: Error preparing merger descriptor. %s", __func__, error.what()); |
| 887 | } |
| 888 | |
surmeh01 | 49b9e10 | 2018-05-17 14:11:25 +0100 | [diff] [blame] | 889 | // Validate the output shape is correct given the input shapes based on the |
| 890 | // only valid concat dimension which is 0 or 1 |
| 891 | if (!ValidateConcatOutputShape(inputShapes, outputShape, concatDim)) |
telsoa01 | 5307bc1 | 2018-03-09 13:51:08 +0000 | [diff] [blame] | 892 | { |
surmeh01 | 49b9e10 | 2018-05-17 14:11:25 +0100 | [diff] [blame] | 893 | return Fail("%s: Error validating the output shape for concat", __func__); |
telsoa01 | 5307bc1 | 2018-03-09 13:51:08 +0000 | [diff] [blame] | 894 | } |
| 895 | |
| 896 | std::vector<const armnn::TensorInfo*> inputTensorInfos; |
| 897 | std::transform(inputHandles.begin(), inputHandles.end(), std::back_inserter(inputTensorInfos), |
| 898 | [](const LayerInputHandle& h) -> const armnn::TensorInfo*{ return &h.GetTensorInfo(); }); |
| 899 | if (!IsLayerSupported(__func__, |
| 900 | armnn::IsMergerSupported, |
| 901 | m_Compute, |
| 902 | inputTensorInfos, |
| 903 | mergerDescriptor)) |
| 904 | { |
| 905 | return false; |
| 906 | } |
| 907 | |
| 908 | armnn::IConnectableLayer* layer = m_Network->AddMergerLayer(mergerDescriptor); |
| 909 | assert(layer != nullptr); |
surmeh01 | 49b9e10 | 2018-05-17 14:11:25 +0100 | [diff] [blame] | 910 | layer->GetOutputSlot(0).SetTensorInfo(outputInfo); |
telsoa01 | 5307bc1 | 2018-03-09 13:51:08 +0000 | [diff] [blame] | 911 | |
| 912 | // Connect inputs to the layer |
| 913 | const int numInputSlots = layer->GetNumInputSlots(); |
| 914 | assert(static_cast<std::size_t>(numInputSlots) == inputHandles.size()); |
| 915 | for (int i = 0; i < numInputSlots; ++i) |
| 916 | { |
surmeh01 | 49b9e10 | 2018-05-17 14:11:25 +0100 | [diff] [blame] | 917 | // connect the input directly to the merge (concat) layer |
telsoa01 | 5307bc1 | 2018-03-09 13:51:08 +0000 | [diff] [blame] | 918 | inputHandles[static_cast<unsigned int>(i)].Connect(layer->GetInputSlot(i)); |
| 919 | } |
| 920 | |
telsoa01 | ce3e84a | 2018-08-31 09:31:35 +0100 | [diff] [blame] | 921 | // Add permutation layer and connect the output to it, the permutation becomes the output layer |
| 922 | armnn::IConnectableLayer& deswizzleLayer = AddPermuteLayer(*m_Network, |
| 923 | layer->GetOutputSlot(0), |
| 924 | permutationPair.second); |
| 925 | layer = &deswizzleLayer; |
| 926 | |
| 927 | if (inputsHaveBeenReshaped) |
surmeh01 | 49b9e10 | 2018-05-17 14:11:25 +0100 | [diff] [blame] | 928 | { |
telsoa01 | ce3e84a | 2018-08-31 09:31:35 +0100 | [diff] [blame] | 929 | armnn::TensorInfo afterConcatInfo = layer->GetOutputSlot(0).GetTensorInfo(); |
| 930 | |
| 931 | // Undo the reshape knowing the amount of dimensions added |
| 932 | if (tensorDimensionsAdded == 1) |
| 933 | { |
| 934 | afterConcatInfo.SetShape(armnn::TensorShape({ afterConcatInfo.GetShape()[1], |
| 935 | afterConcatInfo.GetShape()[2] })); |
| 936 | } |
| 937 | else if (tensorDimensionsAdded == 2) |
| 938 | { |
| 939 | afterConcatInfo.SetShape(armnn::TensorShape({ afterConcatInfo.GetShape()[2], |
| 940 | afterConcatInfo.GetShape()[3] })); |
| 941 | } |
| 942 | |
| 943 | layer = &AddReshapeLayer( |
| 944 | *m_Network, |
| 945 | layer->GetOutputSlot(0), |
| 946 | afterConcatInfo |
| 947 | ); |
surmeh01 | 49b9e10 | 2018-05-17 14:11:25 +0100 | [diff] [blame] | 948 | } |
| 949 | |
telsoa01 | 5307bc1 | 2018-03-09 13:51:08 +0000 | [diff] [blame] | 950 | return SetupAndTrackLayerOutputSlot(operation, 0, *layer); |
| 951 | } |
| 952 | |
kevmay01 | bc5f784 | 2018-08-30 12:34:39 +0100 | [diff] [blame^] | 953 | template<typename HalVersion> |
| 954 | bool ModelToINetworkConverter<HalVersion>::ConvertConv2d(const neuralnetworks::V1_0::Operation& operation) |
telsoa01 | 5307bc1 | 2018-03-09 13:51:08 +0000 | [diff] [blame] | 955 | { |
| 956 | LayerInputHandle input = ConvertToLayerInputHandle(operation, 0); |
| 957 | if (!input.IsValid()) |
| 958 | { |
| 959 | return Fail("%s: Operation has invalid inputs", __func__); |
| 960 | } |
| 961 | |
| 962 | const Operand* output = GetOutputOperand(operation, 0); |
| 963 | if (!output) |
| 964 | { |
| 965 | return Fail("%s: Could not read output 0", __func__); |
| 966 | } |
| 967 | |
| 968 | const armnn::TensorInfo& inputInfo = input.GetTensorInfo(); |
| 969 | const armnn::TensorInfo& outputInfo = GetTensorInfoForOperand(*output); |
| 970 | |
| 971 | const armnn::TensorInfo swizzledInputInfo = armnnUtils::Permuted(inputInfo, NHWCToArmNN); |
| 972 | const armnn::TensorInfo swizzledOutputInfo = armnnUtils::Permuted(outputInfo, NHWCToArmNN); |
| 973 | |
| 974 | // ArmNN does not currently support non-fixed weights or bias |
| 975 | const ConstTensorPin weightsPin = ConvertOperationInputToConstTensorPin(operation, 1, NHWCToArmNN); |
| 976 | const ConstTensorPin biasPin = ConvertOperationInputToConstTensorPin(operation, 2); |
| 977 | |
| 978 | if (!weightsPin.IsValid() || !biasPin.IsValid()) |
| 979 | { |
| 980 | return Fail("%s: Operation has invalid inputs", __func__); |
| 981 | } |
| 982 | |
| 983 | armnn::ConstTensor weights = weightsPin.GetConstTensor(); |
| 984 | armnn::ConstTensor bias = biasPin.GetConstTensor(); |
| 985 | SanitizeBiasQuantizationScale(bias.GetInfo(), weights.GetInfo(), swizzledInputInfo); |
| 986 | |
| 987 | armnn::Convolution2dDescriptor desc; |
| 988 | ActivationFn activation; |
| 989 | |
| 990 | if (operation.inputs.size() == 10) |
| 991 | { |
| 992 | if (!GetInputScalar(operation, 3, OperandType::INT32, desc.m_PadLeft) || |
| 993 | !GetInputScalar(operation, 4, OperandType::INT32, desc.m_PadRight) || |
| 994 | !GetInputScalar(operation, 5, OperandType::INT32, desc.m_PadTop) || |
| 995 | !GetInputScalar(operation, 6, OperandType::INT32, desc.m_PadBottom) || |
| 996 | !GetInputScalar(operation, 7, OperandType::INT32, desc.m_StrideX) || |
| 997 | !GetInputScalar(operation, 8, OperandType::INT32, desc.m_StrideY) || |
| 998 | !GetInputActivationFunction(operation, 9, activation)) |
| 999 | { |
| 1000 | return Fail("%s: Operation has invalid inputs", __func__); |
| 1001 | } |
| 1002 | } |
| 1003 | else if (operation.inputs.size() == 7) |
| 1004 | { |
| 1005 | android::nn::PaddingScheme paddingScheme; |
| 1006 | |
| 1007 | if (!GetInputPaddingScheme(operation, 3, paddingScheme) || |
| 1008 | !GetInputScalar(operation, 4, OperandType::INT32, desc.m_StrideX) || |
| 1009 | !GetInputScalar(operation, 5, OperandType::INT32, desc.m_StrideY) || |
| 1010 | !GetInputActivationFunction(operation, 6, activation)) |
| 1011 | { |
| 1012 | return Fail("%s: Operation has invalid inputs", __func__); |
| 1013 | } |
| 1014 | |
| 1015 | const uint32_t kernelX = weights.GetShape()[3]; |
| 1016 | const uint32_t kernelY = weights.GetShape()[2]; |
| 1017 | const uint32_t inputX = swizzledInputInfo.GetShape()[3]; |
| 1018 | const uint32_t inputY = swizzledInputInfo.GetShape()[2]; |
| 1019 | |
| 1020 | CalcPadding(inputX, kernelX, desc.m_StrideX, desc.m_PadLeft, desc.m_PadRight, paddingScheme); |
| 1021 | CalcPadding(inputY, kernelY, desc.m_StrideY, desc.m_PadTop, desc.m_PadBottom, paddingScheme); |
| 1022 | } |
| 1023 | else |
| 1024 | { |
| 1025 | return Fail("%s: Unsupported number of operation inputs", __func__); |
| 1026 | } |
| 1027 | |
arovir01 | 3b0a2da | 2018-08-29 10:16:58 +0100 | [diff] [blame] | 1028 | desc.m_BiasEnabled = true; |
| 1029 | auto biases = boost::make_optional(bias.GetInfo()); |
telsoa01 | 5307bc1 | 2018-03-09 13:51:08 +0000 | [diff] [blame] | 1030 | |
| 1031 | if (!IsLayerSupported(__func__, |
| 1032 | armnn::IsConvolution2dSupported, |
| 1033 | m_Compute, |
| 1034 | swizzledInputInfo, |
surmeh01 | 49b9e10 | 2018-05-17 14:11:25 +0100 | [diff] [blame] | 1035 | swizzledOutputInfo, |
telsoa01 | 5307bc1 | 2018-03-09 13:51:08 +0000 | [diff] [blame] | 1036 | desc, |
surmeh01 | 49b9e10 | 2018-05-17 14:11:25 +0100 | [diff] [blame] | 1037 | weights.GetInfo(), |
arovir01 | 3b0a2da | 2018-08-29 10:16:58 +0100 | [diff] [blame] | 1038 | biases)) |
telsoa01 | 5307bc1 | 2018-03-09 13:51:08 +0000 | [diff] [blame] | 1039 | { |
| 1040 | return false; |
| 1041 | } |
| 1042 | |
| 1043 | armnn::IConnectableLayer* startLayer = m_Network->AddConvolution2dLayer(desc, weights, bias); |
| 1044 | armnn::IConnectableLayer* endLayer = ProcessActivation(swizzledOutputInfo, activation, startLayer); |
| 1045 | |
| 1046 | if (endLayer != nullptr) |
| 1047 | { |
| 1048 | armnn::IConnectableLayer& outSwizzleLayer = SwizzleInDeswizzleOut(*m_Network, input, *startLayer, *endLayer); |
| 1049 | return SetupAndTrackLayerOutputSlot(operation, 0, outSwizzleLayer); |
| 1050 | } |
| 1051 | else |
| 1052 | { |
| 1053 | return Fail("%s: ProcessActivation failed", __func__); |
| 1054 | } |
| 1055 | } |
| 1056 | |
kevmay01 | bc5f784 | 2018-08-30 12:34:39 +0100 | [diff] [blame^] | 1057 | template<typename HalVersion> |
| 1058 | bool ModelToINetworkConverter<HalVersion>::ConvertDepthwiseConv2d(const neuralnetworks::V1_0::Operation& operation) |
telsoa01 | 5307bc1 | 2018-03-09 13:51:08 +0000 | [diff] [blame] | 1059 | { |
| 1060 | LayerInputHandle input = ConvertToLayerInputHandle(operation, 0); |
| 1061 | if (!input.IsValid()) |
| 1062 | { |
| 1063 | return Fail("%s: Operation has invalid inputs", __func__); |
| 1064 | } |
| 1065 | |
| 1066 | const Operand* output = GetOutputOperand(operation, 0); |
| 1067 | if (!output) |
| 1068 | { |
| 1069 | return Fail("%s: Could not read output 0", __func__); |
| 1070 | } |
| 1071 | |
| 1072 | const armnn::TensorInfo& inputInfo = input.GetTensorInfo(); |
| 1073 | const armnn::TensorInfo& outputInfo = GetTensorInfoForOperand(*output); |
| 1074 | |
| 1075 | const armnn::TensorInfo swizzledInputInfo = armnnUtils::Permuted(inputInfo, NHWCToArmNN); |
| 1076 | const armnn::TensorInfo swizzledOutputInfo = armnnUtils::Permuted(outputInfo, NHWCToArmNN); |
| 1077 | |
| 1078 | // ArmNN does not currently support non-fixed weights or bias |
| 1079 | |
| 1080 | // Find the shape of the weights tensor. In AndroidNN this will be [ 1, H, W, I * M ] |
| 1081 | // but in ArmNN it needs to be [ M, I, H, W ] |
| 1082 | const Operand* weightsOperand = GetInputOperand(operation, 1); |
| 1083 | |
| 1084 | if (weightsOperand == nullptr) |
| 1085 | { |
| 1086 | return Fail("%s: Operand is invalid", __func__); |
| 1087 | } |
| 1088 | |
| 1089 | // Reinterpret weight data as [ H, W, I, M ] |
| 1090 | armnn::TensorShape weightsShape({ weightsOperand->dimensions[1], weightsOperand->dimensions[2], |
| 1091 | inputInfo.GetShape()[3], |
| 1092 | weightsOperand->dimensions[3] / inputInfo.GetShape()[3] }); |
| 1093 | |
| 1094 | // Swizzle weight data [ H, W, I, M ] -> [ M, I, H, W ] |
| 1095 | const armnn::PermutationVector HWIMToMIHW = { 2U, 3U, 1U, 0U }; |
| 1096 | ConstTensorPin weightsPin = ConvertOperationInputToConstTensorPin(operation, 1, HWIMToMIHW, &weightsShape); |
| 1097 | |
| 1098 | // Bias is a 1D tensor |
