telsoa01 | 4fcda01 | 2018-03-09 14:13:49 +0000 | [diff] [blame] | 1 | // |
| 2 | // Copyright © 2017 Arm Ltd. All rights reserved. |
David Beck | ecb56cd | 2018-09-05 12:52:57 +0100 | [diff] [blame] | 3 | // SPDX-License-Identifier: MIT |
telsoa01 | 4fcda01 | 2018-03-09 14:13:49 +0000 | [diff] [blame] | 4 | // |
| 5 | #include "LayerTests.hpp" |
| 6 | |
| 7 | #include "test/TensorHelpers.hpp" |
| 8 | #include "TensorCopyUtils.hpp" |
surmeh01 | 3537c2c | 2018-05-18 16:31:43 +0100 | [diff] [blame] | 9 | #include "Permute.hpp" |
telsoa01 | 4fcda01 | 2018-03-09 14:13:49 +0000 | [diff] [blame] | 10 | |
| 11 | #include <boost/test/unit_test.hpp> |
surmeh01 | 3537c2c | 2018-05-18 16:31:43 +0100 | [diff] [blame] | 12 | #include <boost/assert.hpp> |
telsoa01 | 4fcda01 | 2018-03-09 14:13:49 +0000 | [diff] [blame] | 13 | |
David Beck | 711fa31 | 2018-09-24 10:46:38 +0100 | [diff] [blame] | 14 | #include <armnn/LayerSupport.hpp> |
telsoa01 | 4fcda01 | 2018-03-09 14:13:49 +0000 | [diff] [blame] | 15 | |
David Beck | 711fa31 | 2018-09-24 10:46:38 +0100 | [diff] [blame] | 16 | #include <backends/CpuTensorHandle.hpp> |
| 17 | #include <backends/WorkloadFactory.hpp> |
telsoa01 | 4fcda01 | 2018-03-09 14:13:49 +0000 | [diff] [blame] | 18 | |
| 19 | #ifdef ARMCOMPUTECL_ENABLED |
David Beck | ac42efd | 2018-09-26 17:41:13 +0100 | [diff] [blame] | 20 | #include <backends/cl/ClTensorHandle.hpp> |
David Beck | 711fa31 | 2018-09-24 10:46:38 +0100 | [diff] [blame] | 21 | #include <backends/aclCommon/ArmComputeTensorUtils.hpp> |
telsoa01 | 4fcda01 | 2018-03-09 14:13:49 +0000 | [diff] [blame] | 22 | #endif |
| 23 | |
| 24 | #include <algorithm> |
| 25 | #include <boost/cast.hpp> |
| 26 | |
| 27 | #include "WorkloadTestUtils.hpp" |
| 28 | #include "Conv2dTestImpl.hpp" |
| 29 | #include "BatchNormTestImpl.hpp" |
| 30 | #include "ActivationTestImpl.hpp" |
| 31 | #include "Pooling2dTestImpl.hpp" |
| 32 | #include "ReshapeTestImpl.hpp" |
| 33 | #include "FullyConnectedTestImpl.hpp" |
| 34 | #include "SplitterTestImpl.hpp" |
| 35 | #include "SoftmaxTestImpl.hpp" |
| 36 | #include "NormTestImpl.hpp" |
| 37 | #include "PermuteTestImpl.hpp" |
telsoa01 | c577f2c | 2018-08-31 09:22:23 +0100 | [diff] [blame] | 38 | #include "LstmTestImpl.hpp" |
| 39 | #include "ConvertFp16ToFp32TestImpl.hpp" |
| 40 | #include "ConvertFp32ToFp16TestImpl.hpp" |
telsoa01 | 4fcda01 | 2018-03-09 14:13:49 +0000 | [diff] [blame] | 41 | |
telsoa01 | c577f2c | 2018-08-31 09:22:23 +0100 | [diff] [blame] | 42 | // 3-channel 16x8 image used as common input data for a number of Conv2d tests. |
telsoa01 | 4fcda01 | 2018-03-09 14:13:49 +0000 | [diff] [blame] | 43 | static std::vector<float> ConvInput3x8x16({ |
| 44 | 0.5f, 0.5f, 0.5f, 0.5f, 0.5f, 0.5f, 0.5f, 0.5f, 0.5f, 0.5f, 0.5f, 0.5f, 0.5f, 0.5f, 0.5f, 0.5f, |
| 45 | 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, |
| 46 | 0.5f, 0.5f, 0.5f, 0.5f, 0.5f, 0.5f, 0.5f, 0.5f, 0.5f, 0.5f, 0.5f, 0.5f, 0.5f, 0.5f, 0.5f, 0.5f, |
| 47 | 0.5f, 0.5f, 0.5f, 0.5f, 0.5f, 0.5f, 0.5f, 0.5f, 0.5f, 0.5f, 0.5f, 0.5f, 0.5f, 0.5f, 0.5f, 0.5f, |
| 48 | 0.5f, 0.5f, 0.5f, 0.5f, 0.5f, 0.5f, 0.5f, 0.5f, 0.5f, 0.5f, 0.5f, 0.5f, 0.5f, 0.5f, 0.5f, 0.5f, |
| 49 | 0.5f, 0.5f, 0.5f, 0.5f, 0.5f, 0.5f, 0.5f, 0.5f, 0.5f, 0.5f, 0.5f, 0.5f, 0.5f, 0.5f, 0.5f, 0.5f, |
| 50 | 0.5f, 0.5f, 0.5f, 0.5f, 0.5f, 0.5f, 0.5f, 0.5f, 0.5f, 0.5f, 0.5f, 0.5f, 0.5f, 0.5f, 0.5f, 0.5f, |
| 51 | 0.5f, 0.5f, 0.5f, 0.5f, 0.5f, 0.5f, 0.5f, 0.5f, 0.5f, 0.5f, 0.5f, 0.5f, 0.5f, 0.5f, 0.5f, 0.5f, |
| 52 | 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, |
| 53 | 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, |
| 54 | 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, |
| 55 | 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, |
| 56 | 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, |
| 57 | 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, |
| 58 | 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, |
| 59 | 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, |
| 60 | -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, |
| 61 | -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, |
| 62 | -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, |
| 63 | -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, |
| 64 | -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, |
| 65 | -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, |
| 66 | -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, |
| 67 | -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1 |
| 68 | }); |
| 69 | |
telsoa01 | c577f2c | 2018-08-31 09:22:23 +0100 | [diff] [blame] | 70 | // 2-channel bias used by a number of Conv2d tests. |
