diff options
Diffstat (limited to '1.2/HalPolicy.cpp')
| -rw-r--r-- | 1.2/HalPolicy.cpp | 429 |
1 files changed, 428 insertions, 1 deletions
diff --git a/1.2/HalPolicy.cpp b/1.2/HalPolicy.cpp index 9cad29fa..8dbfd897 100644 --- a/1.2/HalPolicy.cpp +++ b/1.2/HalPolicy.cpp @@ -36,7 +36,6 @@ bool HandledByV1_0(V1_2::OperationType operationType) case V1_0::OperationType::LOCAL_RESPONSE_NORMALIZATION: case V1_0::OperationType::LOGISTIC: case V1_0::OperationType::LSH_PROJECTION: - case V1_0::OperationType::LSTM: case V1_0::OperationType::MUL: case V1_0::OperationType::RESHAPE: case V1_0::OperationType::RNN: @@ -164,6 +163,8 @@ bool HalPolicy::ConvertOperation(const Operation& operation, const Model& model, return ConvertSpaceToDepth(operation, model, data); case V1_2::OperationType::TANH: return ConvertTanH(operation, model, data); + case V1_2::OperationType::LSTM: + return ConvertLstm(operation, model, data); default: return Fail("%s: Operation type %s not supported in ArmnnDriver", __func__, toString(operation.type).c_str()); @@ -1008,5 +1009,431 @@ bool HalPolicy::ConvertTanH(const Operation& operation, const Model& model, Conv return ::ConvertTanH<hal_1_2::HalPolicy>(operation, model, data); } +bool HalPolicy::ConvertLstm(const Operation& operation, const Model& model, ConversionData& data) +{ + // Inputs: + // 00: The input: A 2-D tensor of ANEURALNETWORKS_TENSOR_FLOAT32, of shape [batch_size, input_size], where + // “batch_size” corresponds to the batching dimension, and “input_size” is the size of the input. + LayerInputHandle input = ConvertToLayerInputHandle<hal_1_2::HalPolicy>(operation, 0, model, data); + if (!input.IsValid()) + { + return Fail("%s: Could not read input 0: input", __func__); + } + // 18: The output state: A 2-D tensor of ANEURALNETWORKS_TENSOR_FLOAT32, of shape [batch_size, output_size]. + LayerInputHandle outputStateIn = ConvertToLayerInputHandle<hal_1_2::HalPolicy>(operation, 18, model, data); + if (!outputStateIn.IsValid()) + { + return Fail("%s: Could not read input 18: outputStateIn", __func__); + } + // 19: The cell state: A 2-D tensor of ANEURALNETWORKS_TENSOR_FLOAT32, of shape [batch_size, num_units]. + LayerInputHandle cellStateIn = ConvertToLayerInputHandle<hal_1_2::HalPolicy>(operation, 19, model, data); + if (!cellStateIn.IsValid()) + { + return Fail("%s: Could not read input 19: cellStateIn", __func__); + } + + // Get the mandatory input tensors: + // 02: The input-to-forget weights: A 2-D tensor of ANEURALNETWORKS_TENSOR_FLOAT32, of shape + // [num_units, input_size]. + const ConstTensorPin inputToForgetWeightsPin = + ConvertOperationInputToConstTensorPin<hal_1_2::HalPolicy>(operation, 2, model, data); + // 03: The input-to-cell weights: A 2-D tensor of ANEURALNETWORKS_TENSOR_FLOAT32, of shape + // [num_units, input_size]. + const ConstTensorPin inputToCellWeightsPin = + ConvertOperationInputToConstTensorPin<hal_1_2::HalPolicy>(operation, 3, model, data); + // 04: The input-to-output weights: A 2-D tensor of ANEURALNETWORKS_TENSOR_FLOAT32, of shape + // [num_units, input_size]. + const ConstTensorPin inputToOutputWeightsPin = + ConvertOperationInputToConstTensorPin<hal_1_2::HalPolicy>(operation, 4, model, data); + // 06: The recurrent-to-forget weights: A 2-D tensor of ANEURALNETWORKS_TENSOR_FLOAT32, of