| 1099 | ConstTensorPin biasPin = ConvertOperationInputToConstTensorPin(operation, 2); |
| 1100 | |
| 1101 | if (!weightsPin.IsValid() || !biasPin.IsValid()) |
| 1102 | { |
| 1103 | return Fail("%s: Operation has invalid inputs", __func__); |
| 1104 | } |
| 1105 | |
| 1106 | armnn::ConstTensor weights = weightsPin.GetConstTensor(); |
| 1107 | armnn::ConstTensor bias = biasPin.GetConstTensor(); |
| 1108 | SanitizeBiasQuantizationScale(bias.GetInfo(), weights.GetInfo(), swizzledInputInfo); |
| 1109 | |
| 1110 | armnn::DepthwiseConvolution2dDescriptor desc; |
| 1111 | ActivationFn activation; |
| 1112 | |
| 1113 | if (operation.inputs.size() == 11) |
| 1114 | { |
| 1115 | if (!GetInputScalar(operation, 3, OperandType::INT32, desc.m_PadLeft) || |
| 1116 | !GetInputScalar(operation, 4, OperandType::INT32, desc.m_PadRight) || |
| 1117 | !GetInputScalar(operation, 5, OperandType::INT32, desc.m_PadTop) || |
| 1118 | !GetInputScalar(operation, 6, OperandType::INT32, desc.m_PadBottom) || |
| 1119 | !GetInputScalar(operation, 7, OperandType::INT32, desc.m_StrideX) || |
| 1120 | !GetInputScalar(operation, 8, OperandType::INT32, desc.m_StrideY) || |
| 1121 | !GetInputActivationFunction(operation, 10, activation)) |
| 1122 | { |
| 1123 | return Fail("%s: Operation has invalid inputs", __func__); |
| 1124 | } |
| 1125 | } |
| 1126 | else if (operation.inputs.size() == 8) |
| 1127 | { |
| 1128 | android::nn::PaddingScheme paddingScheme; |
| 1129 | |
| 1130 | if (!GetInputPaddingScheme(operation, 3, paddingScheme) || |
| 1131 | !GetInputScalar(operation, 4, OperandType::INT32, desc.m_StrideX) || |
| 1132 | !GetInputScalar(operation, 5, OperandType::INT32, desc.m_StrideY) || |
| 1133 | !GetInputActivationFunction(operation, 7, activation)) |
| 1134 | { |
| 1135 | return Fail("%s: Operation has invalid inputs", __func__); |
| 1136 | } |
| 1137 | |
| 1138 | const uint32_t kernelX = weights.GetShape()[3]; |
| 1139 | const uint32_t kernelY = weights.GetShape()[2]; |
| 1140 | const uint32_t inputX = swizzledInputInfo.GetShape()[3]; |
| 1141 | const uint32_t inputY = swizzledInputInfo.GetShape()[2]; |
| 1142 | |
| 1143 | CalcPadding(inputX, kernelX, desc.m_StrideX, desc.m_PadLeft, desc.m_PadRight, paddingScheme); |
| 1144 | CalcPadding(inputY, kernelY, desc.m_StrideY, desc.m_PadTop, desc.m_PadBottom, paddingScheme); |
| 1145 | } |
| 1146 | else |
| 1147 | { |
| 1148 | return Fail("%s: Unsupported number of operation inputs", __func__); |
| 1149 | } |
| 1150 | |
| 1151 | desc.m_BiasEnabled = true; |
arovir01 | 3b0a2da | 2018-08-29 10:16:58 +0100 | [diff] [blame] | 1152 | auto biases = boost::make_optional(bias.GetInfo()); |
telsoa01 | 5307bc1 | 2018-03-09 13:51:08 +0000 | [diff] [blame] | 1153 | |
| 1154 | if (!IsLayerSupported(__func__, |
| 1155 | armnn::IsDepthwiseConvolutionSupported, |
| 1156 | m_Compute, |
| 1157 | swizzledInputInfo, |
telsoa01 | ce3e84a | 2018-08-31 09:31:35 +0100 | [diff] [blame] | 1158 | swizzledOutputInfo, |
telsoa01 | 5307bc1 | 2018-03-09 13:51:08 +0000 | [diff] [blame] | 1159 | desc, |
telsoa01 | ce3e84a | 2018-08-31 09:31:35 +0100 | [diff] [blame] | 1160 | weights.GetInfo(), |
arovir01 | 3b0a2da | 2018-08-29 10:16:58 +0100 | [diff] [blame] | 1161 | biases)) |
telsoa01 | 5307bc1 | 2018-03-09 13:51:08 +0000 | [diff] [blame] | 1162 | { |
| 1163 | return false; |
| 1164 | } |
| 1165 | |
| 1166 | armnn::IConnectableLayer* startLayer = m_Network->AddDepthwiseConvolution2dLayer(desc, weights, bias); |
| 1167 | armnn::IConnectableLayer* endLayer = ProcessActivation(swizzledOutputInfo, activation, startLayer); |
| 1168 | |
| 1169 | if (endLayer != nullptr) |
| 1170 | { |
| 1171 | armnn::IConnectableLayer& outSwizzleLayer = SwizzleInDeswizzleOut(*m_Network, input, *startLayer, *endLayer); |
| 1172 | return SetupAndTrackLayerOutputSlot(operation, 0, outSwizzleLayer); |
| 1173 | } |
| 1174 | else |
| 1175 | { |
| 1176 | return Fail("%s: ProcessActivation failed", __func__); |
| 1177 | } |
| 1178 | } |
| 1179 | |
kevmay01 | bc5f784 | 2018-08-30 12:34:39 +0100 | [diff] [blame^] | 1180 | template<typename HalVersion> |
| 1181 | bool ModelToINetworkConverter<HalVersion>::ConvertFloor(const neuralnetworks::V1_0::Operation& operation) |
telsoa01 | 5307bc1 | 2018-03-09 13:51:08 +0000 | [diff] [blame] | 1182 | { |
| 1183 | LayerInputHandle input = ConvertToLayerInputHandle(operation, 0); |
| 1184 | if (!input.IsValid()) |
| 1185 | { |
| 1186 | return Fail("%s: Operation has invalid inputs", __func__); |
| 1187 | } |
| 1188 | |
| 1189 | const Operand* const outputOperand = GetOutputOperand(operation, 0); |
| 1190 | if (!outputOperand) |
| 1191 | { |
| 1192 | return Fail("%s: Operation has invalid outputs", __func__); |
| 1193 | } |
| 1194 | |
| 1195 | if (!IsLayerSupported(__func__, |
| 1196 | armnn::IsFloorSupported, |
| 1197 | m_Compute, |
| 1198 | input.GetTensorInfo(), |
| 1199 | GetTensorInfoForOperand(*outputOperand))) |
| 1200 | { |
| 1201 | return false; |
| 1202 | } |
| 1203 | |
| 1204 | armnn::IConnectableLayer* layer = m_Network->AddFloorLayer(); |
| 1205 | assert(layer != nullptr); |
| 1206 | input.Connect(layer->GetInputSlot(0)); |
| 1207 | |
| 1208 | return SetupAndTrackLayerOutputSlot(operation, 0, *layer); |
| 1209 | } |
| 1210 | |
kevmay01 | bc5f784 | 2018-08-30 12:34:39 +0100 | [diff] [blame^] | 1211 | template<typename HalVersion> |
| 1212 | bool ModelToINetworkConverter<HalVersion>::ConvertFullyConnected(const neuralnetworks::V1_0::Operation& operation) |
telsoa01 | 5307bc1 | 2018-03-09 13:51:08 +0000 | [diff] [blame] | 1213 | { |
| 1214 | LayerInputHandle input = ConvertToLayerInputHandle(operation, 0); |
| 1215 | if (!input.IsValid()) |
| 1216 | { |
| 1217 | return Fail("%s: Operation has invalid inputs", __func__); |
| 1218 | } |
| 1219 | |
| 1220 | const Operand* output = GetOutputOperand(operation, 0); |
| 1221 | if (!output) |
| 1222 | { |
| 1223 | return Fail("%s: Could not read output 0", __func__); |
| 1224 | } |
| 1225 | |
| 1226 | const armnn::TensorInfo& inputInfo = input.GetTensorInfo(); |
| 1227 | const armnn::TensorInfo& outputInfo = GetTensorInfoForOperand(*output); |
| 1228 | |
telsoa01 | 5307bc1 | 2018-03-09 13:51:08 +0000 | [diff] [blame] | 1229 | // ArmNN does not currently support non-fixed weights or bias |
| 1230 | ConstTensorPin weightsPin = ConvertOperationInputToConstTensorPin(operation, 1); // 2D |
| 1231 | ConstTensorPin biasPin = ConvertOperationInputToConstTensorPin(operation, 2); // 1D |
| 1232 | |
| 1233 | if (!weightsPin.IsValid() || !biasPin.IsValid()) |
| 1234 | { |
| 1235 | return Fail("%s: Operation has invalid inputs", __func__); |
| 1236 | } |
| 1237 | |
telsoa01 | 5307bc1 | 2018-03-09 13:51:08 +0000 | [diff] [blame] | 1238 | armnn::ConstTensor weights = weightsPin.GetConstTensor(); |
| 1239 | armnn::ConstTensor bias = biasPin.GetConstTensor(); |
telsoa01 | ce3e84a | 2018-08-31 09:31:35 +0100 | [diff] [blame] | 1240 | |
| 1241 | armnn::TensorInfo reshapedInfo = inputInfo; |
| 1242 | if (inputInfo.GetNumDimensions() > 2U) |
| 1243 | { |
| 1244 | unsigned int dim0 = inputInfo.GetShape()[0]; |
| 1245 | unsigned int dim1 = inputInfo.GetShape()[1]; |
| 1246 | |
| 1247 | for (unsigned int i = 2U; i < inputInfo.GetNumDimensions(); ++i) |
| 1248 | { |
| 1249 | dim1 *= inputInfo.GetShape()[i]; |
| 1250 | } |
| 1251 | |
| 1252 | unsigned int divisor = weights.GetInfo().GetShape()[1] / dim1; |
| 1253 | if(dim0 % divisor != 0) |
| 1254 | { |
| 1255 | return Fail("%s: Failed to deduce tensor shape", __func__); |
| 1256 | } |
| 1257 | |
| 1258 | reshapedInfo.SetShape(armnn::TensorShape({dim0 / divisor, dim1 * divisor})); |
| 1259 | } |
| 1260 | |
| 1261 | // ensuring that the bias value is within 1% of the weights input (small float differences can exist) |
telsoa01 | 5307bc1 | 2018-03-09 13:51:08 +0000 | [diff] [blame] | 1262 | SanitizeBiasQuantizationScale(bias.GetInfo(), weights.GetInfo(), reshapedInfo); |
| 1263 | |
| 1264 | ActivationFn activationFunction; |
| 1265 | if (!GetInputActivationFunction(operation, 3, activationFunction)) |
| 1266 | { |
| 1267 | return Fail("%s: Operation has invalid inputs", __func__); |
| 1268 | } |
| 1269 | |
| 1270 | armnn::FullyConnectedDescriptor desc; |
| 1271 | desc.m_TransposeWeightMatrix = true; |
| 1272 | desc.m_BiasEnabled = true; |
| 1273 | |
| 1274 | if (!IsLayerSupported(__func__, |
| 1275 | armnn::IsFullyConnectedSupported, |
| 1276 | m_Compute, |
telsoa01 | ce3e84a | 2018-08-31 09:31:35 +0100 | [diff] [blame] | 1277 | inputInfo, |
| 1278 | outputInfo, |
| 1279 | weights.GetInfo(), |
| 1280 | bias.GetInfo(), |
telsoa01 | 5307bc1 | 2018-03-09 13:51:08 +0000 | [diff] [blame] | 1281 | desc)) |
| 1282 | { |
| 1283 | return false; |
| 1284 | } |
| 1285 | |
| 1286 | armnn::IConnectableLayer* startLayer = m_Network->AddFullyConnectedLayer(desc, weights, bias); |
| 1287 | armnn::IConnectableLayer* endLayer = ProcessActivation(outputInfo, activationFunction, startLayer); |
| 1288 | |
| 1289 | if (endLayer != nullptr) |
| 1290 | { |
| 1291 | if (inputInfo.GetNumDimensions() > 2U) |
| 1292 | { |
| 1293 | armnn::ReshapeDescriptor reshapeDescriptor; |
| 1294 | reshapeDescriptor.m_TargetShape = reshapedInfo.GetShape(); |
| 1295 | |
| 1296 | armnn::IConnectableLayer* reshapeLayer = m_Network->AddReshapeLayer(reshapeDescriptor); |
| 1297 | assert(reshapeLayer != nullptr); |
| 1298 | input.Connect(reshapeLayer->GetInputSlot(0)); |
| 1299 | reshapeLayer->GetOutputSlot(0).SetTensorInfo(reshapedInfo); |
| 1300 | reshapeLayer->GetOutputSlot(0).Connect(startLayer->GetInputSlot(0)); |
| 1301 | } |
| 1302 | else |
| 1303 | { |
| 1304 | input.Connect(startLayer->GetInputSlot(0)); |
| 1305 | } |
| 1306 | |
| 1307 | return SetupAndTrackLayerOutputSlot(operation, 0, *endLayer); |
| 1308 | } |
| 1309 | else |
| 1310 | { |
| 1311 | return Fail("%s: ProcessActivation failed", __func__); |
| 1312 | } |
| 1313 | } |
| 1314 | |
kevmay01 | bc5f784 | 2018-08-30 12:34:39 +0100 | [diff] [blame^] | 1315 | template<typename HalVersion> |
| 1316 | bool ModelToINetworkConverter<HalVersion>::ConvertLocalResponseNormalization( |
| 1317 | const neuralnetworks::V1_0::Operation& operation) |
telsoa01 | 5307bc1 | 2018-03-09 13:51:08 +0000 | [diff] [blame] | 1318 | { |
| 1319 | LayerInputHandle input = ConvertToLayerInputHandle(operation, 0); |
| 1320 | if (!input.IsValid()) |
| 1321 | { |
| 1322 | return Fail("%s: Operation has invalid inputs", __func__); |
| 1323 | } |
| 1324 | |
| 1325 | const Operand* output = GetOutputOperand(operation, 0); |
| 1326 | if (!output) |
| 1327 | { |
| 1328 | return Fail("%s: Could not read output 0", __func__); |
| 1329 | } |
| 1330 | |
| 1331 | const armnn::TensorInfo& inputInfo = input.GetTensorInfo(); |
| 1332 | const armnn::TensorInfo& outputInfo = GetTensorInfoForOperand(*output); |
| 1333 | |
| 1334 | const armnn::TensorInfo swizzledInputInfo = armnnUtils::Permuted(inputInfo, NHWCToArmNN); |
| 1335 | const armnn::TensorInfo swizzledOutputInfo = armnnUtils::Permuted(outputInfo, NHWCToArmNN); |