telsoa01 | 4fcda01 | 2018-03-09 14:13:49 +0000 | [diff] [blame] | 71 | static std::vector<float> Bias2({0, 2}); |
| 72 | |
telsoa01 | c577f2c | 2018-08-31 09:22:23 +0100 | [diff] [blame] | 73 | // Helper function that returns either Bias2 or an empty vector depending on whether bias is enabled. |
telsoa01 | 4fcda01 | 2018-03-09 14:13:49 +0000 | [diff] [blame] | 74 | template<typename T> |
| 75 | boost::multi_array<T, 1> GetBias2(bool biasEnabled, float qScale, int32_t qOffset) |
| 76 | { |
| 77 | if(biasEnabled) |
| 78 | { |
| 79 | armnn::TensorInfo biasDesc({static_cast<unsigned int>(Bias2.size())}, armnn::GetDataType<T>()); |
| 80 | boost::multi_array<T, 1> bias = MakeTensor<T, 1>(biasDesc, QuantizedVector<T>(qScale, qOffset, Bias2)); |
| 81 | return bias; |
| 82 | } |
| 83 | else |
| 84 | { |
| 85 | return boost::multi_array<T, 1>(); |
| 86 | } |
| 87 | } |
| 88 | |
| 89 | template<typename T> |
| 90 | LayerTestResult<T, 4> SimpleConvolution2d3x5TestCommon(armnn::IWorkloadFactory& workloadFactory, |
| 91 | float qScale, |
| 92 | int32_t qOffset, |
| 93 | bool biasEnabled) |
| 94 | { |
telsoa01 | c577f2c | 2018-08-31 09:22:23 +0100 | [diff] [blame] | 95 | // Use common single-batch 3-channel 16x8 image. |
telsoa01 | 4fcda01 | 2018-03-09 14:13:49 +0000 | [diff] [blame] | 96 | armnn::TensorInfo inputDesc({1, 3, 8, 16}, armnn::GetDataType<T>()); |
| 97 | boost::multi_array<T, 4> input = MakeTensor<T, 4>(inputDesc, QuantizedVector<T>(qScale, qOffset, ConvInput3x8x16)); |
| 98 | |
telsoa01 | c577f2c | 2018-08-31 09:22:23 +0100 | [diff] [blame] | 99 | // Use a 2-element batch with 3-channel 3x5 kernels. |
telsoa01 | 4fcda01 | 2018-03-09 14:13:49 +0000 | [diff] [blame] | 100 | armnn::TensorInfo kernelDesc({2, 3, 5, 3}, armnn::GetDataType<T>()); |
| 101 | boost::multi_array<T, 4> kernel = MakeTensor<T, 4>(kernelDesc, std::vector<T>( |
| 102 | QuantizedVector<T>(qScale, qOffset, { |
| 103 | 1, 1, 1, |
| 104 | 1, -1, 1, |
| 105 | 1, 1, 1, |
| 106 | 1, 1, 1, |
| 107 | 1, 1, 1, |
| 108 | |
| 109 | 0, 0, 0, |
| 110 | 0, 0, 0, |
| 111 | 0, 0, 0, |
| 112 | 0, 0, 0, |
| 113 | 0, 0, 0, |
| 114 | |
| 115 | 2, 2, 2, |
| 116 | 2, 2, 2, |
| 117 | 2, 2, 2, |
| 118 | 2, 2, 2, |
| 119 | 2, 2, 2, |
| 120 | |
| 121 | |
| 122 | 0, 0, 0, |
| 123 | 0, 0, 0, |
| 124 | 0, 0, 0, |
| 125 | 0, 0, 0, |
| 126 | 0, 0, 0, |
| 127 | |
| 128 | 1, 1, 1, |
| 129 | 1, 1, 1, |
| 130 | 1, 1, 1, |
| 131 | 1, 1, 1, |
| 132 | 1, 1, 1, |
| 133 | |
| 134 | 0, 0, 0, |
| 135 | 0, 0, 0, |
| 136 | 0, 0, 0, |
| 137 | 0, 0, 0, |
| 138 | 0, 0, 0 |
| 139 | }))); |
| 140 | |
telsoa01 | c577f2c | 2018-08-31 09:22:23 +0100 | [diff] [blame] | 141 | // Expected output is 2 batch elements of a 1-channel 14x4 image. |
telsoa01 | 4fcda01 | 2018-03-09 14:13:49 +0000 | [diff] [blame] | 142 | armnn::TensorInfo outputDesc({1, 2, 4, 14}, armnn::GetDataType<T>()); |
| 143 | boost::multi_array<T, 4> expectedOutput = MakeTensor<T, 4>(outputDesc, std::vector<T>( |
| 144 | QuantizedVector<T>(qScale, qOffset, { |
| 145 | -24, -24, -24, -24, -24, -24, -24, -24, -24, -24, -24, -24, -24, -24, |
| 146 | -25, -25, -25, -25, -25, -25, -25, -25, -25, -25, -25, -25, -25, -25, |
| 147 | -23.5f, -23.5f, -23.5f, -23.5f, -23.5f, -23.5f, -23.5f, -23.5f, -23.5f, -23.5f, -23.5f, |
| 148 | -23.5f, -23.5f, -23.5f, |
| 149 | -23.5f, -23.5f, -23.5f, -23.5f, -23.5f, -23.5f, -23.5f, -23.5f, -23.5f, -23.5f, -23.5f, |
| 150 | -23.5f, -23.5f, -23.5f, |
| 151 | |
| 152 | 5, 5, 5, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, |
| 153 | 5, 5, 5, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, |
| 154 | 5, 5, 5, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, |
| 155 | 5, 5, 5, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0 |
| 156 | }))); |
| 157 | |
| 158 | return SimpleConvolution2dTestImpl<T>(workloadFactory, |
| 159 | input, |
| 160 | kernel, |
| 161 | GetBias2<typename FullyConnectedBiasTypeForInputType<T>::Type>(biasEnabled, qScale, qOffset), |
| 162 | expectedOutput, |
| 163 | qScale, |
| 164 | qOffset); |
| 165 | } |
| 166 | |
| 167 | template<typename T> |
| 168 | LayerTestResult<T, 4> SimpleConvolution2d3x3TestCommon(armnn::IWorkloadFactory& workloadFactory, |
| 169 | float qScale, |
| 170 | int32_t qOffset, |
| 171 | bool biasEnabled) |
| 172 | { |
telsoa01 | c577f2c | 2018-08-31 09:22:23 +0100 | [diff] [blame] | 173 | // Use a 3x3 kernel, which exercises ArmCompute's direct convolution path. |
telsoa01 | 4fcda01 | 2018-03-09 14:13:49 +0000 | [diff] [blame] | 174 | |
telsoa01 | c577f2c | 2018-08-31 09:22:23 +0100 | [diff] [blame] | 175 | // Use common single-batch 3-channel 16x8 image. |
telsoa01 | 4fcda01 | 2018-03-09 14:13:49 +0000 | [diff] [blame] | 176 | armnn::TensorInfo inputDesc({1, 3, 8, 16}, armnn::GetDataType<T>()); |