shape + // [num_units, output_size]. + const ConstTensorPin recurrentToForgetWeightsPin = + ConvertOperationInputToConstTensorPin<hal_1_2::HalPolicy>(operation, 6, model, data); + // 07: The recurrent-to-cell weights: A 2-D tensor of ANEURALNETWORKS_TENSOR_FLOAT32, of shape + // [num_units, output_size]. + const ConstTensorPin recurrentToCellWeightsPin = + ConvertOperationInputToConstTensorPin<hal_1_2::HalPolicy>(operation, 7, model, data); + // 08: The recurrent-to-output weights: A 2-D tensor of ANEURALNETWORKS_TENSOR_FLOAT32, of shape + // [num_units, output_size]. + const ConstTensorPin recurrentToOutputWeightsPin = + ConvertOperationInputToConstTensorPin<hal_1_2::HalPolicy>(operation, 8, model, data); + // 13: The forget gate bias: A 1-D tensor of ANEURALNETWORKS_TENSOR_FLOAT32, of shape [num_units]. + const ConstTensorPin forgetGateBiasPin = + ConvertOperationInputToConstTensorPin<hal_1_2::HalPolicy>(operation, 13, model, data); + // 14: The cell bias: A 1-D tensor of ANEURALNETWORKS_TENSOR_FLOAT32, of shape [num_units]. + const ConstTensorPin cellBiasPin = + ConvertOperationInputToConstTensorPin<hal_1_2::HalPolicy>(operation, 14, model, data); + // 15: The output gate bias: A 1-D tensor of ANEURALNETWORKS_TENSOR_FLOAT32, of shape [num_units]. + const ConstTensorPin outputGateBiasPin = + ConvertOperationInputToConstTensorPin<hal_1_2::HalPolicy>(operation, 15, model, data); + + if (!inputToForgetWeightsPin.IsValid() || + !inputToCellWeightsPin.IsValid() || + !inputToOutputWeightsPin.IsValid() || + !recurrentToForgetWeightsPin.IsValid() || + !recurrentToCellWeightsPin.IsValid() || + !recurrentToOutputWeightsPin.IsValid() || + !forgetGateBiasPin.IsValid() || + !cellBiasPin.IsValid() || + !outputGateBiasPin.IsValid()) + { + return Fail("%s: Operation has invalid tensor inputs", __func__); + } + + // Get the optional input tensors: + // 01: The input-to-input weights: Optional. A 2-D tensor of ANEURALNETWORKS_TENSOR_FLOAT32, of shape + // [num_units, input_size], where “num_units” corresponds to the number of cell units. + const ConstTensorPin inputToInputWeightsPin = + ConvertOperationInputToConstTensorPin<hal_1_2::HalPolicy>(operation, + 1, + model, + data, + g_DontPermute, + nullptr, + true); + + // 05: The recurrent-to-input weights: Optional. A 2-D tensor of ANEURALNETWORKS_TENSOR_FLOAT32, of shape + // [num_units, output_size], where “output_size” corresponds to either the number of cell units (i.e., + // “num_units”), or the second dimension of the “projection_weights”, if defined. + const ConstTensorPin recurrentToInputWeightsPin = + ConvertOperationInputToConstTensorPin<hal_1_2::HalPolicy>(operation, + 5, + model, + data, + g_DontPermute, + nullptr, + true); + + // 09: The cell-to-input weights: Optional. A 1-D tensor of ANEURALNETWORKS_TENSOR_FLOAT32, of shape [num_units]. + const ConstTensorPin cellToInputWeightsPin = + ConvertOperationInputToConstTensorPin<hal_1_2::HalPolicy>(operation, + 9, + model, + data, + g_DontPermute, + nullptr, + true); + + // 10: The cell-to-forget weights: Optional. A 1-D tensor of ANEURALNETWORKS_TENSOR_FLOAT32, of shape [num_units]. + const ConstTensorPin cellToForgetWeightsPin = + ConvertOperationInputToConstTensorPin<hal_1_2::HalPolicy>(operation, + 