| 1336 | |
| 1337 | armnn::NormalizationDescriptor descriptor; |
| 1338 | |
| 1339 | descriptor.m_NormChannelType = armnn::NormalizationAlgorithmChannel::Across; |
| 1340 | descriptor.m_NormMethodType = armnn::NormalizationAlgorithmMethod::LocalBrightness; |
| 1341 | |
| 1342 | if (!input.IsValid() || |
| 1343 | !GetInputScalar(operation, 1, OperandType::INT32, descriptor.m_NormSize) || |
| 1344 | !GetInputFloat32(operation, 2, descriptor.m_K) || |
| 1345 | !GetInputFloat32(operation, 3, descriptor.m_Alpha) || |
| 1346 | !GetInputFloat32(operation, 4, descriptor.m_Beta)) |
| 1347 | { |
| 1348 | return Fail("%s: Operation has invalid inputs", __func__); |
| 1349 | } |
| 1350 | |
| 1351 | // ArmNN expects normSize to be the full size of the normalization |
| 1352 | // window rather than the radius as in AndroidNN. |
| 1353 | descriptor.m_NormSize = 1 + (2 * descriptor.m_NormSize); |
| 1354 | |
| 1355 | if (!IsLayerSupported(__func__, |
| 1356 | armnn::IsNormalizationSupported, |
| 1357 | m_Compute, |
| 1358 | swizzledInputInfo, |
| 1359 | swizzledOutputInfo, |
| 1360 | descriptor)) |
| 1361 | { |
| 1362 | return false; |
| 1363 | } |
| 1364 | |
| 1365 | |
| 1366 | armnn::IConnectableLayer* layer = m_Network->AddNormalizationLayer(descriptor); |
| 1367 | assert(layer != nullptr); |
| 1368 | layer->GetOutputSlot(0).SetTensorInfo(swizzledOutputInfo); |
| 1369 | |
| 1370 | armnn::IConnectableLayer& outSwizzleLayer = SwizzleInDeswizzleOut(*m_Network, input, *layer); |
| 1371 | |
| 1372 | return SetupAndTrackLayerOutputSlot(operation, 0, outSwizzleLayer); |
| 1373 | } |
| 1374 | |
kevmay01 | bc5f784 | 2018-08-30 12:34:39 +0100 | [diff] [blame^] | 1375 | template<typename HalVersion> |
| 1376 | bool ModelToINetworkConverter<HalVersion>::ConvertLogistic(const neuralnetworks::V1_0::Operation& operation) |
telsoa01 | 5307bc1 | 2018-03-09 13:51:08 +0000 | [diff] [blame] | 1377 | { |
| 1378 | armnn::ActivationDescriptor desc; |
surmeh01 | 49b9e10 | 2018-05-17 14:11:25 +0100 | [diff] [blame] | 1379 | desc.m_Function = armnn::ActivationFunction::Sigmoid; |
telsoa01 | 5307bc1 | 2018-03-09 13:51:08 +0000 | [diff] [blame] | 1380 | |
| 1381 | return ConvertToActivation(operation, __func__, desc); |
| 1382 | } |
| 1383 | |
kevmay01 | bc5f784 | 2018-08-30 12:34:39 +0100 | [diff] [blame^] | 1384 | template<typename HalVersion> |
| 1385 | bool ModelToINetworkConverter<HalVersion>::ConvertL2Normalization(const neuralnetworks::V1_0::Operation& operation) |
telsoa01 | 5307bc1 | 2018-03-09 13:51:08 +0000 | [diff] [blame] | 1386 | { |
| 1387 | LayerInputHandle input = ConvertToLayerInputHandle(operation, 0); |
| 1388 | if (!input.IsValid()) |
| 1389 | { |
| 1390 | return Fail("%s: Operation has invalid inputs", __func__); |
| 1391 | } |
| 1392 | |
| 1393 | const Operand* output = GetOutputOperand(operation, 0); |
| 1394 | if (!output) |
| 1395 | { |
| 1396 | return Fail("%s: Could not read output 0", __func__); |
| 1397 | } |
| 1398 | |
| 1399 | const armnn::TensorInfo& inputInfo = input.GetTensorInfo(); |
| 1400 | const armnn::TensorInfo& outputInfo = GetTensorInfoForOperand(*output); |
| 1401 | |
| 1402 | const armnn::TensorInfo swizzledInputInfo = armnnUtils::Permuted(inputInfo, NHWCToArmNN); |
| 1403 | const armnn::TensorInfo swizzledOutputInfo = armnnUtils::Permuted(outputInfo, NHWCToArmNN); |
| 1404 | |
| 1405 | if (!IsLayerSupported(__func__, |
| 1406 | armnn::IsL2NormalizationSupported, |
| 1407 | m_Compute, |
telsoa01 | ce3e84a | 2018-08-31 09:31:35 +0100 | [diff] [blame] | 1408 | swizzledInputInfo, |
| 1409 | swizzledOutputInfo)) |
telsoa01 | 5307bc1 | 2018-03-09 13:51:08 +0000 | [diff] [blame] | 1410 | { |
| 1411 | return false; |
| 1412 | } |
| 1413 | |
| 1414 | armnn::IConnectableLayer* layer = m_Network->AddL2NormalizationLayer(); |
| 1415 | assert(layer != nullptr); |
| 1416 | layer->GetOutputSlot(0).SetTensorInfo(swizzledOutputInfo); |
| 1417 | |
| 1418 | armnn::IConnectableLayer& outSwizzleLayer = SwizzleInDeswizzleOut(*m_Network, input, *layer); |
| 1419 | |
| 1420 | return SetupAndTrackLayerOutputSlot(operation, 0, outSwizzleLayer); |
| 1421 | } |
| 1422 | |
kevmay01 | bc5f784 | 2018-08-30 12:34:39 +0100 | [diff] [blame^] | 1423 | template<typename HalVersion> |
| 1424 | bool ModelToINetworkConverter<HalVersion>::ConvertL2Pool2d(const neuralnetworks::V1_0::Operation& operation) |
telsoa01 | 5307bc1 | 2018-03-09 13:51:08 +0000 | [diff] [blame] | 1425 | { |
| 1426 | return ConvertPooling2d(operation, __func__, armnn::PoolingAlgorithm::L2); |
| 1427 | } |
| 1428 | |
kevmay01 | bc5f784 | 2018-08-30 12:34:39 +0100 | [diff] [blame^] | 1429 | template<typename HalVersion> |
| 1430 | bool ModelToINetworkConverter<HalVersion>::ConvertMaxPool2d(const neuralnetworks::V1_0::Operation& operation) |
telsoa01 | 5307bc1 | 2018-03-09 13:51:08 +0000 | [diff] [blame] | 1431 | { |
| 1432 | return ConvertPooling2d(operation, __func__, armnn::PoolingAlgorithm::Max); |
| 1433 | } |
| 1434 | |
kevmay01 | bc5f784 | 2018-08-30 12:34:39 +0100 | [diff] [blame^] | 1435 | template<typename HalVersion> |
| 1436 | bool ModelToINetworkConverter<HalVersion>::ConvertMul(const neuralnetworks::V1_0::Operation& operation) |
telsoa01 | 5307bc1 | 2018-03-09 13:51:08 +0000 | [diff] [blame] | 1437 | { |
| 1438 | LayerInputHandle input0 = ConvertToLayerInputHandle(operation, 0); |
| 1439 | LayerInputHandle input1 = ConvertToLayerInputHandle(operation, 1); |
| 1440 | |
| 1441 | if (!input0.IsValid() || !input1.IsValid()) |
| 1442 | { |
| 1443 | return Fail("%s: Operation has invalid inputs", __func__); |
| 1444 | } |
| 1445 | |
telsoa01 | ce3e84a | 2018-08-31 09:31:35 +0100 | [diff] [blame] | 1446 | // The FuseActivation parameter is always the input index 2 |
| 1447 | // and it should be optional |
telsoa01 | 5307bc1 | 2018-03-09 13:51:08 +0000 | [diff] [blame] | 1448 | ActivationFn activationFunction; |
telsoa01 | ce3e84a | 2018-08-31 09:31:35 +0100 | [diff] [blame] | 1449 | if (!GetOptionalInputActivation(operation, 2, activationFunction)) |
telsoa01 | 5307bc1 | 2018-03-09 13:51:08 +0000 | [diff] [blame] | 1450 | { |
| 1451 | return Fail("%s: Operation has invalid inputs", __func__); |
| 1452 | } |
| 1453 | |
telsoa01 | 5307bc1 | 2018-03-09 13:51:08 +0000 | [diff] [blame] | 1454 | const Operand* outputOperand = GetOutputOperand(operation, 0); |
| 1455 | |
| 1456 | if (outputOperand == nullptr) |
| 1457 | { |
| 1458 | return false; |
| 1459 | } |
| 1460 | |
| 1461 | const armnn::TensorInfo& outInfo = GetTensorInfoForOperand(*outputOperand); |
| 1462 | |
telsoa01 | ce3e84a | 2018-08-31 09:31:35 +0100 | [diff] [blame] | 1463 | if (!IsLayerSupported(__func__, |
| 1464 | armnn::IsMultiplicationSupported, |
| 1465 | m_Compute, |
| 1466 | input0.GetTensorInfo(), |
| 1467 | input1.GetTensorInfo(), |
| 1468 | outInfo)) |
| 1469 | { |
| 1470 | return false; |
| 1471 | } |
| 1472 | |
telsoa01 | 5307bc1 | 2018-03-09 13:51:08 +0000 | [diff] [blame] | 1473 | armnn::IConnectableLayer* const startLayer = m_Network->AddMultiplicationLayer(); |
| 1474 | armnn::IConnectableLayer* const endLayer = ProcessActivation(outInfo, activationFunction, startLayer); |
| 1475 | |
telsoa01 | ce3e84a | 2018-08-31 09:31:35 +0100 | [diff] [blame] | 1476 | const armnn::TensorInfo& inputTensorInfo0 = input0.GetTensorInfo(); |
| 1477 | const armnn::TensorInfo& inputTensorInfo1 = input1.GetTensorInfo(); |
| 1478 | |
telsoa01 | 5307bc1 | 2018-03-09 13:51:08 +0000 | [diff] [blame] | 1479 | if (endLayer != nullptr) |
| 1480 | { |
telsoa01 | ce3e84a | 2018-08-31 09:31:35 +0100 | [diff] [blame] | 1481 | BroadcastTensor(input0, input1, startLayer, *m_Network); |
telsoa01 | 5307bc1 | 2018-03-09 13:51:08 +0000 | [diff] [blame] | 1482 | return SetupAndTrackLayerOutputSlot(operation, 0, *endLayer); |
| 1483 | } |
| 1484 | else |
| 1485 | { |
| 1486 | return Fail("%s: ProcessActivation failed", __func__); |
| 1487 | } |
| 1488 | } |
| 1489 | |
kevmay01 | bc5f784 | 2018-08-30 12:34:39 +0100 | [diff] [blame^] | 1490 | template<typename HalVersion> |
| 1491 | bool ModelToINetworkConverter<HalVersion>::ConvertReLu(const neuralnetworks::V1_0::Operation& operation) |
telsoa01 | 5307bc1 | 2018-03-09 13:51:08 +0000 | [diff] [blame] | 1492 | { |
| 1493 | armnn::ActivationDescriptor desc; |
| 1494 | desc.m_Function = armnn::ActivationFunction::ReLu; |
| 1495 | |
| 1496 | return ConvertToActivation(operation, __func__, desc); |
| 1497 | } |
| 1498 | |
kevmay01 | bc5f784 | 2018-08-30 12:34:39 +0100 | [diff] [blame^] | 1499 | template<typename HalVersion> |
| 1500 | bool ModelToINetworkConverter<HalVersion>::ConvertReLu1(const neuralnetworks::V1_0::Operation& operation) |
telsoa01 | 5307bc1 | 2018-03-09 13:51:08 +0000 | [diff] [blame] | 1501 | { |
| 1502 | armnn::ActivationDescriptor desc; |
| 1503 | desc.m_Function = armnn::ActivationFunction::BoundedReLu; |
| 1504 | desc.m_A = 1.0f; |
| 1505 | desc.m_B = -1.0f; |
| 1506 | |
| 1507 | return ConvertToActivation(operation, __func__, desc); |
| 1508 | } |
| 1509 | |
kevmay01 | bc5f784 | 2018-08-30 12:34:39 +0100 | [diff] [blame^] | 1510 | template<typename HalVersion> |
| 1511 | bool ModelToINetworkConverter<HalVersion>::ConvertReLu6(const neuralnetworks::V1_0::Operation& operation) |
telsoa01 | 5307bc1 | 2018-03-09 13:51:08 +0000 | [diff] [blame] | 1512 | { |
| 1513 | armnn::ActivationDescriptor desc; |
| 1514 | desc.m_Function = armnn::ActivationFunction::BoundedReLu; |
| 1515 | desc.m_A = 6.0f; |
| 1516 | |
| 1517 | return ConvertToActivation(operation, __func__, desc); |
| 1518 | } |
| 1519 | |
kevmay01 | bc5f784 | 2018-08-30 12:34:39 +0100 | [diff] [blame^] | 1520 | template<typename HalVersion> |
| 1521 | bool ModelToINetworkConverter<HalVersion>::ConvertSoftmax(const neuralnetworks::V1_0::Operation& operation) |
telsoa01 | 5307bc1 | 2018-03-09 13:51:08 +0000 | [diff] [blame] | 1522 | { |
| 1523 | LayerInputHandle input = ConvertToLayerInputHandle(operation, 0); |
| 1524 | if (!input.IsValid()) |
| 1525 | { |
| 1526 | return Fail("%s: Operation has invalid inputs", __func__); |
| 1527 | } |
| 1528 | |
telsoa01 | ce3e84a | 2018-08-31 09:31:35 +0100 | [diff] [blame] | 1529 | const Operand* outputOperand = GetOutputOperand(operation, 0); |
| 1530 | if (!outputOperand) |
| 1531 | { |
| 1532 | return Fail("%s: Operation has no outputs", __func__); |
| 1533 | } |
| 1534 | |
| 1535 | const armnn::TensorInfo outInfo = GetTensorInfoForOperand(*outputOperand); |
| 1536 | |
telsoa01 | 5307bc1 | 2018-03-09 13:51:08 +0000 | [diff] [blame] | 1537 | armnn::SoftmaxDescriptor desc; |
| 1538 | if (!GetInputFloat32(operation, 1, desc.m_Beta)) |
| 1539 | { |
| 1540 | return Fail("%s: Operation has invalid inputs", __func__); |
| 1541 | } |
| 1542 | |
| 1543 | if (!IsLayerSupported(__func__, |
| 1544 | armnn::IsSoftmaxSupported, |
| 1545 | m_Compute, |
| 1546 | input.GetTensorInfo(), |
telsoa01 | ce3e84a | 2018-08-31 09:31:35 +0100 | [diff] [blame] | 1547 | outInfo, |