| 177 | boost::multi_array<T, 4> input = MakeTensor<T, 4>(inputDesc, QuantizedVector<T>(qScale, qOffset, ConvInput3x8x16)); |
| 178 | |
telsoa01 | c577f2c | 2018-08-31 09:22:23 +0100 | [diff] [blame] | 179 | // Use a 2-element batch of 3-channel 3x3 kernels. |
telsoa01 | 4fcda01 | 2018-03-09 14:13:49 +0000 | [diff] [blame] | 180 | armnn::TensorInfo kernelDesc({2, 3, 3, 3}, armnn::GetDataType<T>()); |
| 181 | boost::multi_array<T, 4> kernel = MakeTensor<T, 4>(kernelDesc, std::vector<T>( |
| 182 | QuantizedVector<T>(qScale, qOffset, { |
| 183 | 1, 1, 1, |
| 184 | 1, -1, 1, |
| 185 | 1, 1, 1, |
| 186 | |
| 187 | 0, 0, 0, |
| 188 | 0, 0, 0, |
| 189 | 0, 0, 0, |
| 190 | |
| 191 | 2, 2, 2, |
| 192 | 2, 2, 2, |
| 193 | 2, 2, 2, |
| 194 | |
| 195 | |
| 196 | 0, 0, 0, |
| 197 | 0, 0, 0, |
| 198 | 0, 0, 0, |
| 199 | |
| 200 | 1, 1, 1, |
| 201 | 1, 1, 1, |
| 202 | 1, 1, 1, |
| 203 | |
| 204 | 0, 0, 0, |
| 205 | 0, 0, 0, |
| 206 | 0, 0, 0 |
| 207 | }))); |
| 208 | |
telsoa01 | c577f2c | 2018-08-31 09:22:23 +0100 | [diff] [blame] | 209 | // Expected output is 1 batch of a 2-channel 14x6 image. |
telsoa01 | 4fcda01 | 2018-03-09 14:13:49 +0000 | [diff] [blame] | 210 | armnn::TensorInfo outputDesc({1, 2, 6, 14}, armnn::GetDataType<T>()); |
| 211 | boost::multi_array<T, 4> expectedOutput = MakeTensor<T, 4>(outputDesc, std::vector<T>( |
| 212 | QuantizedVector<T>(qScale, qOffset, { |
| 213 | -15, -15, -15, -15, -15, -15, -15, -15, -15, -15, -15, -15, -15, -15, |
| 214 | -16, -16, -16, -16, -16, -16, -16, -16, -16, -16, -16, -16, -16, -16, |
| 215 | -14.5f,-14.5f,-14.5f,-14.5f,-14.5f,-14.5f,-14.5f,-14.5f,-14.5f,-14.5f,-14.5f,-14.5f,-14.5f,-14.5f, |
| 216 | -14.5f,-14.5f,-14.5f,-14.5f,-14.5f,-14.5f,-14.5f,-14.5f,-14.5f,-14.5f,-14.5f,-14.5f,-14.5f,-14.5f, |
| 217 | -14.5f,-14.5f,-14.5f,-14.5f,-14.5f,-14.5f,-14.5f,-14.5f,-14.5f,-14.5f,-14.5f,-14.5f,-14.5f,-14.5f, |
| 218 | -14.5f,-14.5f,-14.5f,-14.5f,-14.5f,-14.5f,-14.5f,-14.5f,-14.5f,-14.5f,-14.5f,-14.5f,-14.5f,-14.5f, |
| 219 | |
| 220 | 3, 3, 3, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, |
| 221 | 3, 3, 3, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, |
| 222 | 3, 3, 3, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, |
| 223 | 3, 3, 3, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, |
| 224 | 3, 3, 3, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, |
| 225 | 3, 3, 3, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0 |
| 226 | }))); |
| 227 | |
| 228 | return SimpleConvolution2dTestImpl<T>(workloadFactory, |
| 229 | input, |
| 230 | kernel, |
| 231 | GetBias2<typename FullyConnectedBiasTypeForInputType<T>::Type>(biasEnabled, qScale, qOffset), |
| 232 | expectedOutput, |
| 233 | qScale, |
| 234 | qOffset); |
| 235 | } |
| 236 | |
| 237 | LayerTestResult<float, 4> SimpleConvolution2d3x5Test(armnn::IWorkloadFactory& workloadFactory, |
| 238 | bool biasEnabled) |
| 239 | { |
| 240 | return SimpleConvolution2d3x5TestCommon<float>(workloadFactory, 0.f, 0, biasEnabled); |
| 241 | } |
| 242 | |
| 243 | LayerTestResult<uint8_t, 4> SimpleConvolution2d3x5Uint8Test(armnn::IWorkloadFactory& workloadFactory, |
| 244 | bool biasEnabled) |
| 245 | { |
| 246 | return SimpleConvolution2d3x5TestCommon<uint8_t>(workloadFactory, 0.5f, 50, biasEnabled); |
| 247 | } |
| 248 | |
| 249 | LayerTestResult<float, 4> SimpleConvolution2d3x3Test(armnn::IWorkloadFactory& workloadFactory, |
| 250 | bool biasEnabled) |
| 251 | { |
| 252 | return SimpleConvolution2d3x3TestCommon<float>(workloadFactory, 0.f, 0, biasEnabled); |
| 253 | } |
| 254 | |
| 255 | LayerTestResult<uint8_t, 4> SimpleConvolution2d3x3Uint8Test(armnn::IWorkloadFactory& workloadFactory, |
| 256 | bool biasEnabled) |
| 257 | { |
| 258 | return SimpleConvolution2d3x3TestCommon<uint8_t>(workloadFactory, 0.5f, 50, biasEnabled); |
| 259 | } |
| 260 | |
| 261 | template<typename T> |
| 262 | LayerTestResult<T, 4> Convolution2dAsymmetricPaddingLargerThanHalfKernelSizeTestCommon( |
| 263 | armnn::IWorkloadFactory& workloadFactory, |
| 264 | float qScale, |
| 265 | int32_t qOffset) |
| 266 | { |
telsoa01 | c577f2c | 2018-08-31 09:22:23 +0100 | [diff] [blame] | 267 | // Use a single-batch 1-channel 3x3 image as input. |
telsoa01 | 4fcda01 | 2018-03-09 14:13:49 +0000 | [diff] [blame] | 268 | armnn::TensorInfo inputDesc({1, 1, 3, 3}, armnn::GetDataType<T>()); |
| 269 | boost::multi_array<T, 4> input = MakeTensor<T, 4>(inputDesc, std::vector<T>( |
| 270 | QuantizedVector<T>(qScale, qOffset, { |
| 271 | 11,21,31, |
| 272 | 12,22,32, |
| 273 | 13,23,33 |
| 274 | }))); |
| 275 | |
telsoa01 | c577f2c | 2018-08-31 09:22:23 +0100 | [diff] [blame] | 276 | // Use 1 batch of a 1-channel 2x2 kernel. |