10, + model, + data, + g_DontPermute, + nullptr, + true); + + // 11: The cell-to-output weights: Optional. A 1-D tensor of ANEURALNETWORKS_TENSOR_FLOAT32, of shape [num_units]. + const ConstTensorPin cellToOutputWeightsPin = + ConvertOperationInputToConstTensorPin<hal_1_2::HalPolicy>(operation, + 11, + model, + data, + g_DontPermute, + nullptr, + true); + + // 12: The input gate bias: Optional. A 1-D tensor of ANEURALNETWORKS_TENSOR_FLOAT32, of shape [num_units]. + const ConstTensorPin inputGateBiasPin = + ConvertOperationInputToConstTensorPin<hal_1_2::HalPolicy>(operation, + 12, + model, + data, + g_DontPermute, + nullptr, + true); + + // 16: The projection weights: Optional. A 2-D tensor of ANEURALNETWORKS_TENSOR_FLOAT32, of shape + // [output_size, num_units]. + const ConstTensorPin projectionWeightsPin = + ConvertOperationInputToConstTensorPin<hal_1_2::HalPolicy>(operation, + 16, + model, + data, + g_DontPermute, + nullptr, + true); + + // 17: The projection bias: Optional. A 1-D tensor of ANEURALNETWORKS_TENSOR_FLOAT32, of shape [output_size]. + const ConstTensorPin projectionBiasPin = + ConvertOperationInputToConstTensorPin<hal_1_2::HalPolicy>(operation, + 17, + model, + data, + g_DontPermute, + nullptr, + true); + + if ((!inputToInputWeightsPin.IsValid() && !inputToInputWeightsPin.IsOptional()) || + (!recurrentToInputWeightsPin.IsValid() && !recurrentToInputWeightsPin.IsOptional()) || + (!cellToInputWeightsPin.IsValid() && !cellToInputWeightsPin.IsOptional()) || + (!cellToForgetWeightsPin.IsValid() && !cellToForgetWeightsPin.IsOptional()) || + (!cellToOutputWeightsPin.IsValid() && !cellToOutputWeightsPin.IsOptional()) || + (!inputGateBiasPin.IsValid() && !inputGateBiasPin.IsOptional()) || + (!projectionWeightsPin.IsValid() && !projectionWeightsPin.IsOptional()) || + (!projectionBiasPin.IsValid() && !projectionBiasPin.IsOptional())) + { + return Fail("%s: Operation has invalid tensor inputs", __func__); + } + + // Get the mandatory input scalars (actually 1-D tensors of size 1): + // 20: The activation function: A value indicating the activation function: + // 0: None; 1: Relu; 3: Relu6; 4: Tanh; 6: Sigmoid. + // 21: The clipping threshold: for the cell state, such that values are bound within [-cell_clip, cell_clip]. + // If set to 0.0 then clipping is disabled. + // 22: The clipping threshold: for the output from the projection layer, such that values are bound within + // [-proj_clip, proj_clip]. If set to 0.0 then clipping is disabled. + ActivationFn activation; + float cellClip; + float projClip; + if (!GetInputActivationFunctionFromTensor<hal_1_2::HalPolicy>(operation, 20, activation, model, data) || + !GetInputScalar<hal_1_2::HalPolicy>(operation, 21, OperandType::FLOAT32, cellClip, model, data) || + !GetInputScalar<hal_1_2::HalPolicy>(operation, 22, OperandType::FLOAT32, projClip, model, data)) + { + return Fail("%s: Operation has invalid scalar inputs", __func__); + } + + // Get the normalization tensors + // 23: The input layer normalization weights. A 1-D tensor of shape [num_units]. + // Used to rescale normalized inputs to activation at input gate. + const ConstTensorPin inputLayerNormWeightsPin = + ConvertOperationInputToConstTensorPin<hal_1_2::HalPolicy>(operation, + 23, + model, + data, + g_DontPermute, + nullptr, + true); + + // 