telsoa01 | 5307bc1 | 2018-03-09 13:51:08 +0000 | [diff] [blame] | 1548 | desc)) |
| 1549 | { |
| 1550 | return false; |
| 1551 | } |
| 1552 | |
| 1553 | armnn::IConnectableLayer* layer = m_Network->AddSoftmaxLayer(desc); |
| 1554 | assert(layer != nullptr); |
| 1555 | input.Connect(layer->GetInputSlot(0)); |
| 1556 | |
| 1557 | return SetupAndTrackLayerOutputSlot(operation, 0, *layer); |
| 1558 | } |
| 1559 | |
kevmay01 | bc5f784 | 2018-08-30 12:34:39 +0100 | [diff] [blame^] | 1560 | template<typename HalVersion> |
| 1561 | bool ModelToINetworkConverter<HalVersion>::ConvertTanH(const neuralnetworks::V1_0::Operation& operation) |
telsoa01 | 5307bc1 | 2018-03-09 13:51:08 +0000 | [diff] [blame] | 1562 | { |
| 1563 | armnn::ActivationDescriptor desc; |
| 1564 | desc.m_Function = armnn::ActivationFunction::TanH; |
| 1565 | desc.m_A = 1.0f; // android nn does not support tanH parameters |
| 1566 | desc.m_B = 1.0f; // set to 1.0f for unity scaling |
| 1567 | |
| 1568 | return ConvertToActivation(operation, __func__, desc); |
| 1569 | } |
| 1570 | |
kevmay01 | bc5f784 | 2018-08-30 12:34:39 +0100 | [diff] [blame^] | 1571 | template<typename HalVersion> |
| 1572 | bool ModelToINetworkConverter<HalVersion>::ConvertReshape(const neuralnetworks::V1_0::Operation& operation) |
telsoa01 | 5307bc1 | 2018-03-09 13:51:08 +0000 | [diff] [blame] | 1573 | { |
| 1574 | const Operand* inputOperand = GetInputOperand(operation, 0); |
| 1575 | const Operand* requestedShapeOperand = GetInputOperand(operation, 1); |
| 1576 | const Operand* outputOperand = GetOutputOperand(operation, 0); |
| 1577 | |
| 1578 | if (inputOperand == nullptr |
| 1579 | || requestedShapeOperand == nullptr |
| 1580 | || outputOperand == nullptr) |
| 1581 | { |
| 1582 | return Fail("%s: Operation has invalid inputs", __func__); |
| 1583 | } |
| 1584 | |
| 1585 | |
| 1586 | if (requestedShapeOperand->dimensions.size() != 1) |
| 1587 | { |
| 1588 | return Fail("%s: Input 1 expected to be one-dimensional (found %i dimensions)", |
| 1589 | __func__, requestedShapeOperand->dimensions.size()); |
| 1590 | } |
| 1591 | |
| 1592 | std::vector<int32_t> targetDimensions; |
| 1593 | if (!GetTensorInt32Values(*requestedShapeOperand, targetDimensions)) |
| 1594 | { |
| 1595 | return Fail("%s: Could not read values of input 1", __func__); |
| 1596 | } |
| 1597 | |
| 1598 | const Shape inputOperandShape = GetOperandShape(*inputOperand); |
| 1599 | |
| 1600 | Shape requestedShape; |
| 1601 | // targetDimensions may contain special values (e.g. -1). reshapePrepare() is an AndroidNN provided utility |
| 1602 | // function that resolves these values into a fully specified tensor shape. |
| 1603 | if (!reshapePrepare(inputOperandShape, targetDimensions.data(), targetDimensions.size(), &requestedShape)) |
| 1604 | { |
| 1605 | return Fail("%s: Failed to resolve the requested shape", __func__); |
| 1606 | } |
| 1607 | |
| 1608 | const Shape outputOperandShape = GetOperandShape(*outputOperand); |
| 1609 | if (!SameShape(requestedShape, outputOperandShape)) |
| 1610 | { |
| 1611 | return Fail("%s: Shape of output operand does not match resolved requested shape", __func__); |
| 1612 | } |
| 1613 | |
| 1614 | LayerInputHandle input = ConvertToLayerInputHandle(operation, 0); |
| 1615 | if (!input.IsValid()) |
| 1616 | { |
| 1617 | return Fail("%s: Could not read input 0", __func__); |
| 1618 | } |
| 1619 | |
| 1620 | if (!IsLayerSupported(__func__, |
| 1621 | armnn::IsReshapeSupported, |
| 1622 | m_Compute, |
| 1623 | input.GetTensorInfo())) |
| 1624 | { |
| 1625 | return false; |
| 1626 | } |
| 1627 | |
| 1628 | |
| 1629 | armnn::ReshapeDescriptor reshapeDescriptor; |
| 1630 | reshapeDescriptor.m_TargetShape = armnn::TensorShape(requestedShape.dimensions.size(), |
| 1631 | requestedShape.dimensions.data()); |
| 1632 | |
| 1633 | armnn::IConnectableLayer* layer = m_Network->AddReshapeLayer(reshapeDescriptor); |
| 1634 | assert(layer != nullptr); |
| 1635 | input.Connect(layer->GetInputSlot(0)); |
| 1636 | |
| 1637 | return SetupAndTrackLayerOutputSlot(operation, 0, *layer); |
| 1638 | } |
| 1639 | |
kevmay01 | bc5f784 | 2018-08-30 12:34:39 +0100 | [diff] [blame^] | 1640 | template<typename HalVersion> |
| 1641 | bool ModelToINetworkConverter<HalVersion>::ConvertResizeBilinear(const neuralnetworks::V1_0::Operation& operation) |
telsoa01 | 5307bc1 | 2018-03-09 13:51:08 +0000 | [diff] [blame] | 1642 | { |
| 1643 | LayerInputHandle input = ConvertToLayerInputHandle(operation, 0); |
| 1644 | if (!input.IsValid()) |
| 1645 | { |
| 1646 | return Fail("%s: Could not read input 0", __func__); |
| 1647 | } |
| 1648 | |
| 1649 | const Operand* output = GetOutputOperand(operation, 0); |
| 1650 | if (!output) |
| 1651 | { |
| 1652 | return Fail("%s: Could not read output 0", __func__); |
| 1653 | } |
| 1654 | |
| 1655 | const armnn::TensorInfo& inputInfo = input.GetTensorInfo(); |
| 1656 | const armnn::TensorInfo& outputInfo = GetTensorInfoForOperand(*output); |
| 1657 | |
| 1658 | const armnn::TensorInfo swizzledInputInfo = armnnUtils::Permuted(inputInfo, NHWCToArmNN); |
| 1659 | const armnn::TensorInfo swizzledOutputInfo = armnnUtils::Permuted(outputInfo, NHWCToArmNN); |
| 1660 | |
| 1661 | if (!IsLayerSupported(__func__, |
| 1662 | armnn::IsResizeBilinearSupported, |
| 1663 | m_Compute, |
| 1664 | swizzledInputInfo)) |
| 1665 | { |
| 1666 | return false; |
| 1667 | } |
| 1668 | |
| 1669 | armnn::ResizeBilinearDescriptor desc; |
| 1670 | |
| 1671 | if ( !GetInputScalar(operation, 1, OperandType::INT32, desc.m_TargetHeight) |
| 1672 | || !GetInputScalar(operation, 2, OperandType::INT32, desc.m_TargetWidth)) |
| 1673 | { |
| 1674 | return Fail("%s: Operation has invalid inputs", __func__); |
| 1675 | } |
| 1676 | |
| 1677 | armnn::IConnectableLayer* layer = m_Network->AddResizeBilinearLayer(desc); |
| 1678 | assert(layer != nullptr); |
| 1679 | layer->GetOutputSlot(0).SetTensorInfo(swizzledOutputInfo); |
| 1680 | |
| 1681 | armnn::IConnectableLayer& outSwizzleLayer = SwizzleInDeswizzleOut(*m_Network, input, *layer); |
| 1682 | |
| 1683 | return SetupAndTrackLayerOutputSlot(operation, 0, outSwizzleLayer); |
| 1684 | |
| 1685 | } |
| 1686 | |
kevmay01 | bc5f784 | 2018-08-30 12:34:39 +0100 | [diff] [blame^] | 1687 | template<typename HalVersion> |
| 1688 | bool ModelToINetworkConverter<HalVersion>::ConvertLstm(const neuralnetworks::V1_0::Operation& operation) |
telsoa01 | ce3e84a | 2018-08-31 09:31:35 +0100 | [diff] [blame] | 1689 | { |
| 1690 | // Inputs: |
| 1691 | // 00: The input: A 2-D tensor of ANEURALNETWORKS_TENSOR_FLOAT32, of shape [batch_size, input_size], where |
| 1692 | // “batch_size” corresponds to the batching dimension, and “input_size” is the size of the input. |
| 1693 | LayerInputHandle input = ConvertToLayerInputHandle(operation, 0); |
| 1694 | if (!input.IsValid()) |
| 1695 | { |
| 1696 | return Fail("%s: Could not read input 0: input", __func__); |
| 1697 | } |
| 1698 | // 18: The output state: A 2-D tensor of ANEURALNETWORKS_TENSOR_FLOAT32, of shape [batch_size, output_size]. |
| 1699 | LayerInputHandle outputStateIn = ConvertToLayerInputHandle(operation, 18); |
| 1700 | if (!outputStateIn.IsValid()) |
| 1701 | { |
| 1702 | return Fail("%s: Could not read input 18: outputStateIn", __func__); |
| 1703 | } |
| 1704 | // 19: The cell state: A 2-D tensor of ANEURALNETWORKS_TENSOR_FLOAT32, of shape [batch_size, num_units]. |
| 1705 | LayerInputHandle cellStateIn = ConvertToLayerInputHandle(operation, 19); |
| 1706 | if (!cellStateIn.IsValid()) |
| 1707 | { |
| 1708 | return Fail("%s: Could not read input 19: cellStateIn", __func__); |
| 1709 | } |
| 1710 | |
| 1711 | // Get the mandatory input tensors: |
| 1712 | // 02: The input-to-forget weights: A 2-D tensor of ANEURALNETWORKS_TENSOR_FLOAT32, of shape |
| 1713 | // [num_units, input_size]. |
| 1714 | const ConstTensorPin inputToForgetWeightsPin = ConvertOperationInputToConstTensorPin(operation, 2); |
| 1715 | // 03: The input-to-cell weights: A 2-D tensor of ANEURALNETWORKS_TENSOR_FLOAT32, of shape [num_units, input_size]. |
| 1716 | const ConstTensorPin inputToCellWeightsPin = ConvertOperationInputToConstTensorPin(operation, 3); |
| 1717 | // 04: The input-to-output weights: A 2-D tensor of ANEURALNETWORKS_TENSOR_FLOAT32, of shape |
| 1718 | // [num_units, input_size]. |
| 1719 | const ConstTensorPin inputToOutputWeightsPin = ConvertOperationInputToConstTensorPin(operation, 4); |
| 1720 | // 06: The recurrent-to-forget weights: A 2-D tensor of ANEURALNETWORKS_TENSOR_FLOAT32, of shape |
| 1721 | // [num_units, output_size]. |
| 1722 | const ConstTensorPin recurrentToForgetWeightsPin = ConvertOperationInputToConstTensorPin(operation, 6); |
| 1723 | // 07: The recurrent-to-cell weights: A 2-D tensor of ANEURALNETWORKS_TENSOR_FLOAT32, of shape |
| 1724 | // [num_units, output_size]. |
| 1725 | const ConstTensorPin recurrentToCellWeightsPin = ConvertOperationInputToConstTensorPin(operation, 7); |
| 1726 | // 08: The recurrent-to-output weights: A 2-D tensor of ANEURALNETWORKS_TENSOR_FLOAT32, of shape |
| 1727 | // [num_units, output_size]. |
| 1728 | const ConstTensorPin recurrentToOutputWeightsPin = ConvertOperationInputToConstTensorPin(operation, 8); |
| 1729 | // 13: The forget gate bias: A 1-D tensor of ANEURALNETWORKS_TENSOR_FLOAT32, of shape [num_units]. |
| 1730 | const ConstTensorPin forgetGateBiasPin = ConvertOperationInputToConstTensorPin(operation, 13); |
| 1731 | // 14: The cell bias: A 1-D tensor of ANEURALNETWORKS_TENSOR_FLOAT32, of shape [num_units]. |
| 1732 | const ConstTensorPin cellBiasPin = ConvertOperationInputToConstTensorPin(operation, 14); |
| 1733 | // 15: The output gate bias: A 1-D tensor of ANEURALNETWORKS_TENSOR_FLOAT32, of shape [num_units]. |
| 1734 | const ConstTensorPin outputGateBiasPin = ConvertOperationInputToConstTensorPin(operation, 15); |
| 1735 | |
| 1736 | if (!inputToForgetWeightsPin.IsValid() || |
| 1737 | !inputToCellWeightsPin.IsValid() || |
| 1738 | !inputToOutputWeightsPin.IsValid() || |
| 1739 | !recurrentToForgetWeightsPin.IsValid() || |
| 1740 | !recurrentToCellWeightsPin.IsValid() || |
| 1741 | !recurrentToOutputWeightsPin.IsValid() || |
| 1742 | !forgetGateBiasPin.IsValid() || |
| 1743 | !cellBiasPin.IsValid() || |
| 1744 | !outputGateBiasPin.IsValid()) |
| 1745 | { |
| 1746 | return Fail("%s: Operation has invalid tensor inputs", __func__); |
| 1747 | } |
| 1748 | |
| 1749 | // Get the optional input tensors: |
| 1750 | // 01: The input-to-input weights: Optional. A 2-D tensor of ANEURALNETWORKS_TENSOR_FLOAT32, of shape |
| 1751 | // [num_units, input_size], where “num_units” corresponds to the number of cell units. |
| 1752 | const ConstTensorPin inputToInputWeightsPin = ConvertOperationInputToConstTensorPin(operation, 1); |
| 1753 | // 05: The recurrent-to-input weights: Optional. A 2-D tensor of ANEURALNETWORKS_TENSOR_FLOAT32, of shape |