telsoa01 | 4fcda01 | 2018-03-09 14:13:49 +0000 | [diff] [blame] | 277 | armnn::TensorInfo kernelDesc({1, 1, 2, 2}, armnn::GetDataType<T>()); |
| 278 | boost::multi_array<T, 4> kernel = MakeTensor<T, 4>(kernelDesc, std::vector<T>( |
| 279 | QuantizedVector<T>(qScale, qOffset, { |
| 280 | -11,-21, |
| 281 | -12,-22, |
| 282 | }))); |
| 283 | |
telsoa01 | c577f2c | 2018-08-31 09:22:23 +0100 | [diff] [blame] | 284 | // Expected output is 1 batch of a 1-channel 6x8 image. |
telsoa01 | 4fcda01 | 2018-03-09 14:13:49 +0000 | [diff] [blame] | 285 | // Manually calculated like this: |
| 286 | //[-11*0 -21*0 -12*0 -22*0 ; -11*0 -21*0 -12*0 -22*0 ; -11*0 -21*0 -12*0 -22*0 ; -11*0 -21*0 -12*0 -22*0 ..] |
| 287 | //[-11*0 -21*0 -12*0 -22*11 ; -11*0 -21*0 -12*11 -22*21 ; -11*0 -21*0 -12*21 -22*31 ; -11*0 -21*0 -12*31 -22*0 ..] |
| 288 | //[-11*0 -21*11 -12*0 -22*12 ; -11*11 -21*21 -12*12 -22*22 ; -11*21 -21*31 -12*22 -22*32 ; -11*31 -21*0 -12*32 -22*0 ..] |
| 289 | //[-11*0 -21*12 -12*0 -22*13 ; -11*12 -21*22 -12*13 -22*23 ; -11*22 -21*32 -12*23 -22*33 ; -11*32 -21*0 -12*33 -22*0 ..] |
| 290 | //[-11*0 -21*13 -12*0 -22*0 ; -11*13 -21*23 -12*0 -22*0 ; -11*23 -21*33 -12*0 -22*0 ; -11*33 -21*0 -12*0 -22*0 ..] |
| 291 | //[-11*0 -21*0 -12*0 -22*0 ; -11*0 -21*0 -12*0 -22*0 ; -11*0 -21*0 -12*0 -22*0 ; -11*0 -21*0 -12*0 -22*0 ..] |
| 292 | //[..... ..... ..... ..... ; ..... ..... ..... ..... ; ..... ..... ..... ..... ; ..... ..... ..... ..... ..] |
| 293 | armnn::TensorInfo outputDesc({1, 1, 8, 6}, armnn::GetDataType<T>()); |
| 294 | boost::multi_array<T, 4> expectedOutput = MakeTensor<T, 4>(outputDesc, std::vector<T>( |
| 295 | QuantizedVector<T>(qScale, qOffset, { |
| 296 | 0, 0, 0, 0, 0, 0, |
| 297 | -242, -594, -934, -372, 0, 0, |
| 298 | -495, -1190, -1850, -725, 0, 0, |
| 299 | -538, -1256, -1916, -748, 0, 0, |
| 300 | -273, -626, -946, -363, 0, 0, |
| 301 | 0, 0, 0, 0, 0, 0, |
| 302 | 0, 0, 0, 0, 0, 0, |
| 303 | 0, 0, 0, 0, 0, 0 |
| 304 | }))); |
| 305 | |
| 306 | return SimpleConvolution2dTestImpl<T>(workloadFactory, |
| 307 | input, |
| 308 | kernel, |
| 309 | GetBias2<typename FullyConnectedBiasTypeForInputType<T>::Type>(false, qScale, qOffset), |
| 310 | expectedOutput, |
| 311 | qScale, |
| 312 | qOffset, |
telsoa01 | c577f2c | 2018-08-31 09:22:23 +0100 | [diff] [blame] | 313 | 1, // Padding left. |
| 314 | 2, // Padding top. |
| 315 | 3, // Padding right. |
| 316 | 4); // Padding bottom. |
telsoa01 | 4fcda01 | 2018-03-09 14:13:49 +0000 | [diff] [blame] | 317 | } |
| 318 | |
| 319 | template<typename T> |
| 320 | LayerTestResult<T, 4> SimpleConvolution2dAsymmetricPaddingTestCommon(armnn::IWorkloadFactory& workloadFactory, |
| 321 | float qScale, |
| 322 | int32_t qOffset) |
| 323 | { |
telsoa01 | c577f2c | 2018-08-31 09:22:23 +0100 | [diff] [blame] | 324 | // Use a single-batch 1-channel 5x5 image as input. |
telsoa01 | 4fcda01 | 2018-03-09 14:13:49 +0000 | [diff] [blame] | 325 | armnn::TensorInfo inputDesc({ 1, 1, 5, 5 }, armnn::GetDataType<T>()); |
| 326 | boost::multi_array<T, 4> input = MakeTensor<T, 4>(inputDesc, std::vector<T>( |
| 327 | QuantizedVector<T>(qScale, qOffset, { |
| 328 | 11,21,31,41,51, |
| 329 | 12,22,32,42,52, |
| 330 | 13,23,33,43,53, |
| 331 | 14,24,34,44,54, |
| 332 | 15,25,35,45,55, |
| 333 | }))); |
| 334 | |
telsoa01 | c577f2c | 2018-08-31 09:22:23 +0100 | [diff] [blame] | 335 | // Use 1 batch of a 1-channel 4x4 kernel. |
telsoa01 | 4fcda01 | 2018-03-09 14:13:49 +0000 | [diff] [blame] | 336 | armnn::TensorInfo kernelDesc({ 1, 1, 4, 4 }, armnn::GetDataType<T>()); |
| 337 | boost::multi_array<T, 4> kernel = MakeTensor<T, 4>(kernelDesc, std::vector<T>( |
| 338 | QuantizedVector<T>(qScale, qOffset, { |
| 339 | -11,-21,-31,-41, |
| 340 | -12,-22,-32,-42, |
| 341 | -13,-23,-33,-43, |
| 342 | -14,-24,-34,-44, |
| 343 | }))); |
| 344 | |
telsoa01 | c577f2c | 2018-08-31 09:22:23 +0100 | [diff] [blame] | 345 | // Expected output is 1 batch of a 1-channel 5x5 image. |
telsoa01 | 4fcda01 | 2018-03-09 14:13:49 +0000 | [diff] [blame] | 346 | armnn::TensorInfo outputDesc({ 1, 1, 5, 5 }, armnn::GetDataType<T>()); |
| 347 | std::vector<T> myVec(outputDesc.GetNumElements(), 0); |
| 348 | boost::multi_array<T, 4> expectedOutput = MakeTensor<T, 4>(outputDesc, std::vector<T>( |
| 349 | QuantizedVector<T>(qScale, qOffset, { |
telsoa01 | 4fcda01 | 2018-03-09 14:13:49 +0000 | [diff] [blame] | 350 | -7140, -10580, -13940, -9300, -5230, |
| 351 | -9590, -14120, -18520, -12290, -6860, |
| 352 | -9980, -14560, -18960, -12560, -7000, |
| 353 | -7518, -10904, -14144, -9318, -5152, |
surmeh01 | 3537c2c | 2018-05-18 16:31:43 +0100 | [diff] [blame] | 354 | -5032, -7256, -9376, -6142, -3368, |
telsoa01 | 4fcda01 | 2018-03-09 14:13:49 +0000 | [diff] [blame] | 355 | }))); |
