Teresa Charlin | 5841c74 | 2022-05-15 14:07:05 +0100 | [diff] [blame^] | 1 | // |
| 2 | // Copyright © 2022 Arm Ltd and Contributors. All rights reserved. |
| 3 | // SPDX-License-Identifier: MIT |
| 4 | // |
| 5 | #pragma once |
| 6 | |
| 7 | #include "Optimization.hpp" |
| 8 | #include "NetworkUtils.hpp" |
| 9 | |
| 10 | namespace armnn |
| 11 | { |
| 12 | namespace optimizations |
| 13 | { |
| 14 | |
| 15 | class ConvertConstDequantisationLayersToConstLayersImpl |
| 16 | { |
| 17 | public: |
| 18 | void Run(Graph& graph, InputSlot& connection) const |
| 19 | { |
| 20 | Layer& base = connection.GetConnectedOutputSlot()->GetOwningLayer(); |
| 21 | Layer& child = connection.GetOwningLayer(); |
| 22 | |
| 23 | ARMNN_ASSERT(base.GetType() == LayerType::Constant); |
| 24 | ARMNN_ASSERT(child.GetType() == LayerType::Dequantize); |
| 25 | |
| 26 | ReplaceConstDequantisationLayer(graph, |
| 27 | PolymorphicDowncast<ConstantLayer*>(&base), |
| 28 | PolymorphicDowncast<DequantizeLayer*>(&child)); |
| 29 | |
| 30 | } |
| 31 | protected: |
| 32 | ConvertConstDequantisationLayersToConstLayersImpl() = default; |
| 33 | ~ConvertConstDequantisationLayersToConstLayersImpl() = default; |
| 34 | private: |
| 35 | |
| 36 | static void ReplaceConstDequantisationLayer(Graph& graph, |
| 37 | ConstantLayer* constantLayer, |
| 38 | DequantizeLayer* dequantizeLayer) |
| 39 | { |
| 40 | IgnoreUnused(graph); |
| 41 | /** |
| 42 | * This optimisation is to find situations where a constant set of inputs is being provided to a Dequantization |
| 43 | * layer. In this case we don't want the overhead of Dequantizing the values on every inference, instead we |
| 44 | * want to Dequantize them once and store them in a Const layer to be used everytime as they will not change. |
| 45 | */ |
| 46 | TensorInfo constantInfo = constantLayer->GetOutputSlot(0).GetTensorInfo(); |
| 47 | TensorInfo inputDequantizeInfo = dequantizeLayer->GetInputSlot(0).GetConnectedOutputSlot()->GetTensorInfo(); |
| 48 | TensorInfo outputDequantizeInfo = dequantizeLayer->GetOutputSlot(0).GetTensorInfo(); |
| 49 | |
| 50 | ARMNN_ASSERT(constantLayer->GetNumOutputSlots() == 1); |
| 51 | auto numConnections = constantLayer->GetOutputSlot(0).GetNumConnections(); |
| 52 | |
| 53 | std::vector<float> newValues(outputDequantizeInfo.GetNumElements()); |
| 54 | if (constantInfo.GetDataType() == DataType::Float16 && |
| 55 | inputDequantizeInfo.GetDataType() == DataType::Float16 && |
| 56 | outputDequantizeInfo.GetDataType() == DataType::Float32) |
| 57 | { |
| 58 | armnnUtils::FloatingPointConverter::ConvertFloat16To32(constantLayer->m_LayerOutput->Map(true), |
| 59 | outputDequantizeInfo.GetNumElements(), |
| 60 | newValues.data()); |
| 61 | } |
| 62 | else if (constantInfo.GetDataType() == DataType::QAsymmS8 && |
| 63 | inputDequantizeInfo.GetDataType() == DataType::QAsymmS8 && |
| 64 | outputDequantizeInfo.GetDataType() == DataType::Float32) |
| 65 | { |
| 66 | ConvertInt8To32(constantLayer->m_LayerOutput->Map(true), |
| 67 | outputDequantizeInfo.GetNumElements(), |
| 68 | newValues.data()); |
| 69 | } |
| 70 | |
| 71 | TensorInfo newInfo = outputDequantizeInfo; |
| 72 | newInfo.SetConstant(true); |
| 73 | ConstTensor newInput(newInfo, newValues); |
| 74 | constantLayer->m_LayerOutput.reset(new ScopedTensorHandle(newInput)); |
| 75 | |
| 76 | // Moves connections in dequantize output to the constant layer. |
| 77 | // Dequantize layer will be removed if left unconnected. |
| 78 | dequantizeLayer->GetOutputSlot().MoveAllConnections(constantLayer->GetOutputSlot()); |
| 79 | |
| 80 | // Updating the output tensor |
| 81 | constantLayer->GetOutputSlot(0).SetTensorInfo(newInfo); |
| 82 | ARMNN_ASSERT(constantLayer->GetOutputSlot(0).GetTensorInfo().IsConstant() == true); |
| 83 | |
| 84 | // Set isConstant to true in all input tensor infos where constantLayer is now connected to |
| 85 | for (unsigned int i = numConnections; i < constantLayer->GetOutputSlot(0).GetNumConnections(); ++i) |
| 86 | { |
| 87 | auto info = constantLayer->GetOutputSlot(0).GetConnection(i)->GetOwningLayer().GetInputSlot(0) |
| 88 | .GetConnectedOutputSlot()->GetTensorInfo(); |
| 89 | info.SetConstant(); |
| 90 | constantLayer->GetOutputSlot(0).GetConnection(i)->GetOwningLayer().GetInputSlot(0) |
| 91 | .GetConnectedOutputSlot()->SetTensorInfo(info); |
| 92 | } |
| 93 | } |
| 94 | |
| 95 | |
| 96 | static void ConvertInt8To32(const void* srcInt8Buffer, |
| 97 | size_t numElements, |
| 98 | float* dstFloat32Buffer) |
| 99 | { |
| 100 | ARMNN_ASSERT(srcInt8Buffer != nullptr); |
| 101 | ARMNN_ASSERT(dstFloat32Buffer != nullptr); |
| 102 | |
| 103 | const auto* pInt8 = static_cast<const int8_t*>(srcInt8Buffer); |
| 104 | |
| 105 | for (size_t i = 0; i < numElements; ++i) |
| 106 | { |
| 107 | dstFloat32Buffer[i] = pInt8[i]; |
| 108 | } |
| 109 | } |
| 110 | |
| 111 | }; |
| 112 | |
| 113 | using ConvertConstDequantisationLayersToConstLayers |
| 114 | = OptimizeForConnection<ConstantLayer, |
| 115 | DequantizeLayer, |
| 116 | ConvertConstDequantisationLayersToConstLayersImpl>; |
| 117 | |
| 118 | } // namespace optimizations |
| 119 | } // namespace armnn |