| 1754 | // [num_units, output_size], where “output_size” corresponds to either the number of cell units (i.e., |
| 1755 | // “num_units”), or the second dimension of the “projection_weights”, if defined. |
| 1756 | const ConstTensorPin recurrentToInputWeightsPin = ConvertOperationInputToConstTensorPin(operation, 5); |
| 1757 | // 09: The cell-to-input weights: Optional. A 1-D tensor of ANEURALNETWORKS_TENSOR_FLOAT32, of shape [num_units]. |
| 1758 | const ConstTensorPin cellToInputWeightsPin = ConvertOperationInputToConstTensorPin(operation, 9); |
| 1759 | // 10: The cell-to-forget weights: Optional. A 1-D tensor of ANEURALNETWORKS_TENSOR_FLOAT32, of shape [num_units]. |
| 1760 | const ConstTensorPin cellToForgetWeightsPin = ConvertOperationInputToConstTensorPin(operation, 10); |
| 1761 | // 11: The cell-to-output weights: Optional. A 1-D tensor of ANEURALNETWORKS_TENSOR_FLOAT32, of shape [num_units]. |
| 1762 | const ConstTensorPin cellToOutputWeightsPin = ConvertOperationInputToConstTensorPin(operation, 11); |
| 1763 | // 12: The input gate bias: Optional. A 1-D tensor of ANEURALNETWORKS_TENSOR_FLOAT32, of shape [num_units]. |
| 1764 | const ConstTensorPin inputGateBiasPin = ConvertOperationInputToConstTensorPin(operation, 12); |
| 1765 | // 16: The projection weights: Optional. A 2-D tensor of ANEURALNETWORKS_TENSOR_FLOAT32, of shape |
| 1766 | // [output_size, num_units]. |
| 1767 | const ConstTensorPin projectionWeightsPin = ConvertOperationInputToConstTensorPin(operation, 16); |
| 1768 | // 17: The projection bias: Optional. A 1-D tensor of ANEURALNETWORKS_TENSOR_FLOAT32, of shape [output_size]. |
| 1769 | const ConstTensorPin projectionBiasPin = ConvertOperationInputToConstTensorPin(operation, 17); |
| 1770 | |
| 1771 | if ((!inputToInputWeightsPin.IsValid() && !inputToInputWeightsPin.IsOptional()) || |
| 1772 | (!recurrentToInputWeightsPin.IsValid() && !recurrentToInputWeightsPin.IsOptional()) || |
| 1773 | (!cellToInputWeightsPin.IsValid() && !cellToInputWeightsPin.IsOptional()) || |
| 1774 | (!cellToForgetWeightsPin.IsValid() && !cellToForgetWeightsPin.IsOptional()) || |
| 1775 | (!cellToOutputWeightsPin.IsValid() && !cellToOutputWeightsPin.IsOptional()) || |
| 1776 | (!inputGateBiasPin.IsValid() && !inputGateBiasPin.IsOptional()) || |
| 1777 | (!projectionWeightsPin.IsValid() && !projectionWeightsPin.IsOptional()) || |
| 1778 | (!projectionBiasPin.IsValid() && !projectionBiasPin.IsOptional())) |
| 1779 | { |
| 1780 | return Fail("%s: Operation has invalid tensor inputs", __func__); |
| 1781 | } |
| 1782 | |
| 1783 | // Get the mandatory input scalars (actually 1-D tensors of size 1): |
| 1784 | // 20: The activation function: A value indicating the activation function: |
| 1785 | // 0: None; 1: Relu; 3: Relu6; 4: Tanh; 6: Sigmoid. |
| 1786 | // 21: The clipping threshold: for the cell state, such that values are bound within [-cell_clip, cell_clip]. |
| 1787 | // If set to 0.0 then clipping is disabled. |
| 1788 | // 22: The clipping threshold: for the output from the projection layer, such that values are bound within |
| 1789 | // [-proj_clip, proj_clip]. If set to 0.0 then clipping is disabled. |
| 1790 | ActivationFn activation; |
| 1791 | float cellClip; |
| 1792 | float projClip; |
| 1793 | if (!GetInputActivationFunctionFromTensor(operation, 20, activation) || |
| 1794 | !GetInputScalar(operation, 21, OperandType::FLOAT32, cellClip) || |
| 1795 | !GetInputScalar(operation, 22, OperandType::FLOAT32, projClip)) |
| 1796 | { |
| 1797 | return Fail("%s: Operation has invalid scalar inputs", __func__); |
| 1798 | } |
| 1799 | |
| 1800 | // Outputs: |
| 1801 | // 00: The scratch buffer: A 2-D tensor of ANEURALNETWORKS_TENSOR_FLOAT32, of shape [batch_size, num_units * 4] with |
| 1802 | // CIFG, or [batch_size, num_units * 3] without CIFG. |
| 1803 | const Operand* scratchBuffer = GetOutputOperand(operation, 0); |
| 1804 | if (!scratchBuffer) |
| 1805 | { |
| 1806 | return Fail("%s: Could not read output 0: scratchBuffer", __func__); |
| 1807 | } |
| 1808 | // 01: The output state (out): A 2-D tensor of ANEURALNETWORKS_TENSOR_FLOAT32, of shape [batch_size, output_size]. |
| 1809 | const Operand* outputStateOut = GetOutputOperand(operation, 1); |
| 1810 | if (!outputStateOut) |
| 1811 | { |
| 1812 | return Fail("%s: Could not read output 1: outputStateOut", __func__); |
| 1813 | } |
| 1814 | // 02: The cell state (out): A 2-D tensor of ANEURALNETWORKS_TENSOR_FLOAT32, of shape [batch_size, num_units]. |
| 1815 | const Operand* cellStateOut = GetOutputOperand(operation, 2); |
| 1816 | if (!cellStateOut) |
| 1817 | { |
| 1818 | return Fail("%s: Could not read output 2: cellStateOut", __func__); |
| 1819 | } |
| 1820 | // 03: The output: A 2-D tensor of ANEURALNETWORKS_TENSOR_FLOAT32, of shape [batch_size, output_size]. This is |
| 1821 | // effectively the same as the current “output state (out)” value. |
| 1822 | const Operand* output = GetOutputOperand(operation, 3); |
| 1823 | if (!output) |
| 1824 | { |
| 1825 | return Fail("%s: Could not read output 3: output", __func__); |
| 1826 | } |
| 1827 | |
| 1828 | // set the params structure for the AddLstmLayer call |
| 1829 | armnn::LstmInputParams params; |
| 1830 | params.m_InputToInputWeights = inputToInputWeightsPin.GetConstTensorPtr(); |
| 1831 | params.m_InputToForgetWeights = inputToForgetWeightsPin.GetConstTensorPtr(); |
| 1832 | params.m_InputToCellWeights = inputToCellWeightsPin.GetConstTensorPtr(); |
| 1833 | params.m_InputToOutputWeights = inputToOutputWeightsPin.GetConstTensorPtr(); |
| 1834 | params.m_RecurrentToInputWeights = recurrentToInputWeightsPin.GetConstTensorPtr(); |
| 1835 | params.m_RecurrentToForgetWeights = recurrentToForgetWeightsPin.GetConstTensorPtr(); |
| 1836 | params.m_RecurrentToCellWeights = recurrentToCellWeightsPin.GetConstTensorPtr(); |
| 1837 | params.m_RecurrentToOutputWeights = recurrentToOutputWeightsPin.GetConstTensorPtr(); |
| 1838 | params.m_CellToInputWeights = cellToInputWeightsPin.GetConstTensorPtr(); |
| 1839 | params.m_CellToForgetWeights = cellToForgetWeightsPin.GetConstTensorPtr(); |
| 1840 | params.m_CellToOutputWeights = cellToOutputWeightsPin.GetConstTensorPtr(); |
| 1841 | params.m_InputGateBias = inputGateBiasPin.GetConstTensorPtr(); |
| 1842 | params.m_ForgetGateBias = forgetGateBiasPin.GetConstTensorPtr(); |
| 1843 | params.m_CellBias = cellBiasPin.GetConstTensorPtr(); |
| 1844 | params.m_OutputGateBias = outputGateBiasPin.GetConstTensorPtr(); |
| 1845 | params.m_ProjectionWeights = projectionWeightsPin.GetConstTensorPtr(); |
| 1846 | params.m_ProjectionBias = projectionBiasPin.GetConstTensorPtr(); |
| 1847 | |
| 1848 | // set the layer descriptor |
| 1849 | armnn::LstmDescriptor desc; |
| 1850 | desc.m_ActivationFunc = activation; |
| 1851 | desc.m_ClippingThresCell = cellClip; |
| 1852 | desc.m_ClippingThresProj = projClip; |
| 1853 | desc.m_CifgEnabled = (params.m_InputToInputWeights == nullptr || |
| 1854 | params.m_RecurrentToInputWeights == nullptr || |
| 1855 | params.m_InputGateBias == nullptr); |
| 1856 | desc.m_PeepholeEnabled = (params.m_CellToForgetWeights != nullptr || |
| 1857 | params.m_CellToOutputWeights != nullptr); |
| 1858 | desc.m_ProjectionEnabled = (params.m_ProjectionWeights != nullptr); |
| 1859 | |
| 1860 | // validate the optional input groups |
| 1861 | if (desc.m_CifgEnabled && |
| 1862 | (params.m_InputToInputWeights != nullptr || |
| 1863 | params.m_RecurrentToInputWeights != nullptr || |
| 1864 | params.m_InputGateBias != nullptr)) |
| 1865 | { |
| 1866 | return Fail("%s: All, or none, of input-to-input weights, recurrent-to-input weights," |
| 1867 | " and input gate bias must be provided", __func__); |
| 1868 | } |
| 1869 | |
| 1870 | if (!desc.m_ProjectionEnabled && params.m_ProjectionBias != nullptr) |
| 1871 | { |
| 1872 | return Fail("%s: projection bias should not be provided without projection weights", __func__); |
| 1873 | } |
| 1874 | |
| 1875 | if (desc.m_PeepholeEnabled && |
| 1876 | (params.m_CellToForgetWeights == nullptr || |
| 1877 | params.m_CellToOutputWeights == nullptr || |
| 1878 | (!desc.m_CifgEnabled && params.m_CellToInputWeights == nullptr))) |
| 1879 | { |
| 1880 | return Fail("%s: All, or none, of cell-to-forget weights and cell-to-output weights must be provided" |
| 1881 | " and, if CIFG is not enabled, cell-to-input weights must also be provided", __func__); |
| 1882 | } |
| 1883 | |
| 1884 | // Check if the layer is supported |
| 1885 | // Inputs |
| 1886 | const armnn::TensorInfo& inputInfo = input.GetTensorInfo(); |
| 1887 | const armnn::TensorInfo& outputStateInInfo = outputStateIn.GetTensorInfo(); |
| 1888 | const armnn::TensorInfo& cellStateInInfo = cellStateIn.GetTensorInfo(); |
| 1889 | |
| 1890 | // Outputs |
| 1891 | const armnn::TensorInfo& scratchBufferInfo = GetTensorInfoForOperand(*scratchBuffer); |
| 1892 | const armnn::TensorInfo& outputStateOutInfo = GetTensorInfoForOperand(*outputStateOut); |
| 1893 | const armnn::TensorInfo& cellStateOutInfo = GetTensorInfoForOperand(*cellStateOut); |
| 1894 | const armnn::TensorInfo& outputInfo = GetTensorInfoForOperand(*output); |
| 1895 | |
| 1896 | // Basic parameters |
| 1897 | const armnn::TensorInfo& inputToForgetWeights = params.m_InputToForgetWeights->GetInfo(); |
| 1898 | const armnn::TensorInfo& inputToCellWeights = params.m_InputToCellWeights->GetInfo(); |
| 1899 | const armnn::TensorInfo& inputToOutputWeights = params.m_InputToOutputWeights->GetInfo(); |
| 1900 | const armnn::TensorInfo& recurrentToForgetWeights = params.m_RecurrentToForgetWeights->GetInfo(); |
| 1901 | const armnn::TensorInfo& recurrentToCellWeights = params.m_RecurrentToCellWeights->GetInfo(); |
| 1902 | const armnn::TensorInfo& recurrentToOutputWeights = params.m_RecurrentToOutputWeights->GetInfo(); |
| 1903 | const armnn::TensorInfo& forgetGateBias = params.m_ForgetGateBias->GetInfo(); |
| 1904 | const armnn::TensorInfo& cellBias = params.m_CellBias->GetInfo(); |
| 1905 | const armnn::TensorInfo& outputGateBias = params.m_OutputGateBias->GetInfo(); |
| 1906 | |
| 1907 | //Optional parameters |
| 1908 | const armnn::TensorInfo* inputToInputWeights = nullptr; |
| 1909 | const armnn::TensorInfo* recurrentToInputWeights = nullptr; |
| 1910 | const armnn::TensorInfo* cellToInputWeights = nullptr; |
| 1911 | const armnn::TensorInfo* inputGateBias = nullptr; |
| 1912 | const armnn::TensorInfo* projectionWeights = nullptr; |
| 1913 | const armnn::TensorInfo* projectionBias = nullptr; |
| 1914 | const armnn::TensorInfo* cellToForgetWeights = nullptr; |
| 1915 | const armnn::TensorInfo* cellToOutputWeights = nullptr; |
| 1916 | |
| 1917 | if(!desc.m_CifgEnabled) |
| 1918 | { |
| 1919 | inputToInputWeights = &(params.m_InputToInputWeights->GetInfo()); |
| 1920 | recurrentToInputWeights = &(params.m_RecurrentToInputWeights->GetInfo()); |
| 1921 | if (params.m_CellToInputWeights != nullptr) |
| 1922 | { |