| 356 | |
| 357 | return SimpleConvolution2dTestImpl<T>(workloadFactory, |
| 358 | input, |
| 359 | kernel, |
| 360 | GetBias2<typename FullyConnectedBiasTypeForInputType<T>::Type>(false, qScale, qOffset), |
| 361 | expectedOutput, |
| 362 | qScale, |
| 363 | qOffset, |
telsoa01 | c577f2c | 2018-08-31 09:22:23 +0100 | [diff] [blame] | 364 | 1, // Padding left. |
| 365 | 1, // Padding top. |
| 366 | 2, // Padding right. |
| 367 | 2); // Padding bottom. |
surmeh01 | 3537c2c | 2018-05-18 16:31:43 +0100 | [diff] [blame] | 368 | } |
| 369 | |
| 370 | template<typename T> |
| 371 | LayerTestResult<T, 4> DepthwiseConvolution2dAsymmetricTestCommon(armnn::IWorkloadFactory& workloadFactory, |
| 372 | float qScale, |
| 373 | int32_t qOffset, |
| 374 | bool biasEnabled) |
| 375 | { |
telsoa01 | c577f2c | 2018-08-31 09:22:23 +0100 | [diff] [blame] | 376 | // Use a single-batch 2-channel 5x5 image as input. |
surmeh01 | 3537c2c | 2018-05-18 16:31:43 +0100 | [diff] [blame] | 377 | armnn::TensorInfo inputTensorInfo({ 1, 2, 5, 5 }, armnn::GetDataType<T>()); |
| 378 | auto input = MakeTensor<T, 4>(inputTensorInfo, std::vector<T>( |
| 379 | QuantizedVector<T>(inputTensorInfo.GetQuantizationScale(), inputTensorInfo.GetQuantizationOffset(), { |
| 380 | 0, 1, 2, 3, 4, |
| 381 | 5, 6, 7, 8, 9, |
| 382 | 10, 11, 12, 13, 14, |
| 383 | 15, 16, 17, 18, 19, |
| 384 | 20, 21, 22, 23, 24, |
| 385 | |
| 386 | 25, 26, 27, 28, 29, |
| 387 | 30, 31, 32, 33, 34, |
| 388 | 35, 36, 37, 38, 39, |
| 389 | 40, 41, 42, 43, 44, |
| 390 | 45, 46, 47, 48, 49 |
| 391 | }))); |
| 392 | |
telsoa01 | c577f2c | 2018-08-31 09:22:23 +0100 | [diff] [blame] | 393 | // Use a depth multiplier of 1 on a 2-channel 4x4 kernel. |
surmeh01 | 3537c2c | 2018-05-18 16:31:43 +0100 | [diff] [blame] | 394 | armnn::TensorInfo kernelTensorInfo({ 1, 2, 4, 4 }, armnn::GetDataType<T>()); |
| 395 | auto kernel = MakeTensor<T, 4>(kernelTensorInfo, std::vector<T>( |
| 396 | QuantizedVector<T>(kernelTensorInfo.GetQuantizationScale(), kernelTensorInfo.GetQuantizationOffset(), { |
| 397 | 32, 31, 30, 29, |
| 398 | 28, 27, 26, 25, |
| 399 | 24, 23, 22, 21, |
| 400 | 20, 19, 18, 17, |
| 401 | |
| 402 | 16, 15, 14, 13, |
| 403 | 12, 11, 10, 9, |
| 404 | 8, 7, 6, 5, |
| 405 | 4, 3, 2, 1 |
| 406 | }))); |
| 407 | |
telsoa01 | c577f2c | 2018-08-31 09:22:23 +0100 | [diff] [blame] | 408 | // Expected output is 1 batch of a 2-channel 5x5 image. |
| 409 | // Calculated using the python tensorflow library with strideX=1, strideY=1. |
surmeh01 | 3537c2c | 2018-05-18 16:31:43 +0100 | [diff] [blame] | 410 | armnn::TensorInfo outputTensorInfo({ 1, 2, 5, 5 }, armnn::GetDataType<T>()); |
| 411 | boost::multi_array<T, 4> expectedOutput = MakeTensor<T, 4>(outputTensorInfo, std::vector<T>( |
| 412 | QuantizedVector<T>(outputTensorInfo.GetQuantizationScale(), outputTensorInfo.GetQuantizationOffset(), { |
| 413 | 1062, 1580, 1850, 1530, 1117, |
| 414 | 2140, 3108, 3500, 2842, 2042, |
| 415 | 3580, 5068, 5460, 4342, 3062, |
| 416 | 3618, 5072, 5390, 4248, 2971, |
| 417 | 3074, 4282, 4510, 3533, 2457, |
| 418 | 1550, 2284, 2362, 1955, 1428, |
| 419 | 2910, 4206, 4342, 3528, 2536, |
| 420 | 3390, 4886, 5022, 4068, 2916, |
| 421 | 3566, 5056, 5182, 4133, 2922, |
| 422 | 3100, 4352, 4452, 3517, 2465 |
| 423 | }))); |
| 424 | |
| 425 | return DepthwiseConvolution2dAsymmetricTestImpl<T>(workloadFactory, |
| 426 | input, |
| 427 | kernel, |
| 428 | GetBias2<typename FullyConnectedBiasTypeForInputType<T>::Type>(biasEnabled, qScale, qOffset), |
| 429 | expectedOutput, |
| 430 | qScale, |
| 431 | qOffset, |
telsoa01 | c577f2c | 2018-08-31 09:22:23 +0100 | [diff] [blame] | 432 | 1, // Padding left. |
| 433 | 1, // Padding top. |
| 434 | 2, // Padding right. |
| 435 | 2, // Padding bottom. |
surmeh01 | 3537c2c | 2018-05-18 16:31:43 +0100 | [diff] [blame] | 436 | 1, // strideX |
| 437 | 1); // strideY |
telsoa01 | 4fcda01 | 2018-03-09 14:13:49 +0000 | [diff] [blame] | 438 | } |
| 439 | |
| 440 | LayerTestResult<float, 4> |
| 441 | Convolution2dAsymmetricPaddingLargerThanHalfKernelSizeTest(armnn::IWorkloadFactory& workloadFactory) |
| 442 | { |
| 443 | return Convolution2dAsymmetricPaddingLargerThanHalfKernelSizeTestCommon<float>(workloadFactory, 0.0f, 0); |
| 444 | } |
| 445 | |
| 446 | LayerTestResult<float, 4> Convolution2dAsymmetricPaddingTest(armnn::IWorkloadFactory& workloadFactory) |
| 447 | { |
| 448 | return SimpleConvolution2dAsymmetricPaddingTestCommon<float>(workloadFactory, 0.0f, 0); |
| 449 | } |
| 450 | |
| 451 | LayerTestResult<float, 4> DepthwiseConvolution2dTest(armnn::IWorkloadFactory& workloadFactory, |
| 452 | bool biasEnabled) |
| 453 | { |
