Matthew Sloyan | c5fe6e7 | 2022-11-25 16:10:00 +0000 | [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 "EndToEndTestImpl.hpp" |
| 8 | #include <armnnUtils/QuantizeHelper.hpp> |
| 9 | |
| 10 | #include <ResolveType.hpp> |
| 11 | |
| 12 | #include <CommonTestUtils.hpp> |
| 13 | #include <armnnTestUtils/DataLayoutUtils.hpp> |
| 14 | |
| 15 | #include <map> |
| 16 | #include <vector> |
| 17 | |
| 18 | namespace |
| 19 | { |
| 20 | |
| 21 | armnn::INetworkPtr CreateConstConvolution2dNetwork(const armnn::Convolution2dDescriptor& descriptor, |
| 22 | const armnn::TensorInfo& inputInfo, |
| 23 | const armnn::TensorInfo& weightsInfo, |
| 24 | const armnn::TensorInfo& biasInfo, |
| 25 | const armnn::TensorInfo& outputInfo, |
| 26 | const armnn::ConstTensor& weights, |
| 27 | const armnn::ConstTensor& biases, |
| 28 | bool biasEnabled) |
| 29 | { |
| 30 | using namespace armnn; |
| 31 | |
| 32 | INetworkPtr network(INetwork::Create()); |
| 33 | IConnectableLayer* input = network->AddInputLayer(0, "input"); |
| 34 | IConnectableLayer* weightsLayer = network->AddConstantLayer(weights, "Weights"); |
| 35 | IConnectableLayer* convolution2d = network->AddConvolution2dLayer(descriptor, "convolution2d"); |
| 36 | IConnectableLayer* output = network->AddOutputLayer(0, "output"); |
| 37 | |
| 38 | Connect(input, convolution2d, inputInfo, 0, 0); |
| 39 | Connect(weightsLayer, convolution2d, weightsInfo, 0, 1); |
| 40 | |
| 41 | if(biasEnabled) |
| 42 | { |
| 43 | armnn::IConnectableLayer* biasLayer = network->AddConstantLayer(biases, "Bias"); |
| 44 | Connect(biasLayer, convolution2d, biasInfo, 0, 2); |
| 45 | } |
| 46 | |
| 47 | Connect(convolution2d, output, outputInfo, 0, 0); |
| 48 | |
| 49 | return network; |
| 50 | } |
| 51 | |
| 52 | template<armnn::DataType ArmnnType, typename T = armnn::ResolveType<ArmnnType>> |
| 53 | void Convolution2dEndToEnd(const std::vector<armnn::BackendId>& backends, |
| 54 | armnn::DataLayout dataLayout, |
| 55 | bool biasEnabled = true) |
| 56 | { |
| 57 | using namespace armnn; |
| 58 | |
| 59 | const float qScale = IsQuantizedType<T>() ? 0.25f : 1.0f; |
| 60 | const int32_t qOffset = IsQuantizedType<T>() ? 50 : 0; |
| 61 | |
| 62 | TensorInfo inputInfo({ 1, 5, 5, 1 }, ArmnnType, qScale, qOffset, true); |
| 63 | TensorInfo outputInfo({ 1, 3, 3, 1 }, ArmnnType, qScale, qOffset); |
| 64 | TensorInfo weightsInfo({ 1, 3, 3, 1 }, ArmnnType, qScale, qOffset, true); |
| 65 | TensorInfo biasesInfo({ 1 }, ArmnnType, qScale * qScale, 0, true); |
| 66 | |
| 67 | std::vector<float> inputData = |
| 68 | { |
| 69 | 1.0f, 5.0f, 2.0f, 3.0f, 5.0f, |
| 70 | 8.0f, 7.0f, 3.0f, 6.0f, 3.0f, |
| 71 | 3.0f, 3.0f, 9.0f, 1.0f, 9.0f, |
| 72 | 4.0f, 1.0f, 8.0f, 1.0f, 3.0f, |
| 73 | 6.0f, 8.0f, 1.0f, 9.0f, 2.0f |
| 74 | }; |
| 75 | |
| 76 | std::vector<float> weightsData = |
| 77 | { |
| 78 | 4.0f, 5.0f, 6.0f, |
| 79 | 0.0f, 0.0f, 0.0f, |
| 80 | 3.0f, 2.0f, 1.0f |
| 81 | }; |
| 82 | |
| 83 | std::vector<float> biasesData = { 1.0f }; |
| 84 | |
| 85 | float bias = biasEnabled ? biasesData[0] : 0.0f; |
| 86 | std::vector<float> expectedOutputData = |
| 87 | { |
| 88 | 65.0f + bias, 76.0f + bias, 91.0f + bias, |
| 89 | 107.0f + bias, 99.0f + bias, 89.0f + bias, |
| 90 | 116.0f + bias, 98.0f + bias, 118.0f + bias, |
| 91 | }; |
| 92 | |
| 93 | Convolution2dDescriptor descriptor; |
| 94 | descriptor.m_PadLeft = 0; |
| 95 | descriptor.m_PadRight = 0; |
| 96 | descriptor.m_PadTop = 0; |
| 97 | descriptor.m_PadBottom = 0; |
| 98 | descriptor.m_StrideX = 1; |
| 99 | descriptor.m_StrideY = 1; |
| 100 | descriptor.m_BiasEnabled = biasEnabled; |
| 101 | descriptor.m_DataLayout = dataLayout; |
| 102 | |
| 103 | if (dataLayout == DataLayout::NCHW) |
| 104 | { |
| 105 | PermuteTensorNhwcToNchw(inputInfo, inputData); |
| 106 | PermuteTensorNhwcToNchw(weightsInfo, weightsData); |
| 107 | PermuteTensorNhwcToNchw(outputInfo, expectedOutputData); |
| 108 | } |
| 109 | |
| 110 | // Quantize data |
| 111 | std::vector<T> qInputData = armnnUtils::QuantizedVector<T>(inputData, qScale, qOffset); |
| 112 | std::vector<T> qWeightsData = armnnUtils::QuantizedVector<T>(weightsData, qScale, qOffset); |
| 113 | std::vector<T> qExpectedOutputData = armnnUtils::QuantizedVector<T>(expectedOutputData, qScale, qOffset); |
| 114 | std::vector<T> qBiasesData = armnnUtils::QuantizedVector<T>(biasesData, qScale * qScale, 0); |
| 115 | |
| 116 | ConstTensor weights(weightsInfo, qWeightsData); |
| 117 | ConstTensor biases(biasesInfo, qBiasesData); |
| 118 | |
| 119 | INetworkPtr network = CreateConstConvolution2dNetwork(descriptor, |
| 120 | inputInfo, |
| 121 | weightsInfo, |
| 122 | biasesInfo, |
| 123 | outputInfo, |
| 124 | weights, |
| 125 | biases, |
| 126 | biasEnabled); |
| 127 | |
| 128 | EndToEndLayerTestImpl<ArmnnType, ArmnnType>(std::move(network), |
| 129 | {{ 0, qInputData }}, |
| 130 | {{ 0, qExpectedOutputData }}, |
| 131 | backends); |
| 132 | } |
| 133 | |
| 134 | } // anonymous namespace |