| 1923 | cellToInputWeights = &(params.m_CellToInputWeights->GetInfo()); |
| 1924 | } |
| 1925 | inputGateBias = &(params.m_InputGateBias->GetInfo()); |
| 1926 | } |
| 1927 | |
| 1928 | if(desc.m_ProjectionEnabled) |
| 1929 | { |
| 1930 | projectionWeights = &(params.m_ProjectionWeights->GetInfo()); |
| 1931 | if (params.m_ProjectionBias != nullptr) |
| 1932 | { |
| 1933 | projectionBias = &(params.m_ProjectionBias->GetInfo()); |
| 1934 | } |
| 1935 | } |
| 1936 | |
| 1937 | if(desc.m_PeepholeEnabled) |
| 1938 | { |
| 1939 | cellToForgetWeights = &(params.m_CellToForgetWeights->GetInfo()); |
| 1940 | cellToOutputWeights = &(params.m_CellToOutputWeights->GetInfo()); |
| 1941 | } |
| 1942 | |
| 1943 | if (!IsLayerSupported(__func__, |
| 1944 | armnn::IsLstmSupported, |
| 1945 | m_Compute, |
| 1946 | inputInfo, |
| 1947 | outputStateInInfo, |
| 1948 | cellStateInInfo, |
| 1949 | scratchBufferInfo, |
| 1950 | outputStateOutInfo, |
| 1951 | cellStateOutInfo, |
| 1952 | outputInfo, |
| 1953 | desc, |
| 1954 | inputToForgetWeights, |
| 1955 | inputToCellWeights, |
| 1956 | inputToOutputWeights, |
| 1957 | recurrentToForgetWeights, |
| 1958 | recurrentToCellWeights, |
| 1959 | recurrentToOutputWeights, |
| 1960 | forgetGateBias, |
| 1961 | cellBias, |
| 1962 | outputGateBias, |
| 1963 | inputToInputWeights, |
| 1964 | recurrentToInputWeights, |
| 1965 | cellToInputWeights, |
| 1966 | inputGateBias, |
| 1967 | projectionWeights, |
| 1968 | projectionBias, |
| 1969 | cellToForgetWeights, |
| 1970 | cellToOutputWeights)) |
| 1971 | { |
| 1972 | return false; |
| 1973 | } |
| 1974 | |
| 1975 | // Add the layer |
| 1976 | armnn::IConnectableLayer* layer = m_Network->AddLstmLayer(desc, params, "Lstm"); |
| 1977 | |
| 1978 | input.Connect(layer->GetInputSlot(0)); |
| 1979 | outputStateIn.Connect(layer->GetInputSlot(1)); |
| 1980 | cellStateIn.Connect(layer->GetInputSlot(2)); |
| 1981 | |
| 1982 | return (SetupAndTrackLayerOutputSlot(operation, 0, *layer, 0) && |
| 1983 | SetupAndTrackLayerOutputSlot(operation, 1, *layer, 1) && |
| 1984 | SetupAndTrackLayerOutputSlot(operation, 2, *layer, 2) && |
| 1985 | SetupAndTrackLayerOutputSlot(operation, 3, *layer, 3)); |
| 1986 | } |
| 1987 | |
kevmay01 | bc5f784 | 2018-08-30 12:34:39 +0100 | [diff] [blame^] | 1988 | template<typename HalVersion> |
| 1989 | bool ModelToINetworkConverter<HalVersion>::ConvertToActivation(const neuralnetworks::V1_0::Operation& operation, |
telsoa01 | 5307bc1 | 2018-03-09 13:51:08 +0000 | [diff] [blame] | 1990 | const char* operationName, |
| 1991 | const armnn::ActivationDescriptor& activationDesc) |
| 1992 | { |
| 1993 | LayerInputHandle input = ConvertToLayerInputHandle(operation, 0); |
| 1994 | if (!input.IsValid()) |
| 1995 | { |
| 1996 | return Fail("%s: Input 0 is invalid", operationName); |
| 1997 | } |
| 1998 | |
telsoa01 | ce3e84a | 2018-08-31 09:31:35 +0100 | [diff] [blame] | 1999 | const Operand* outputOperand = GetOutputOperand(operation, 0); |
| 2000 | if (!outputOperand) |
| 2001 | { |
| 2002 | return false; |
| 2003 | } |
| 2004 | const armnn::TensorInfo outInfo = GetTensorInfoForOperand(*outputOperand); |
telsoa01 | 5307bc1 | 2018-03-09 13:51:08 +0000 | [diff] [blame] | 2005 | if (!IsLayerSupported(__func__, |
| 2006 | armnn::IsActivationSupported, |
| 2007 | m_Compute, |
| 2008 | input.GetTensorInfo(), |
telsoa01 | ce3e84a | 2018-08-31 09:31:35 +0100 | [diff] [blame] | 2009 | outInfo, |
telsoa01 | 5307bc1 | 2018-03-09 13:51:08 +0000 | [diff] [blame] | 2010 | activationDesc)) |
| 2011 | { |
| 2012 | return false; |
| 2013 | } |
| 2014 | |
| 2015 | armnn::IConnectableLayer* layer = m_Network->AddActivationLayer(activationDesc); |
| 2016 | assert(layer != nullptr); |
| 2017 | input.Connect(layer->GetInputSlot(0)); |
| 2018 | |
| 2019 | return SetupAndTrackLayerOutputSlot(operation, 0, *layer); |
| 2020 | } |
| 2021 | |
kevmay01 | bc5f784 | 2018-08-30 12:34:39 +0100 | [diff] [blame^] | 2022 | template<typename HalVersion> |
| 2023 | bool ModelToINetworkConverter<HalVersion>::ConvertPooling2d(const neuralnetworks::V1_0::Operation& operation, |
telsoa01 | 5307bc1 | 2018-03-09 13:51:08 +0000 | [diff] [blame] | 2024 | const char* operationName, |
| 2025 | armnn::PoolingAlgorithm poolType) |
| 2026 | { |
| 2027 | LayerInputHandle input = ConvertToLayerInputHandle(operation, 0); |
| 2028 | if (!input.IsValid()) |
| 2029 | { |
| 2030 | return Fail("%s: Could not read input 0", operationName); |
| 2031 | } |
| 2032 | |
| 2033 | const Operand* output = GetOutputOperand(operation, 0); |
| 2034 | if (!output) |
| 2035 | { |
| 2036 | return Fail("%s: Could not read output 0", __func__); |
| 2037 | } |
| 2038 | |
| 2039 | const armnn::TensorInfo& inputInfo = input.GetTensorInfo(); |
| 2040 | const armnn::TensorInfo& outputInfo = GetTensorInfoForOperand(*output); |
| 2041 | |
| 2042 | const armnn::TensorInfo swizzledInputInfo = armnnUtils::Permuted(inputInfo, NHWCToArmNN); |
| 2043 | const armnn::TensorInfo swizzledOutputInfo = armnnUtils::Permuted(outputInfo, NHWCToArmNN); |
| 2044 | |
| 2045 | armnn::Pooling2dDescriptor desc; |
| 2046 | desc.m_PoolType = poolType; |
| 2047 | desc.m_OutputShapeRounding = armnn::OutputShapeRounding::Floor; |
| 2048 | |
| 2049 | ActivationFn activation; |
| 2050 | |
| 2051 | if (operation.inputs.size() == 7) |
| 2052 | { |
| 2053 | // one input, 6 parameters (padding, stridex, stridey, width, height, activation type) |
| 2054 | android::nn::PaddingScheme scheme; |
| 2055 | |
| 2056 | if ( !GetInputPaddingScheme(operation, 1, scheme) |
| 2057 | || !GetInputScalar(operation, 2, OperandType::INT32, desc.m_StrideX) |
| 2058 | || !GetInputScalar(operation, 3, OperandType::INT32, desc.m_StrideY) |
| 2059 | || !GetInputScalar(operation, 4, OperandType::INT32, desc.m_PoolWidth) |
| 2060 | || !GetInputScalar(operation, 5, OperandType::INT32, desc.m_PoolHeight) |
| 2061 | || !GetInputActivationFunction(operation, 6, activation)) |
| 2062 | { |
| 2063 | return Fail("%s: Operation has invalid inputs", operationName); |
| 2064 | } |
| 2065 | |
| 2066 | const unsigned int inputWidth = swizzledInputInfo.GetShape()[3]; |
| 2067 | const unsigned int inputHeight = swizzledInputInfo.GetShape()[2]; |
| 2068 | |
| 2069 | CalcPadding(inputWidth, desc.m_PoolWidth, desc.m_StrideX, desc.m_PadLeft, desc.m_PadRight, scheme); |
| 2070 | CalcPadding(inputHeight, desc.m_PoolHeight, desc.m_StrideY, desc.m_PadTop, desc.m_PadBottom, scheme); |
| 2071 | } |
| 2072 | else |
| 2073 | { |
| 2074 | // one input, 9 parameters (padding l r t b, stridex, stridey, width, height, activation type) |
| 2075 | if ( !GetInputScalar(operation, 1, OperandType::INT32, desc.m_PadLeft) |
| 2076 | || !GetInputScalar(operation, 2, OperandType::INT32, desc.m_PadRight) |
| 2077 | || !GetInputScalar(operation, 3, OperandType::INT32, desc.m_PadTop) |
| 2078 | || !GetInputScalar(operation, 4, OperandType::INT32, desc.m_PadBottom) |
| 2079 | || !GetInputScalar(operation, 5, OperandType::INT32, desc.m_StrideX) |
| 2080 | || !GetInputScalar(operation, 6, OperandType::INT32, desc.m_StrideY) |
| 2081 | || !GetInputScalar(operation, 7, OperandType::INT32, desc.m_PoolWidth) |
| 2082 | || !GetInputScalar(operation, 8, OperandType::INT32, desc.m_PoolHeight) |
| 2083 | || !GetInputActivationFunction(operation, 9, activation)) |
| 2084 | { |
| 2085 | return Fail("%s: Operation has invalid inputs", operationName); |
| 2086 | } |
| 2087 | } |
| 2088 | |
| 2089 | // ArmNN does not accept a pool size of 1, but the ArmNN driver is expected to cope. |
| 2090 | // This is mapped to a trivial splitter instead. |
| 2091 | armnn::IConnectableLayer* startLayer = nullptr; |
| 2092 | if (desc.m_PoolWidth != 1 || desc.m_PoolHeight != 1) |
| 2093 | { |
| 2094 | if (!IsLayerSupported(__func__, |
| 2095 | armnn::IsPooling2dSupported, |
| 2096 | m_Compute, |
| 2097 | swizzledInputInfo, |
| 2098 | swizzledOutputInfo, |
| 2099 | desc)) |
| 2100 | { |
| 2101 | return false; |
| 2102 | } |
| 2103 | |
| 2104 | startLayer = m_Network->AddPooling2dLayer(desc); |
| 2105 | } |
| 2106 | else |
| 2107 | { |
| 2108 | const unsigned int numDims = swizzledOutputInfo.GetNumDimensions(); |
| 2109 | |
| 2110 | armnn::ViewsDescriptor viewsDesc(1, numDims); |
| 2111 | |
| 2112 | for (unsigned int i = 0; i < numDims; ++i) |
| 2113 | { |
| 2114 | viewsDesc.SetViewOriginCoord(0, i, 0); |
| 2115 | viewsDesc.SetViewSize(0, i, swizzledOutputInfo.GetShape()[i]); |
| 2116 | } |
| 2117 | |
| 2118 | if (!IsLayerSupported(__func__, |
| 2119 | armnn::IsSplitterSupported, |
| 2120 | m_Compute, |
| 2121 | swizzledInputInfo, |
| 2122 | viewsDesc)) |
| 2123 | { |
| 2124 | return false; |
| 2125 | } |
| 2126 | |
| 2127 | startLayer = m_Network->AddSplitterLayer(viewsDesc); |
| 2128 | } |
| 2129 | |
| 2130 | armnn::IConnectableLayer* endLayer = ProcessActivation(swizzledOutputInfo, activation, startLayer); |
| 2131 | |
| 2132 | if (endLayer != nullptr) |
| 2133 | { |
| 2134 | armnn::IConnectableLayer& outSwizzleLayer = SwizzleInDeswizzleOut(*m_Network, input, *startLayer, *endLayer); |
| 2135 | return SetupAndTrackLayerOutputSlot(operation, 0, outSwizzleLayer); |
| 2136 | } |
| 2137 | else |
| 2138 | { |
| 2139 | return Fail("%s: ProcessActivation failed", operationName); |
| 2140 | } |
| 2141 | } |
| 2142 | |
kevmay01 | bc5f784 | 2018-08-30 12:34:39 +0100 | [diff] [blame^] | 2143 | template<typename HalVersion> |
| 2144 | const void* ModelToINetworkConverter<HalVersion>::GetOperandValueReadOnlyAddress(const Operand& operand) const |
telsoa01 | 5307bc1 | 2018-03-09 13:51:08 +0000 | [diff] [blame] | 2145 | { |
| 2146 | const void* valueStart = nullptr; |
| 2147 | |
| 2148 | switch (operand.lifetime) |
| 2149 | { |
| 2150 | case OperandLifeTime::CONSTANT_COPY: |
| 2151 | { |
| 2152 | // Constant found in model.operandValues |
| 2153 | valueStart = &m_Model.operandValues[operand.location.offset]; |
| 2154 | break; |
| 2155 | } |
| 2156 | case OperandLifeTime::CONSTANT_REFERENCE: |
| 2157 | { |
| 2158 | // Constant specified via a Memory object |
| 2159 | valueStart = GetMemoryFromPool(operand.location, m_MemPools); |
| 2160 | break; |
| 2161 | } |
| 2162 | default: |
| 2163 | { |
| 2164 | // Unsupported/invalid (e.g. can't get value of an input to the model) |
| 2165 | Fail("%s: unsupported/invalid operand lifetime: %s", |
| 2166 | __func__, toString(operand.lifetime).c_str()); |
| 2167 | valueStart = nullptr; |
| 2168 | } |
| 2169 | } |
| 2170 | |
| 2171 | return valueStart; |
| 2172 | } |
| 2173 | |
kevmay01 | bc5f784 | 2018-08-30 12:34:39 +0100 | [diff] [blame^] | 2174 | template<typename HalVersion> |
| 2175 | const Operand* ModelToINetworkConverter<HalVersion>::GetInputOperand(const neuralnetworks::V1_0::Operation& operation, |
| 2176 | uint32_t inputIndex) const |
telsoa01 | 5307bc1 | 2018-03-09 13:51:08 +0000 | [diff] [blame] | 2177 | { |
| 2178 | if (inputIndex >= operation.inputs.size()) |
| 2179 | { |