| 454 | return DepthwiseConvolution2dTestImpl<float, float>(workloadFactory, 0.0f, 0, biasEnabled); |
| 455 | } |
| 456 | |
| 457 | LayerTestResult<float, 4> DepthwiseConvolution2dDepthMul1Test(armnn::IWorkloadFactory& workloadFactory, |
| 458 | bool biasEnabled) |
| 459 | { |
| 460 | return DepthwiseConvolution2dDepthMul1TestImpl<float, float>(workloadFactory, 0.0f, 0, biasEnabled); |
| 461 | } |
| 462 | |
surmeh01 | 3537c2c | 2018-05-18 16:31:43 +0100 | [diff] [blame] | 463 | LayerTestResult<float, 4> DepthwiseConvolution2dAsymmetricTest(armnn::IWorkloadFactory& workloadFactory, |
| 464 | bool biasEnabled) |
| 465 | { |
| 466 | return DepthwiseConvolution2dAsymmetricTestCommon<float>(workloadFactory, 0.0f, 0, biasEnabled); |
| 467 | } |
| 468 | |
telsoa01 | 4fcda01 | 2018-03-09 14:13:49 +0000 | [diff] [blame] | 469 | LayerTestResult<uint8_t, 4> DepthwiseConvolution2dUint8Test(armnn::IWorkloadFactory& workloadFactory, |
| 470 | bool biasEnabled) |
| 471 | { |
| 472 | return DepthwiseConvolution2dTestImpl<uint8_t, int32_t>(workloadFactory, 0.5f, 50, biasEnabled); |
| 473 | } |
| 474 | |
| 475 | LayerTestResult<uint8_t, 4> DepthwiseConvolution2dDepthMul1Uint8Test(armnn::IWorkloadFactory& workloadFactory, |
| 476 | bool biasEnabled) |
| 477 | { |
| 478 | return DepthwiseConvolution2dDepthMul1TestImpl<uint8_t, int32_t>(workloadFactory, 0.5f, 50, biasEnabled); |
| 479 | } |
| 480 | |
| 481 | LayerTestResult<float, 4> Convolution1dTest(armnn::IWorkloadFactory& workloadFactory, bool biasEnabled) |
| 482 | { |
| 483 | return Convolution1dTestImpl<float>(workloadFactory, 0.0f, 0, biasEnabled); |
| 484 | } |
| 485 | |
| 486 | LayerTestResult<uint8_t, 4> Convolution1dUint8Test(armnn::IWorkloadFactory& workloadFactory, bool biasEnabled) |
| 487 | { |
| 488 | return Convolution1dTestImpl<uint8_t>(workloadFactory, 0.1f, 128, biasEnabled); |
| 489 | } |
| 490 | |
| 491 | LayerTestResult<float,4> CompareConvolution2dTest(armnn::IWorkloadFactory& workloadFactory, |
| 492 | armnn::IWorkloadFactory& refWorkloadFactory) |
| 493 | { |
| 494 | return CompareConvolution2dTestImpl<float>(workloadFactory, refWorkloadFactory); |
| 495 | } |
| 496 | |
| 497 | template<typename T> |
| 498 | LayerTestResult<T,4> CompareDepthwiseConvolution2dTest(armnn::IWorkloadFactory& workloadFactory, |
| 499 | armnn::IWorkloadFactory& refWorkloadFactory) |
| 500 | { |
| 501 | return CompareDepthwiseConvolution2dTestImpl<T>(workloadFactory, refWorkloadFactory); |
| 502 | } |
| 503 | |
| 504 | template LayerTestResult<float, 4> CompareDepthwiseConvolution2dTest<float>( |
| 505 | armnn::IWorkloadFactory&, armnn::IWorkloadFactory&); |
| 506 | template LayerTestResult<uint8_t, 4> CompareDepthwiseConvolution2dTest<uint8_t>( |
| 507 | armnn::IWorkloadFactory&, armnn::IWorkloadFactory&); |
| 508 | |
| 509 | LayerTestResult<float,4> SimpleNormalizationAcrossTest(armnn::IWorkloadFactory& workloadFactory) |
| 510 | { |
| 511 | auto normMethod = armnn::NormalizationAlgorithmMethod::LocalBrightness; |
| 512 | auto normChannel = armnn::NormalizationAlgorithmChannel::Across; |
| 513 | return SimpleNormalizationTestImpl(workloadFactory, normChannel, normMethod); |
| 514 | } |
| 515 | |
| 516 | LayerTestResult<float,4> SimpleNormalizationWithinTest(armnn::IWorkloadFactory& workloadFactory) |
| 517 | { |
| 518 | auto normMethod = armnn::NormalizationAlgorithmMethod::LocalBrightness; |
| 519 | auto normChannel = armnn::NormalizationAlgorithmChannel::Within; |
| 520 | return SimpleNormalizationTestImpl(workloadFactory, normChannel, normMethod); |
| 521 | } |
| 522 | |
narpra01 | 55a97bc | 2018-10-02 14:35:53 +0100 | [diff] [blame] | 523 | LayerTestResult<float,4> SimpleNormalizationAcrossNhwcTest(armnn::IWorkloadFactory& workloadFactory) |
| 524 | { |
| 525 | auto normMethod = armnn::NormalizationAlgorithmMethod::LocalBrightness; |
| 526 | auto normChannel = armnn::NormalizationAlgorithmChannel::Across; |
| 527 | return SimpleNormalizationNhwcClNeonTestImpl(workloadFactory, normChannel, normMethod); |
| 528 | } |
| 529 | |
telsoa01 | 4fcda01 | 2018-03-09 14:13:49 +0000 | [diff] [blame] | 530 | LayerTestResult<float,2> SimpleSoftmaxTest(armnn::IWorkloadFactory& workloadFactory, float beta) |
| 531 | { |
| 532 | return SimpleSoftmaxTestImpl<float>(workloadFactory, beta); |
| 533 | } |
| 534 | |
| 535 | LayerTestResult<uint8_t,2> SimpleSoftmaxUint8Test(armnn::IWorkloadFactory& workloadFactory, float beta) |
| 536 | { |
| 537 | return SimpleSoftmaxTestImpl<uint8_t>(workloadFactory, beta); |
| 538 | } |
| 539 | |
| 540 | LayerTestResult<float,4> CompareNormalizationTest(armnn::IWorkloadFactory& workloadFactory, |
| 541 | armnn::IWorkloadFactory& refWorkloadFactory, |
| 542 | armnn::NormalizationAlgorithmChannel normChannel, |