24: The forget layer normalization weights. A 1-D tensor of shape [num_units]. + // Used to rescale normalized inputs to activation at forget gate. + const ConstTensorPin forgetLayerNormWeightsPin = + ConvertOperationInputToConstTensorPin<hal_1_2::HalPolicy>(operation, + 24, + model, + data, + g_DontPermute, + nullptr, + true); + + // 25: The cell layer normalization weights. A 1-D tensor of shape [num_units]. + // Used to rescale normalized inputs to activation at cell gate. + const ConstTensorPin cellLayerNormWeightsPin = + ConvertOperationInputToConstTensorPin<hal_1_2::HalPolicy>(operation, + 25, + model, + data, + g_DontPermute, + nullptr, + true); + + // 26: The output layer normalization weights. A 1-D tensor of shape [num_units]. + // Used to rescale normalized inputs to activation at output gate. + const ConstTensorPin outputLayerNormWeightsPin = + ConvertOperationInputToConstTensorPin<hal_1_2::HalPolicy>(operation, + 26, + model, + data, + g_DontPermute, + nullptr, + true); + + // Outputs: + // 00: The scratch buffer: A 2-D tensor of ANEURALNETWORKS_TENSOR_FLOAT32, of shape [batch_size, num_units * 4] + // with CIFG, or [batch_size, num_units * 3] without CIFG. + const Operand* scratchBuffer = GetOutputOperand<hal_1_2::HalPolicy>(operation, 0, model); + if (!scratchBuffer) + { + return Fail("%s: Could not read output 0: scratchBuffer", __func__); + } + // 01: The output state (out): A 2-D tensor of ANEURALNETWORKS_TENSOR_FLOAT32, of shape [batch_size, output_size]. + const Operand* outputStateOut = GetOutputOperand<hal_1_2::HalPolicy>(operation, 1, model); + if (!outputStateOut) + { + return Fail("%s: Could not read output 1: outputStateOut", __func__); + } + // 02: The cell state (out): A 2-D tensor of ANEURALNETWORKS_TENSOR_FLOAT32, of shape [batch_size, num_units]. + const Operand* cellStateOut = GetOutputOperand<hal_1_2::HalPolicy>(operation, 2, model); + if (!cellStateOut) + { + return Fail("%s: Could not read output 2: cellStateOut", __func__); + } + // 03: The output: A 2-D tensor of ANEURALNETWORKS_TENSOR_FLOAT32, of shape [batch_size, output_size]. This is + // effectively the same as the current “output state (out)” value. + const Operand* output = GetOutputOperand<hal_1_2::HalPolicy>(operation, 3, model); + if (!output) + { + return Fail("%s: Could not read output 3: output", __func__); + } + + // set the params structure for the AddLstmLayer call + armnn::LstmInputParams params; + params.m_InputToInputWeights = inputToInputWeightsPin.GetConstTensorPtr(); + params.m_InputToForgetWeights = inputToForgetWeightsPin.GetConstTensorPtr(); + params.m_InputToCellWeights = inputToCellWeightsPin.GetConstTensorPtr(); + params.m_InputToOutputWeights = inputToOutputWeightsPin.GetConstTensorPtr(); + params.m_RecurrentToInputWeights = recurrentToInputWeightsPin.GetConstTensorPtr(); + params.m_RecurrentToForgetWeights = recurrentToForgetWeightsPin.GetConstTensorPtr(); + params.m_RecurrentToCellWeights = recurrentToCellWeightsPin.GetConstTensorPtr(); + params.m_RecurrentToOutputWeights = recurrentToOutputWeightsPin.GetConstTensorPtr(); + params.m_CellToInputWeights = cellToInputWeightsPin.GetConstTensorPtr(); + params.m_CellToForgetWeights = cellToForgetWeightsPin.GetConstTensorPtr(); + params.m_CellToOutputWeights = cellToOutputWeightsPin.GetConstTensorPtr(); + params.m_InputGateBias = inputGateBiasPin.GetConstTensorPtr(); + params.m_ForgetGateBias = forgetGateBiasPin.GetConstTensorPtr(); + params.m_CellBias = cellBiasPin.GetConstTensorPtr(); + params.m_OutputGateBias = outputGateBiasPin.GetConstTensorPtr(); + params.m_ProjectionWeights = projectionWeightsPin.GetConstTensorPtr(); + params.m_ProjectionBias = projectionBiasPin.GetConstTensorPtr(); + params.m_InputLayerNormWeights = inputLayerNormWeightsPin.GetConstTensorPtr(); + params.m_ForgetLayerNormWeights = forgetLayerNormWeightsPin.GetConstTensorPtr(); + params.m_CellLayerNormWeights = cellLayerNormWeightsPin.GetConstTensorPtr(); + params.m_OutputLayerNormWeights = outputLayerNormWeightsPin.GetConstTensorPtr(); + + // set the layer descriptor + armnn::LstmDescriptor desc; + desc.m_ActivationFunc = activation; + desc.m_ClippingThresCell = cellClip; + desc.m_ClippingThresProj = projClip; + desc.m_CifgEnabled = (params.m_InputToInputWeights == nullptr || + params.m_RecurrentToInputWeights == nullptr || + params.m_InputGateBias == nullptr); + desc.m_PeepholeEnabled = (params.m_CellToForgetWeights != nullptr || + params.m_CellToOutputWeights != nullptr); + desc.m_ProjectionEnabled = (params.m_ProjectionWeights != nullptr); + desc.m_LayerNormEnabled = (params.m_InputLayerNormWeights != nullptr || + params.m_ForgetLayerNormWeights != nullptr || + params.m_CellLayerNormWeights != nullptr || + params.m_OutputLayerNormWeights != nullptr); + + // validate the optional input groups + if (desc.m_CifgEnabled && + (params.m_InputToInputWeights != nullptr || + params.m_RecurrentToInputWeights != nullptr || + params.m_InputGateBias != nullptr)) + { + return Fail("%s: All, or none, of input-to-input weights, recurrent-to-input weights," + " and input gate bias must be provided", __func__); + } + + if (!desc.m_ProjectionEnabled && params.m_ProjectionBias != nullptr) + { + return Fail("%s: projection bias should not be provided without projection weights", __func__); + } + + if (desc.m_PeepholeEnabled && + (params.m_CellToForgetWeights == nullptr || + params.m_CellToOutputWeights == nullptr || + (!desc.m_CifgEnabled && params.m_CellToInputWeights == nullptr))) + { + return Fail("%s: All, or none, of cell-to-forget weights and cell-to-output weights must be provided" + " and, if CIFG is not enabled, cell-to-input weights must also be provided", __func__); + } + + if (desc.m_LayerNormEnabled && + (params.m_ForgetLayerNormWeights == nullptr || + params.m_CellLayerNormWeights == nullptr || + params.m_OutputLayerNormWeights == nullptr || + (!desc.m_CifgEnabled && params.m_InputLayerNormWeights == nullptr))) + { + return Fail("%s: All, or none, of forget-norm weights, cell-norm weights and output-norm weights must be" + " provided and, if CIFG is not enabled, input-norm weights must also be provided", __func__); + } + + // Check if the layer is supported + // Inputs + const armnn::TensorInfo& inputInfo = input.GetTensorInfo(); + const armnn::TensorInfo& outputStateInInfo = outputStateIn.GetTensorInfo(); + const armnn::TensorInfo& cellStateInInfo = cellStateIn.GetTensorInfo(); + + // Outputs + const armnn::TensorInfo& scratchBufferInfo = GetTensorInfoForOperand(*scratchBuffer); + const armnn::TensorInfo& outputStateOutInfo = GetTensorInfoForOperand(*outputStateOut); + const armnn::TensorInfo& cellStateOutInfo = GetTensorInfoForOperand(*cellStateOut); + const armnn::TensorInfo& outputInfo = GetTensorInfoForOperand(*output); + + // Basic parameters + armnn::LstmInputParamsInfo paramsInfo; + paramsInfo.m_InputToForgetWeights = &(params.m_InputToForgetWeights->GetInfo()); + paramsInfo.m_InputToCellWeights = &(params.m_InputToCellWeights->GetInfo()); + paramsInfo.m_InputToOutputWeights = &(params.m_InputToOutputWeights->GetInfo()); + paramsInfo.m_RecurrentToForgetWeights = &(params.m_RecurrentToForgetWeights->GetInfo()); + paramsInfo.m_RecurrentToCellWeights = &(params.m_RecurrentToCellWeights->GetInfo()); + paramsInfo.m_RecurrentToOutputWeights = &(params.m_RecurrentToOutputWeights->GetInfo()); + paramsInfo.m_ForgetGateBias = &(params.m_ForgetGateBias->GetInfo()); + paramsInfo.m_CellBias = &(params.m_CellBias->GetInfo()); + paramsInfo.m_OutputGateBias = &(params.m_OutputGateBias->GetInfo()); + + // Optional parameters + if(!desc.m_CifgEnabled) + { + paramsInfo.m_InputToInputWeights = &(params.m_InputToInputWeights->GetInfo()); + paramsInfo.m_RecurrentToInputWeights = &(params.m_RecurrentToInputWeights->GetInfo()); + if (params.m_CellToInputWeights != nullptr) + { + paramsInfo.m_CellToInputWeights = &(params.m_CellToInputWeights->GetInfo()); + } + paramsInfo.m_InputGateBias = &(params.m_InputGateBias->GetInfo()); + } + + if(desc.m_ProjectionEnabled) + { + paramsInfo.m_ProjectionWeights = &(params.m_ProjectionWeights->GetInfo()); + if (params.m_ProjectionBias != nullptr) + { + paramsInfo.m_ProjectionBias = &(params.m_ProjectionBias->GetInfo()); + } + } + + if(desc.m_PeepholeEnabled) + { + paramsInfo.m_CellToForgetWeights = &(params.m_CellToForgetWeights->GetInfo()); + paramsInfo.m_CellToOutputWeights = &(params.m_CellToOutputWeights->GetInfo()); + } + + if (desc.m_LayerNormEnabled) + { + if(!desc.m_CifgEnabled) + { + paramsInfo.m_InputLayerNormWeights = &(params.m_InputLayerNormWeights->GetInfo()); + } + paramsInfo.m_ForgetLayerNormWeights = &(params.m_ForgetLayerNormWeights->GetInfo()); + paramsInfo.m_CellLayerNormWeights = &(params.m_CellLayerNormWeights->GetInfo()); + paramsInfo.m_OutputLayerNormWeights = &(params.m_OutputLayerNormWeights->GetInfo()); + } + + bool isSupported = false; + FORWARD_LAYER_SUPPORT_FUNC(__func__, + IsLstmSupported, + data.m_Backends, + isSupported, + inputInfo, + outputStateInInfo, + cellStateInInfo, + scratchBufferInfo, + outputStateOutInfo, + cellStateOutInfo, + outputInfo, + desc, + paramsInfo); + if (!isSupported) + { + return false; + } + + // Add the layer + armnn::IConnectableLayer* layer = data.m_Network->AddLstmLayer(desc, params, "Lstm"); + + input.Connect(layer->GetInputSlot(0)); + outputStateIn.Connect(layer->GetInputSlot(1)); + cellStateIn.Connect(layer->GetInputSlot(2)); + + return (SetupAndTrackLayerOutputSlot<hal_1_2::HalPolicy>(operation, 0, *layer, 0, model, data) && + SetupAndTrackLayerOutputSlot<hal_1_2::HalPolicy>(operation, 1, *layer, 1, model, data) && + SetupAndTrackLayerOutputSlot<hal_1_2::HalPolicy>(operation, 2, *layer, 2, model, data) && + SetupAndTrackLayerOutputSlot<hal_1_2::HalPolicy>(operation, 3, *layer, 3, model, data)); +} + } // namespace hal_1_2 } // namespace armnn_driver |