| 2180 | Fail("%s: invalid input index: %i out of %i", __func__, inputIndex, operation.inputs.size()); |
| 2181 | return nullptr; |
| 2182 | } |
| 2183 | |
| 2184 | assert(operation.inputs[inputIndex] < m_Model.operands.size()); // Model should have been validated beforehand |
| 2185 | return &m_Model.operands[operation.inputs[inputIndex]]; |
| 2186 | } |
| 2187 | |
kevmay01 | bc5f784 | 2018-08-30 12:34:39 +0100 | [diff] [blame^] | 2188 | template<typename HalVersion> |
| 2189 | const Operand* ModelToINetworkConverter<HalVersion>::GetOutputOperand(const neuralnetworks::V1_0::Operation& operation, |
| 2190 | uint32_t outputIndex) const |
telsoa01 | 5307bc1 | 2018-03-09 13:51:08 +0000 | [diff] [blame] | 2191 | { |
| 2192 | if (outputIndex >= operation.outputs.size()) |
| 2193 | { |
| 2194 | Fail("%s: invalid output index: %i out of %i", __func__, outputIndex, operation.outputs.size()); |
| 2195 | return nullptr; |
| 2196 | } |
| 2197 | |
| 2198 | assert(operation.outputs[outputIndex] < m_Model.operands.size()); // Model should have been validated beforehand |
| 2199 | return &m_Model.operands[operation.outputs[outputIndex]]; |
| 2200 | } |
| 2201 | |
kevmay01 | bc5f784 | 2018-08-30 12:34:39 +0100 | [diff] [blame^] | 2202 | template<typename HalVersion> |
telsoa01 | 5307bc1 | 2018-03-09 13:51:08 +0000 | [diff] [blame] | 2203 | template<typename T> |
kevmay01 | bc5f784 | 2018-08-30 12:34:39 +0100 | [diff] [blame^] | 2204 | bool ModelToINetworkConverter<HalVersion>::GetInputScalar(const neuralnetworks::V1_0::Operation& operation, |
| 2205 | uint32_t inputIndex, |
telsoa01 | 5307bc1 | 2018-03-09 13:51:08 +0000 | [diff] [blame] | 2206 | OperandType type, T& outValue) const |
| 2207 | { |
| 2208 | const Operand* operand = GetInputOperand(operation, inputIndex); |
| 2209 | if (!operand) |
| 2210 | { |
| 2211 | return Fail("%s: invalid input operand at index %i", __func__, inputIndex); |
| 2212 | } |
| 2213 | |
| 2214 | if (operand->type != type) |
| 2215 | { |
| 2216 | return Fail("%s: unexpected operand type: %s (should be %s)", |
| 2217 | __func__, toString(operand->type).c_str(), toString(type).c_str()); |
| 2218 | } |
| 2219 | |
| 2220 | if (operand->location.length != sizeof(T)) |
| 2221 | { |
| 2222 | return Fail("%s: incorrect operand location length: %i (should be %i)", |
| 2223 | __func__, operand->location.length, sizeof(T)); |
| 2224 | } |
| 2225 | |
| 2226 | const void* valueAddress = GetOperandValueReadOnlyAddress(*operand); |
| 2227 | if (!valueAddress) |
| 2228 | { |
| 2229 | return Fail("%s: failed to get address for operand", __func__); |
| 2230 | } |
| 2231 | |
| 2232 | outValue = *(static_cast<const T*>(valueAddress)); |
| 2233 | return true; |
| 2234 | } |
| 2235 | |
kevmay01 | bc5f784 | 2018-08-30 12:34:39 +0100 | [diff] [blame^] | 2236 | template<typename HalVersion> |
| 2237 | bool ModelToINetworkConverter<HalVersion>::GetInputInt32(const neuralnetworks::V1_0::Operation& operation, |
surmeh01 | deb3bdb | 2018-07-05 12:06:04 +0100 | [diff] [blame] | 2238 | uint32_t inputIndex, int32_t& outValue) const |
telsoa01 | 5307bc1 | 2018-03-09 13:51:08 +0000 | [diff] [blame] | 2239 | { |
| 2240 | return GetInputScalar(operation, inputIndex, OperandType::INT32, outValue); |
| 2241 | } |
| 2242 | |
kevmay01 | bc5f784 | 2018-08-30 12:34:39 +0100 | [diff] [blame^] | 2243 | template<typename HalVersion> |
| 2244 | bool ModelToINetworkConverter<HalVersion>::GetInputFloat32(const neuralnetworks::V1_0::Operation& operation, |
surmeh01 | deb3bdb | 2018-07-05 12:06:04 +0100 | [diff] [blame] | 2245 | uint32_t inputIndex, float& outValue) const |
telsoa01 | 5307bc1 | 2018-03-09 13:51:08 +0000 | [diff] [blame] | 2246 | { |
| 2247 | return GetInputScalar(operation, inputIndex, OperandType::FLOAT32, outValue); |
| 2248 | } |
| 2249 | |
kevmay01 | bc5f784 | 2018-08-30 12:34:39 +0100 | [diff] [blame^] | 2250 | template<typename HalVersion> |
| 2251 | bool ModelToINetworkConverter<HalVersion>::GetInputActivationFunctionImpl( |
| 2252 | const neuralnetworks::V1_0::Operation& operation, |
| 2253 | uint32_t inputIndex, |
| 2254 | OperandType type, |
| 2255 | ActivationFn& outActivationFunction) const |
telsoa01 | 5307bc1 | 2018-03-09 13:51:08 +0000 | [diff] [blame] | 2256 | { |
telsoa01 | ce3e84a | 2018-08-31 09:31:35 +0100 | [diff] [blame] | 2257 | if (type != OperandType::INT32 && type != OperandType::TENSOR_INT32) |
| 2258 | { |
| 2259 | return Fail("%s: unexpected operand type: %s (should be %s or %s)", |
| 2260 | __func__, |
| 2261 | toString(type).c_str(), |
| 2262 | toString(OperandType::INT32).c_str(), |
| 2263 | toString(OperandType::TENSOR_INT32).c_str()); |
| 2264 | } |
| 2265 | |
telsoa01 | 5307bc1 | 2018-03-09 13:51:08 +0000 | [diff] [blame] | 2266 | int32_t activationFunctionAsInt; |
telsoa01 | ce3e84a | 2018-08-31 09:31:35 +0100 | [diff] [blame] | 2267 | if (!GetInputScalar(operation, inputIndex, type, activationFunctionAsInt)) |
telsoa01 | 5307bc1 | 2018-03-09 13:51:08 +0000 | [diff] [blame] | 2268 | { |
| 2269 | return Fail("%s: failed to get activation input value", __func__); |
| 2270 | } |
telsoa01 | 5307bc1 | 2018-03-09 13:51:08 +0000 | [diff] [blame] | 2271 | outActivationFunction = static_cast<ActivationFn>(activationFunctionAsInt); |
| 2272 | return true; |
| 2273 | } |
| 2274 | |
kevmay01 | bc5f784 | 2018-08-30 12:34:39 +0100 | [diff] [blame^] | 2275 | template<typename HalVersion> |
| 2276 | bool ModelToINetworkConverter<HalVersion>::GetInputActivationFunction( |
| 2277 | const neuralnetworks::V1_0::Operation& operation, |
| 2278 | uint32_t inputIndex, |
| 2279 | ActivationFn& outActivationFunction) const |
telsoa01 | ce3e84a | 2018-08-31 09:31:35 +0100 | [diff] [blame] | 2280 | { |
| 2281 | return GetInputActivationFunctionImpl(operation, inputIndex, OperandType::INT32, outActivationFunction); |
| 2282 | } |
| 2283 | |
kevmay01 | bc5f784 | 2018-08-30 12:34:39 +0100 | [diff] [blame^] | 2284 | template<typename HalVersion> |
| 2285 | bool ModelToINetworkConverter<HalVersion>::GetInputActivationFunctionFromTensor( |
| 2286 | const neuralnetworks::V1_0::Operation& operation, |
| 2287 | uint32_t inputIndex, |
| 2288 | ActivationFn& outActivationFunction) const |
telsoa01 | ce3e84a | 2018-08-31 09:31:35 +0100 | [diff] [blame] | 2289 | { |
| 2290 | // This only accepts a 1-D tensor of size 1 |
| 2291 | return GetInputActivationFunctionImpl(operation, inputIndex, OperandType::INT32, outActivationFunction); |
| 2292 | } |
| 2293 | |
kevmay01 | bc5f784 | 2018-08-30 12:34:39 +0100 | [diff] [blame^] | 2294 | template<typename HalVersion> |
| 2295 | bool ModelToINetworkConverter<HalVersion>::GetOptionalInputActivation(const neuralnetworks::V1_0::Operation& operation, |
| 2296 | uint32_t inputIndex, |
| 2297 | ActivationFn& activationFunction) const |
telsoa01 | ce3e84a | 2018-08-31 09:31:35 +0100 | [diff] [blame] | 2298 | { |
| 2299 | if (operation.inputs.size() <= inputIndex) |
| 2300 | { |
| 2301 | activationFunction = ActivationFn::kActivationNone; |
| 2302 | } |
| 2303 | else |
| 2304 | { |
| 2305 | if (!GetInputActivationFunction(operation, inputIndex, activationFunction)) |
| 2306 | { |
| 2307 | return Fail("%s: Operation has invalid inputs", __func__); |
| 2308 | } |
| 2309 | } |
| 2310 | return true; |
| 2311 | } |
| 2312 | |
kevmay01 | bc5f784 | 2018-08-30 12:34:39 +0100 | [diff] [blame^] | 2313 | template<typename HalVersion> |
| 2314 | bool ModelToINetworkConverter<HalVersion>::GetInputPaddingScheme(const neuralnetworks::V1_0::Operation& operation, |
telsoa01 | 5307bc1 | 2018-03-09 13:51:08 +0000 | [diff] [blame] | 2315 | uint32_t inputIndex, |
| 2316 | android::nn::PaddingScheme& outPaddingScheme) const |
| 2317 | { |
| 2318 | int32_t paddingSchemeAsInt; |
| 2319 | if (!GetInputInt32(operation, inputIndex, paddingSchemeAsInt)) |
| 2320 | { |
| 2321 | return Fail("%s: failed to get padding scheme input value", __func__); |
| 2322 | } |
| 2323 | |
| 2324 | outPaddingScheme = static_cast<android::nn::PaddingScheme>(paddingSchemeAsInt); |
| 2325 | return true; |
| 2326 | } |
| 2327 | |
kevmay01 | bc5f784 | 2018-08-30 12:34:39 +0100 | [diff] [blame^] | 2328 | template<typename HalVersion> |
| 2329 | LayerInputHandle ModelToINetworkConverter<HalVersion>::ConvertToLayerInputHandle( |
telsoa01 | ce3e84a | 2018-08-31 09:31:35 +0100 | [diff] [blame] | 2330 | const neuralnetworks::V1_0::Operation& operation, |
telsoa01 | 5307bc1 | 2018-03-09 13:51:08 +0000 | [diff] [blame] | 2331 | uint32_t inputIndex) |
| 2332 | { |
| 2333 | const Operand* operand = GetInputOperand(operation, inputIndex); |
| 2334 | if (!operand) |
| 2335 | { |
| 2336 | Fail("%s: failed to get input operand %i", __func__, inputIndex); |
| 2337 | return LayerInputHandle(); |
| 2338 | } |
| 2339 | |
| 2340 | if (!IsOperandTypeSupportedForTensors(operand->type)) |
| 2341 | { |
| 2342 | Fail("%s: unsupported operand type for tensor %s", __func__, toString(operand->type).c_str()); |
| 2343 | return LayerInputHandle(); |
| 2344 | } |
| 2345 | |
| 2346 | armnn::TensorInfo operandTensorInfo = GetTensorInfoForOperand(*operand); |
| 2347 | |
| 2348 | switch (operand->lifetime) |
| 2349 | { |
| 2350 | case OperandLifeTime::TEMPORARY_VARIABLE: // intentional fallthrough |
| 2351 | case OperandLifeTime::MODEL_INPUT: |
| 2352 | { |
| 2353 | // The tensor is either an operand internal to the model, or a model input. |
| 2354 | // It can be associated with an ArmNN output slot for an existing layer. |
| 2355 | |
| 2356 | // m_OutputSlotForOperand[...] can be nullptr if the previous layer could not be converted |
| 2357 | const uint32_t operandIndex = operation.inputs[inputIndex]; |
| 2358 | return LayerInputHandle(true, m_OutputSlotForOperand[operandIndex], operandTensorInfo); |
| 2359 | break; |
| 2360 | } |
| 2361 | case OperandLifeTime::CONSTANT_COPY: |
| 2362 | case OperandLifeTime::CONSTANT_REFERENCE: |
| 2363 | { |
| 2364 | // The tensor has an already known constant value, and can be converted into an ArmNN Constant layer. |
| 2365 | ConstTensorPin tensorPin = ConvertOperandToConstTensorPin(*operand); |
| 2366 | if (tensorPin.IsValid()) |
| 2367 | { |
| 2368 | if (!IsLayerSupported(__func__, |
| 2369 | armnn::IsConstantSupported, |
| 2370 | m_Compute, |
| 2371 | tensorPin.GetConstTensor().GetInfo())) |
| 2372 | { |
| 2373 | return LayerInputHandle(); |
| 2374 | } |
| 2375 | |
| 2376 | armnn::IConnectableLayer* constantLayer = m_Network->AddConstantLayer(tensorPin.GetConstTensor()); |
| 2377 | armnn::IOutputSlot& outputSlot = constantLayer->GetOutputSlot(0); |
| 2378 | outputSlot.SetTensorInfo(tensorPin.GetConstTensor().GetInfo()); |
| 2379 | |
| 2380 | return LayerInputHandle(true, &outputSlot, operandTensorInfo); |
| 2381 | } |
| 2382 | else |
| 2383 | { |
| 2384 | Fail("%s: invalid operand tensor", __func__); |
| 2385 | return LayerInputHandle(); |
| 2386 | } |
| 2387 | break; |
| 2388 | } |
| 2389 | default: |
| 2390 | { |
| 2391 | // Unsupported lifetime for an input tensor |