| 543 | armnn::NormalizationAlgorithmMethod normMethod) |
| 544 | { |
| 545 | return CompareNormalizationTestImpl(workloadFactory, refWorkloadFactory, normChannel, normMethod); |
| 546 | } |
| 547 | |
| 548 | LayerTestResult<float,2> CompareSoftmaxTest(armnn::IWorkloadFactory& workloadFactory, |
| 549 | armnn::IWorkloadFactory& refWorkloadFactory, |
| 550 | float beta) |
| 551 | { |
| 552 | return CompareSoftmaxTestImpl<float>(workloadFactory, refWorkloadFactory, beta); |
| 553 | } |
| 554 | |
| 555 | LayerTestResult<uint8_t,2> CompareSoftmaxUint8Test(armnn::IWorkloadFactory& workloadFactory, |
| 556 | armnn::IWorkloadFactory& refWorkloadFactory, |
| 557 | float beta) |
| 558 | { |
| 559 | return CompareSoftmaxTestImpl<uint8_t>(workloadFactory, refWorkloadFactory, beta); |
| 560 | } |
| 561 | |
| 562 | std::vector<LayerTestResult<float,3>> SplitterTest(armnn::IWorkloadFactory& workloadFactory) |
| 563 | { |
| 564 | return SplitterTestCommon<float>(workloadFactory); |
| 565 | } |
| 566 | |
| 567 | std::vector<LayerTestResult<uint8_t,3>> SplitterUint8Test(armnn::IWorkloadFactory& workloadFactory) |
| 568 | { |
| 569 | return SplitterTestCommon<uint8_t>(workloadFactory, 1.0f, 0); |
| 570 | } |
| 571 | |
| 572 | LayerTestResult<float, 3> CopyViaSplitterTest(armnn::IWorkloadFactory& workloadFactory) |
| 573 | { |
| 574 | return CopyViaSplitterTestImpl<float>(workloadFactory, 0.0f, 0); |
| 575 | } |
| 576 | |
| 577 | LayerTestResult<uint8_t, 3> CopyViaSplitterUint8Test(armnn::IWorkloadFactory& workloadFactory) |
| 578 | { |
| 579 | return CopyViaSplitterTestImpl<uint8_t>(workloadFactory, 1.0f, 0); |
| 580 | } |
| 581 | |
telsoa01 | c577f2c | 2018-08-31 09:22:23 +0100 | [diff] [blame] | 582 | LayerTestResult<float, 2> LstmLayerFloat32WithCifgWithPeepholeNoProjectionTest( |
| 583 | armnn::IWorkloadFactory& workloadFactory) |
| 584 | { |
| 585 | armnn::TensorInfo inputDesc({ 2, 2 }, armnn::GetDataType<float>()); |
| 586 | boost::multi_array<float, 2> input = MakeTensor<float, 2>(inputDesc, std::vector<float>( |
| 587 | { 2., 3., 3., 4. })); |
| 588 | |
| 589 | armnn::TensorInfo outputDesc({ 2, 4 }, armnn::GetDataType<float>()); |
| 590 | boost::multi_array<float, 2> expectedOutput = MakeTensor<float, 2>(outputDesc, std::vector<float>( |
| 591 | {-0.36444446f, -0.00352185f, 0.12886585f, -0.05163646f, |
| 592 | -0.42734814f, -0.00478661f, 0.13455015f, -0.03560682f})); |
| 593 | return LstmLayerWithCifgWithPeepholeNoProjectionTestImpl(workloadFactory, input, expectedOutput); |
| 594 | } |
| 595 | |
| 596 | LayerTestResult<float, 2> LstmLayerFloat32NoCifgWithPeepholeWithProjectionTest( |
| 597 | armnn::IWorkloadFactory& workloadFactory) |
| 598 | { |
| 599 | armnn::TensorInfo inputDesc({ 2, 5 }, armnn::GetDataType<float>()); |
| 600 | boost::multi_array<float, 2> input = MakeTensor<float, 2>(inputDesc, std::vector<float>( |
| 601 | {0.787926f, 0.151646f, 0.071352f, 0.118426f, 0.458058f, |
| 602 | 0.295743f, 0.544053f, 0.690064f, 0.858138f, 0.497181f})); |
| 603 | |
| 604 | armnn::TensorInfo outputDesc({ 2, 16 }, armnn::GetDataType<float>()); |
| 605 | boost::multi_array<float, 2> expectedOutput = MakeTensor<float, 2>(outputDesc, std::vector<float>( |
| 606 | {-0.00396806f, 0.029352f, -0.00279226f, 0.0159977f, -0.00835576f, |
| 607 | -0.0211779f, 0.0283512f, -0.0114597f, 0.00907307f, -0.0244004f, |
| 608 | -0.0152191f, -0.0259063f, 0.00914318f, 0.00415118f, 0.017147f, |
| 609 | 0.0134203f, -0.013869f, 0.0287268f, -0.00334693f, 0.00733398f, -0.0287926f, |
| 610 | -0.0186926f, 0.0193662f, -0.0115437f, 0.00422612f, -0.0345232f, |
| 611 | 0.00223253f, -0.00957321f, 0.0210624f, 0.013331f, 0.0150954f, |
| 612 | 0.02168f})); |
| 613 | return LstmLayerFloat32NoCifgWithPeepholeWithProjectionTestImpl(workloadFactory, input, expectedOutput); |
| 614 | } |
| 615 | |
| 616 | LayerTestResult<float, 2> LstmLayerFloat32NoCifgNoPeepholeNoProjectionTest(armnn::IWorkloadFactory& workloadFactory) |
| 617 | { |
| 618 | armnn::TensorInfo inputDesc({2, 2}, armnn::GetDataType<float>()); |
| 619 | boost::multi_array<float, 2> input = MakeTensor<float, 2>(inputDesc, std::vector<float>( |
| 620 | {2., 3., 3., 4.})); |
| 621 | |
| 622 | |
| 623 | armnn::TensorInfo outputDesc({2, 4}, armnn::GetDataType<float>()); |
| 624 | boost::multi_array<float, 2> expectedOutput = MakeTensor<float, 2>(outputDesc, std::vector<float>( |
| 625 | {{-0.02973187f, 0.1229473f, 0.20885126f, -0.15358765f, |
| 626 | -0.0185422f, 0.11281417f, 0.24466537f, -0.1826292f}})); |
| 627 | |
| 628 | return LstmNoCifgNoPeepholeNoProjectionTestImpl(workloadFactory, input, expectedOutput); |
| 629 | } |
| 630 | |
telsoa01 | 4fcda01 | 2018-03-09 14:13:49 +0000 | [diff] [blame] | 631 | LayerTestResult<float,3> MergerTest(armnn::IWorkloadFactory& workloadFactory) |
