Teresa Charlin | ca5a23a | 2023-12-15 14:20:47 +0000 | [diff] [blame] | 1 | // |
| 2 | // Copyright © 2023 Arm Ltd and Contributors. All rights reserved. |
| 3 | // SPDX-License-Identifier: MIT |
| 4 | // |
| 5 | // Copyright © 2020 The TensorFlow Authors. All Rights Reserved. |
| 6 | // SPDX-License-Identifier: Apache-2.0 |
| 7 | // |
| 8 | |
| 9 | #include "QuantizeOperator.hpp" |
| 10 | |
| 11 | // This function is paraphrased from: |
| 12 | // tensorflow/compiler/mlir/tosa/transforms/legalize_common.cc from function convertQuantizeOp |
| 13 | TosaSerializationBasicBlock* ConvertQuantizeToTosaOperator(const Layer* layer, |
| 14 | const std::vector<const TensorInfo*>& inputs, |
| 15 | const std::vector<const TensorInfo*>& outputs) |
| 16 | { |
| 17 | ARMNN_THROW_INVALIDARG_MSG_IF_FALSE( inputs.size() == 1, |
| 18 | "ConvertQuantizeToTosaOperator: Quantize must have only one input" ); |
| 19 | ARMNN_THROW_INVALIDARG_MSG_IF_FALSE( outputs.size() == 1, |
| 20 | "ConvertQuantizeToTosaOperator: Quantize must have only one output" ); |
| 21 | |
| 22 | std::string inputName = std::string("input0_"); |
| 23 | std::string outputNameZeroPoint = std::string("intermediate0_") + GetUniqueTosaMappingID(); |
| 24 | std::string outputNameScale = std::string("intermediate1_") + GetUniqueTosaMappingID(); |
| 25 | std::string outputNameMul = std::string("intermediate2_") + GetUniqueTosaMappingID(); |
| 26 | std::string outputNameAdd = std::string("intermediate3_") + GetUniqueTosaMappingID(); |
| 27 | std::string outputName = std::string("output0_"); |
| 28 | std::string blockName = std::string("Op_QUANTIZE_block_") + GetUniqueTosaMappingID(); |
| 29 | |
| 30 | // If a layer is present then the block will be used for execution, so input and output names need to be determined |
| 31 | // using the previous and following layers so the graph is connected correctly. For validation this doesn't matter. |
| 32 | if(layer != nullptr) |
| 33 | { |
| 34 | // Get the layers connected to the input slots and determine unique tensor names. |
| 35 | Layer& connectedLayer = layer->GetInputSlot(0).GetConnectedOutputSlot()->GetOwningLayer(); |
| 36 | inputName = GenerateUniqueName(connectedLayer, 0); |
| 37 | |
| 38 | // Determine unique output tensor name. |
| 39 | outputName = GenerateUniqueOutputName(*layer, 0); |
| 40 | } |
| 41 | |
| 42 | const TensorInfo inputInfo = *inputs[0]; |
| 43 | const TensorInfo outputInfo = *outputs[0]; |
| 44 | |
| 45 | // Extract quantization detail from Tensor |
| 46 | float zeroPoint = static_cast<float>(outputInfo.GetQuantizationOffset()); |
| 47 | // No per axis support in Tensorflow TOSA code |
| 48 | float scale = outputInfo.GetQuantizationScale(); |
| 49 | |
| 50 | // As per the Tensorflow quantization specification |
| 51 | // Tensorflow TOSA code calculates quantization using multiplication by scale |
| 52 | // Armnn code calculates quantization using division by scale |
| 53 | // Invert scale factor passed from Armnn for tf TOSA code |
| 54 | scale = (scale != 0) ? (1 / scale) : scale; |
| 55 | |
| 56 | std::vector<TosaSerializationTensor*> tensors; |
| 57 | |
| 58 | // Only add input tensors if connected layer is an input layer. |
| 59 | // As intermediate or constant tensors will be created separately. |
| 60 | // There also can't be duplicate tensor. |