| 2392 | Fail("%s: unsupported lifetime for input tensor: %s", |
| 2393 | __func__, toString(operand->lifetime).c_str()); |
| 2394 | return LayerInputHandle(); |
| 2395 | } |
| 2396 | } |
| 2397 | } |
| 2398 | |
kevmay01 | bc5f784 | 2018-08-30 12:34:39 +0100 | [diff] [blame^] | 2399 | template<typename HalVersion> |
| 2400 | ConstTensorPin ModelToINetworkConverter<HalVersion>::ConvertOperationInputToConstTensorPin( |
telsoa01 | ce3e84a | 2018-08-31 09:31:35 +0100 | [diff] [blame] | 2401 | const neuralnetworks::V1_0::Operation& operation, |
| 2402 | uint32_t inputIndex, const armnn::PermutationVector& dimensionMappings, |
| 2403 | const armnn::TensorShape* overrideTensorShape, bool optional) |
telsoa01 | 5307bc1 | 2018-03-09 13:51:08 +0000 | [diff] [blame] | 2404 | { |
| 2405 | const Operand* operand = GetInputOperand(operation, inputIndex); |
| 2406 | if (!operand) |
| 2407 | { |
telsoa01 | ce3e84a | 2018-08-31 09:31:35 +0100 | [diff] [blame] | 2408 | Fail("%s: failed to get input operand: index=%u", __func__, inputIndex); |
telsoa01 | 5307bc1 | 2018-03-09 13:51:08 +0000 | [diff] [blame] | 2409 | return ConstTensorPin(); |
| 2410 | } |
telsoa01 | ce3e84a | 2018-08-31 09:31:35 +0100 | [diff] [blame] | 2411 | return ConvertOperandToConstTensorPin(*operand, dimensionMappings, overrideTensorShape, optional); |
telsoa01 | 5307bc1 | 2018-03-09 13:51:08 +0000 | [diff] [blame] | 2412 | } |
| 2413 | |
kevmay01 | bc5f784 | 2018-08-30 12:34:39 +0100 | [diff] [blame^] | 2414 | template<typename HalVersion> |
| 2415 | ConstTensorPin ModelToINetworkConverter<HalVersion>::ConvertOperandToConstTensorPin(const Operand& operand, |
telsoa01 | ce3e84a | 2018-08-31 09:31:35 +0100 | [diff] [blame] | 2416 | const armnn::PermutationVector& dimensionMappings, const armnn::TensorShape* overrideTensorShape, bool optional) |
telsoa01 | 5307bc1 | 2018-03-09 13:51:08 +0000 | [diff] [blame] | 2417 | { |
| 2418 | if (!IsOperandTypeSupportedForTensors(operand.type)) |
| 2419 | { |
| 2420 | Fail("%s: unsupported operand type for tensor %s", __func__, toString(operand.type).c_str()); |
| 2421 | return ConstTensorPin(); |
| 2422 | } |
| 2423 | |
| 2424 | if (operand.lifetime != OperandLifeTime::CONSTANT_COPY && operand.lifetime != OperandLifeTime::CONSTANT_REFERENCE) |
| 2425 | { |
| 2426 | Fail("%s: invalid operand lifetime: %s", __func__, toString(operand.lifetime).c_str()); |
| 2427 | return ConstTensorPin(); |
| 2428 | } |
| 2429 | |
| 2430 | const void* const valueStart = GetOperandValueReadOnlyAddress(operand); |
| 2431 | if (!valueStart) |
| 2432 | { |
telsoa01 | ce3e84a | 2018-08-31 09:31:35 +0100 | [diff] [blame] | 2433 | if (optional) |
| 2434 | { |
| 2435 | // optional tensor with no values is not really an error; return it as invalid, but marked as optional |
| 2436 | return ConstTensorPin(true); |
| 2437 | } |
| 2438 | // mandatory tensor with no values |
telsoa01 | 5307bc1 | 2018-03-09 13:51:08 +0000 | [diff] [blame] | 2439 | Fail("%s: failed to get operand address", __func__); |
| 2440 | return ConstTensorPin(); |
| 2441 | } |
| 2442 | |
| 2443 | armnn::TensorInfo tensorInfo = GetTensorInfoForOperand(operand); |
| 2444 | if (overrideTensorShape != nullptr) |
| 2445 | { |
| 2446 | tensorInfo.SetShape(*overrideTensorShape); |
| 2447 | } |
| 2448 | return ConstTensorPin(tensorInfo, valueStart, operand.location.length, dimensionMappings); |
| 2449 | } |
| 2450 | |
kevmay01 | bc5f784 | 2018-08-30 12:34:39 +0100 | [diff] [blame^] | 2451 | template<typename HalVersion> |
| 2452 | bool ModelToINetworkConverter<HalVersion>::GetTensorInt32Values(const Operand& operand, |
| 2453 | std::vector<int32_t>& outValues) const |
telsoa01 | 5307bc1 | 2018-03-09 13:51:08 +0000 | [diff] [blame] | 2454 | { |
| 2455 | if (operand.type != OperandType::TENSOR_INT32) |
| 2456 | { |
| 2457 | return Fail("%s: invalid operand type: %s", __func__, toString(operand.type).c_str()); |
| 2458 | } |
| 2459 | |
| 2460 | const void* startAddress = GetOperandValueReadOnlyAddress(operand); |
| 2461 | if (!startAddress) |
| 2462 | { |
| 2463 | return Fail("%s: failed to get operand address", __func__, operand.type); |
| 2464 | } |
| 2465 | |
| 2466 | // Check number of bytes is sensible |
| 2467 | const uint32_t numBytes = operand.location.length; |
| 2468 | if (numBytes % sizeof(int32_t) != 0) |
| 2469 | { |
| 2470 | return Fail("%s: invalid number of bytes: %i, expected to be a multiple of %i", |
| 2471 | __func__, numBytes, sizeof(int32_t)); |
| 2472 | } |
| 2473 | |
| 2474 | outValues.resize(numBytes / sizeof(int32_t)); |
| 2475 | memcpy(outValues.data(), startAddress, numBytes); |
| 2476 | return true; |
| 2477 | } |
| 2478 | |
| 2479 | // Creates an ArmNN activation layer and connects it to the given layer, if the |
| 2480 | // passed in AndroidNN activation function requires so. |
| 2481 | // @return The end layer of the sequence of layers built for the given AndroidNN |
| 2482 | // activation function or nullptr if an error occurred (e.g. unsupported activation). |
| 2483 | // Note that the end layer matches the input layer if no activation is required |
| 2484 | // (the sequence of layers has length 1). |
kevmay01 | bc5f784 | 2018-08-30 12:34:39 +0100 | [diff] [blame^] | 2485 | template<typename HalVersion> |
| 2486 | armnn::IConnectableLayer* ModelToINetworkConverter<HalVersion>::ProcessActivation(const armnn::TensorInfo& tensorInfo, |
telsoa01 | 5307bc1 | 2018-03-09 13:51:08 +0000 | [diff] [blame] | 2487 | ActivationFn activation, armnn::IConnectableLayer* prevLayer) |
| 2488 | { |
| 2489 | assert(prevLayer->GetNumOutputSlots() == 1); |
| 2490 | |
| 2491 | prevLayer->GetOutputSlot(0).SetTensorInfo(tensorInfo); |
| 2492 | |
| 2493 | armnn::IConnectableLayer* activationLayer = prevLayer; |
| 2494 | |
| 2495 | if (activation != ActivationFn::kActivationNone) |
| 2496 | { |
| 2497 | armnn::ActivationDescriptor activationDesc; |
| 2498 | switch (activation) |
| 2499 | { |
| 2500 | case ActivationFn::kActivationRelu: |
| 2501 | { |
| 2502 | activationDesc.m_Function = armnn::ActivationFunction::ReLu; |
| 2503 | break; |
| 2504 | } |
| 2505 | case ActivationFn::kActivationRelu1: |
| 2506 | { |
| 2507 | activationDesc.m_Function = armnn::ActivationFunction::BoundedReLu; |
| 2508 | activationDesc.m_A = 1.0f; |
| 2509 | activationDesc.m_B = -1.0f; |
| 2510 | break; |
| 2511 | } |
| 2512 | case ActivationFn::kActivationRelu6: |
| 2513 | { |
| 2514 | activationDesc.m_Function = armnn::ActivationFunction::BoundedReLu; |
| 2515 | activationDesc.m_A = 6.0f; |
| 2516 | break; |
| 2517 | } |
| 2518 | case ActivationFn::kActivationSigmoid: |
| 2519 | { |
| 2520 | activationDesc.m_Function = armnn::ActivationFunction::Sigmoid; |
| 2521 | break; |
| 2522 | } |
| 2523 | case ActivationFn::kActivationTanh: |
| 2524 | { |
| 2525 | activationDesc.m_Function = armnn::ActivationFunction::TanH; |
| 2526 | activationDesc.m_A = 1.0f; |
| 2527 | activationDesc.m_B = 1.0f; |
| 2528 | break; |
| 2529 | } |
| 2530 | default: |
| 2531 | { |
| 2532 | Fail("%s: Invalid activation enum value %i", __func__, activation); |
| 2533 | return nullptr; |
| 2534 | } |
| 2535 | } |
| 2536 | |
| 2537 | if (!IsLayerSupported(__func__, armnn::IsActivationSupported, m_Compute, |
telsoa01 | ce3e84a | 2018-08-31 09:31:35 +0100 | [diff] [blame] | 2538 | prevLayer->GetOutputSlot(0).GetTensorInfo(), tensorInfo, activationDesc)) |
telsoa01 | 5307bc1 | 2018-03-09 13:51:08 +0000 | [diff] [blame] | 2539 | { |
| 2540 | return nullptr; |
| 2541 | } |
| 2542 | |
| 2543 | activationLayer = m_Network->AddActivationLayer(activationDesc); |
| 2544 | |
| 2545 | prevLayer->GetOutputSlot(0).Connect(activationLayer->GetInputSlot(0)); |
| 2546 | activationLayer->GetOutputSlot(0).SetTensorInfo(tensorInfo); |
| 2547 | } |
| 2548 | |
| 2549 | return activationLayer; |
| 2550 | } |
| 2551 | |
kevmay01 | bc5f784 | 2018-08-30 12:34:39 +0100 | [diff] [blame^] | 2552 | template<typename HalVersion> |
| 2553 | bool ModelToINetworkConverter<HalVersion>::SetupAndTrackLayerOutputSlot( |
| 2554 | const neuralnetworks::V1_0::Operation& operation, |
| 2555 | uint32_t operationOutputIndex, |
| 2556 | armnn::IConnectableLayer& layer, |
| 2557 | uint32_t layerOutputIndex) |
telsoa01 | 5307bc1 | 2018-03-09 13:51:08 +0000 | [diff] [blame] | 2558 | { |
telsoa01 | ce3e84a | 2018-08-31 09:31:35 +0100 | [diff] [blame] | 2559 | const Operand* outputOperand = GetOutputOperand(operation, operationOutputIndex); |
telsoa01 | 5307bc1 | 2018-03-09 13:51:08 +0000 | [diff] [blame] | 2560 | |
telsoa01 | ce3e84a | 2018-08-31 09:31:35 +0100 | [diff] [blame] | 2561 | if ((outputOperand == nullptr) || (operationOutputIndex >= layer.GetNumOutputSlots())) |
telsoa01 | 5307bc1 | 2018-03-09 13:51:08 +0000 | [diff] [blame] | 2562 | { |
| 2563 | return false; |
| 2564 | } |
| 2565 | |
telsoa01 | ce3e84a | 2018-08-31 09:31:35 +0100 | [diff] [blame] | 2566 | armnn::IOutputSlot& outputSlot = layer.GetOutputSlot(layerOutputIndex); |
telsoa01 | 5307bc1 | 2018-03-09 13:51:08 +0000 | [diff] [blame] | 2567 | |
telsoa01 | ce3e84a | 2018-08-31 09:31:35 +0100 | [diff] [blame] | 2568 | const uint32_t operandIndex = operation.outputs[operationOutputIndex]; |
telsoa01 | 5307bc1 | 2018-03-09 13:51:08 +0000 | [diff] [blame] | 2569 | m_OutputSlotForOperand[operandIndex] = &outputSlot; |
| 2570 | |
| 2571 | outputSlot.SetTensorInfo(GetTensorInfoForOperand(*outputOperand)); |
| 2572 | |
| 2573 | return true; |
| 2574 | } |
| 2575 | |
kevmay01 | bc5f784 | 2018-08-30 12:34:39 +0100 | [diff] [blame^] | 2576 | template<typename HalVersion> |
| 2577 | bool ModelToINetworkConverter<HalVersion>::SetupAndTrackLayerOutputSlot( |
| 2578 | const neuralnetworks::V1_0::Operation& operation, |
| 2579 | uint32_t outputIndex, |
| 2580 | armnn::IConnectableLayer& layer) |
telsoa01 | ce3e84a | 2018-08-31 09:31:35 +0100 | [diff] [blame] | 2581 | { |
| 2582 | return SetupAndTrackLayerOutputSlot(operation, outputIndex, layer, outputIndex); |
| 2583 | } |
| 2584 | |
kevmay01 | bc5f784 | 2018-08-30 12:34:39 +0100 | [diff] [blame^] | 2585 | template<typename HalVersion> |
| 2586 | bool ModelToINetworkConverter<HalVersion>::IsOperationSupported(uint32_t operationIndex) const |
telsoa01 | 5307bc1 | 2018-03-09 13:51:08 +0000 | [diff] [blame] | 2587 | { |
| 2588 | std::map<uint32_t, bool>::const_iterator it = m_OperationSupported.find(operationIndex); |
| 2589 | assert(it != m_OperationSupported.end()); |
| 2590 | return it->second; |
| 2591 | } |
| 2592 | |
kevmay01 | bc5f784 | 2018-08-30 12:34:39 +0100 | [diff] [blame^] | 2593 | template class ModelToINetworkConverter<HalVersion_1_0>; |
telsoa01 | 5307bc1 | 2018-03-09 13:51:08 +0000 | [diff] [blame] | 2594 | |
kevmay01 | bc5f784 | 2018-08-30 12:34:39 +0100 | [diff] [blame^] | 2595 | #if defined(ARMNN_ANDROID_NN_V1_1) |
| 2596 | template class ModelToINetworkConverter<HalVersion_1_1>; |
| 2597 | #endif |
| 2598 | |
| 2599 | } // armnn_driver |