| 61 | std::vector<int32_t> inputShape0; |
| 62 | DType inputDType0 = DType::DType_UNKNOWN; |
| 63 | if(inputName.find("input0_") != std::string::npos) |
| 64 | { |
| 65 | inputShape0 = GetTosaTensorShape(inputInfo.GetShape()); |
| 66 | inputDType0 = ArmNNToDType(inputInfo.GetDataType()); |
| 67 | ARMNN_THROW_INVALIDARG_MSG_IF_FALSE( inputDType0 == DType::DType_FP16 || inputDType0 == DType::DType_FP32, |
| 68 | "ConvertQuantizeToTosaOperator: Quantize input must be of type Float" ); |
| 69 | tensors.push_back(new TosaSerializationTensor(inputName, inputShape0, inputDType0, {})); |
| 70 | } |
| 71 | |
| 72 | std::vector<int32_t> outputShape0 = GetTosaTensorShape(outputInfo.GetShape()); |
| 73 | DType outputDType0 = ArmNNToDType(outputInfo.GetDataType()); |
| 74 | |
| 75 | // quantize: |
| 76 | // const_zeroPoint = constant(zeroPoint) |
| 77 | // const_scale = constant(scale) |
| 78 | // out_mul = mul(input, const_scale) |
| 79 | // out_add = add(out_mul, const_zeroPoint) |
| 80 | // output = cast<output_type>(out_add) |
| 81 | |
| 82 | // const_zeroPoint |
| 83 | TosaSerializationOperator* zeroPointOp = nullptr; |
| 84 | TosaSerializationTensor* zeroPointTensor = nullptr; |
| 85 | CreateConstTosaOperator<float>(outputNameZeroPoint, |
| 86 | zeroPoint, |
| 87 | inputDType0, |
| 88 | inputShape0, |
| 89 | zeroPointOp, |
| 90 | zeroPointTensor); |
| 91 | tensors.push_back(zeroPointTensor); |
| 92 | |
| 93 | // const_scale |
| 94 | TosaSerializationOperator *scaleOp = nullptr; |
| 95 | TosaSerializationTensor* scaleTensor = nullptr; |
| 96 | CreateConstTosaOperator<float>(outputNameScale, |
| 97 | scale, |
| 98 | inputDType0, |
| 99 | inputShape0, |
| 100 | scaleOp, |
| 101 | scaleTensor); |
| 102 | tensors.push_back(scaleTensor); |
| 103 | |
| 104 | // mul |
| 105 | int32_t shift = 0; |
| 106 | TosaMulAttribute mulAttribute(shift); |
| 107 | TosaSerializationOperator* mulOp = new TosaSerializationOperator(Op_MUL, |
| 108 | Attribute_MulAttribute, |
| 109 | &mulAttribute, |
| 110 | {inputName, outputNameScale}, |
| 111 | {outputNameMul}); |
| 112 | tensors.push_back(new TosaSerializationTensor(outputNameMul, inputShape0, inputDType0, {})); |
| 113 | |
| 114 | // add |
| 115 | TosaSerializationOperator* addOp = new TosaSerializationOperator(Op_ADD, |
| 116 | Attribute_NONE, |
| 117 | nullptr, |
| 118 | {outputNameMul, outputNameZeroPoint}, |
| 119 | {outputNameAdd}); |
| 120 | tensors.push_back(new TosaSerializationTensor(outputNameAdd, inputShape0, inputDType0, {})); |
| 121 | |
| 122 | // cast |
| 123 | TosaSerializationOperator* castOp = new TosaSerializationOperator(Op_CAST, |
| 124 | Attribute_NONE, |
| 125 | nullptr, |
| 126 | {outputNameAdd}, |
| 127 | {outputName}); |
| 128 | |
| 129 | tensors.push_back(new TosaSerializationTensor(outputName, outputShape0, outputDType0, {})); |
| 130 | |
| 131 | // operatorInputNames/operatorOutputNames ends up being the same as |
| 132 | // blockInputNames/blockOutputNames for one-to-one ArmNN to TOSA mappings |
| 133 | return new TosaSerializationBasicBlock(blockName, // name |
| 134 | mainName, // region name |
| 135 | {zeroPointOp, scaleOp, mulOp, addOp, castOp}, // operators |
| 136 | tensors, // tensors |
| 137 | {inputName}, // inputs |
| 138 | {outputName}); // outputs |
| 139 | } |