| 632 | { |
surmeh01 | 3537c2c | 2018-05-18 16:31:43 +0100 | [diff] [blame] | 633 | unsigned int outputWidth = 3; |
telsoa01 | 4fcda01 | 2018-03-09 14:13:49 +0000 | [diff] [blame] | 634 | unsigned int outputHeight = 6; |
| 635 | unsigned int outputChannels = 3; |
| 636 | |
surmeh01 | 3537c2c | 2018-05-18 16:31:43 +0100 | [diff] [blame] | 637 | unsigned int inputWidth1 = 3; |
| 638 | unsigned int inputHeight1 = 6; |
| 639 | unsigned int inputChannels1 = 2; |
telsoa01 | 4fcda01 | 2018-03-09 14:13:49 +0000 | [diff] [blame] | 640 | |
surmeh01 | 3537c2c | 2018-05-18 16:31:43 +0100 | [diff] [blame] | 641 | unsigned int inputWidth2 = 3; |
| 642 | unsigned int inputHeight2 = 6; |
| 643 | unsigned int inputChannels2 = 1; |
telsoa01 | 4fcda01 | 2018-03-09 14:13:49 +0000 | [diff] [blame] | 644 | |
telsoa01 | c577f2c | 2018-08-31 09:22:23 +0100 | [diff] [blame] | 645 | // Define the tensor descriptors. |
telsoa01 | 4fcda01 | 2018-03-09 14:13:49 +0000 | [diff] [blame] | 646 | armnn::TensorInfo outputTensorInfo({ outputChannels, outputHeight, outputWidth }, armnn::DataType::Float32); |
| 647 | armnn::TensorInfo inputTensorInfo1({ inputChannels1, inputHeight1, inputWidth1 }, armnn::DataType::Float32); |
| 648 | armnn::TensorInfo inputTensorInfo2({ inputChannels2, inputHeight2, inputWidth2 }, armnn::DataType::Float32); |
telsoa01 | 4fcda01 | 2018-03-09 14:13:49 +0000 | [diff] [blame] | 649 | |
| 650 | LayerTestResult<float,3> ret(outputTensorInfo); |
| 651 | |
telsoa01 | 4fcda01 | 2018-03-09 14:13:49 +0000 | [diff] [blame] | 652 | ret.outputExpected = MakeTensor<float, 3>(outputTensorInfo, std::vector<float>( |
surmeh01 | 3537c2c | 2018-05-18 16:31:43 +0100 | [diff] [blame] | 653 | { |
| 654 | 1.0f, 2.0f, 3.0f, |
| 655 | 4.0f, 5.0f, 6.0f, |
| 656 | 7.0f, 8.0f, 9.0f, |
| 657 | 10.0f, 11.0f, 12.0f, |
| 658 | 13.0f, 14.0f, 15.0f, |
| 659 | 16.0f, 17.0f, 18.0f, |
telsoa01 | 4fcda01 | 2018-03-09 14:13:49 +0000 | [diff] [blame] | 660 | |
surmeh01 | 3537c2c | 2018-05-18 16:31:43 +0100 | [diff] [blame] | 661 | 19.0f, 20.0f, 21.0f, |
| 662 | 22.0f, 23.0f, 24.0f, |
| 663 | 25.0f, 26.0f, 27.0f, |
| 664 | 28.0f, 29.0f, 30.0f, |
| 665 | 31.0f, 32.0f, 33.0f, |
| 666 | 34.0f, 35.0f, 36.0f, |
telsoa01 | 4fcda01 | 2018-03-09 14:13:49 +0000 | [diff] [blame] | 667 | |
surmeh01 | 3537c2c | 2018-05-18 16:31:43 +0100 | [diff] [blame] | 668 | 37.0f, 38.0f, 39.0f, |
| 669 | 40.0f, 41.0f, 42.0f, |
| 670 | 43.0f, 44.0f, 45.0f, |
| 671 | 46.0f, 47.0f, 48.0f, |
| 672 | 49.0f, 50.0f, 51.0f, |
| 673 | 52.0f, 53.0f, 54.0f, |
telsoa01 | 4fcda01 | 2018-03-09 14:13:49 +0000 | [diff] [blame] | 674 | }) |
| 675 | ); |
| 676 | |
telsoa01 | 4fcda01 | 2018-03-09 14:13:49 +0000 | [diff] [blame] | 677 | auto input1 = MakeTensor<float, 3>(inputTensorInfo1, std::vector<float>( |
| 678 | { |
surmeh01 | 3537c2c | 2018-05-18 16:31:43 +0100 | [diff] [blame] | 679 | 1.0f, 2.0f, 3.0f, |
| 680 | 4.0f, 5.0f, 6.0f, |
| 681 | 7.0f, 8.0f, 9.0f, |
| 682 | 10.0f, 11.0f, 12.0f, |
| 683 | 13.0f, 14.0f, 15.0f, |
| 684 | 16.0f, 17.0f, 18.0f, |
telsoa01 | 4fcda01 | 2018-03-09 14:13:49 +0000 | [diff] [blame] | 685 | |
surmeh01 | 3537c2c | 2018-05-18 16:31:43 +0100 | [diff] [blame] | 686 | 19.0f, 20.0f, 21.0f, |
| 687 | 22.0f, 23.0f, 24.0f, |
| 688 | 25.0f, 26.0f, 27.0f, |
| 689 | 28.0f, 29.0f, 30.0f, |
| 690 | 31.0f, 32.0f, 33.0f, |
| 691 | 34.0f, 35.0f, 36.0f, |
telsoa01 | 4fcda01 | 2018-03-09 14:13:49 +0000 | [diff] [blame] | 692 | }) |
| 693 | ); |
| 694 | |
| 695 | auto input2 = MakeTensor<float, 3>(inputTensorInfo2, std::vector<float>( |
| 696 | { |
surmeh01 | 3537c2c | 2018-05-18 16:31:43 +0100 | [diff] [blame] | 697 | 37.0f, 38.0f, 39.0f, |
| 698 | 40.0f, 41.0f, 42.0f, |
telsoa01 | 4fcda01 | 2018-03-09 14:13:49 +0000 | [diff] [blame] | 699 | 43.0f, 44.0f, 45.0f, |
surmeh01 | 3537c2c | 2018-05-18 16:31:43 +0100 | [diff] [blame] | 700 | 46.0f, 47.0f, 48.0f, |
| 701 | 49.0f, 50.0f, 51.0f, |
| 702 | 52.0f, 53.0f, 54.0f, |
telsoa01 | 4fcda01 | 2018-03-09 14:13:49 +0000 | [diff] [blame] | 703 | }) |
| 704 | ); |
| 705 | |
telsoa01 | c577f2c | 2018-08-31 09:22:23 +0100 | [diff] [blame] | 706 | std::vector<unsigned int> wOrigin1 = {0, 0, 0}; //Extent of the window is defined by size of input[0]. |
telsoa01 | 4fcda01 | 2018-03-09 14:13:49 +0000 | [diff] [blame] | 707 | armnn::MergerQueueDescriptor::ViewOrigin window1(wOrigin1); |
| 708 | |
telsoa01 | c577f2c | 2018-08-31 09:22:23 +0100 | [diff] [blame] | 709 | std::vector<unsigned int> wOrigin2 = {2, 0, 0}; //Extent of the window is defined by size of input[1]. |
telsoa01 | 4fcda01 | 2018-03-09 14:13:49 +0000 | [diff] [blame] | 710 | armnn::MergerQueueDescriptor::ViewOrigin window2(wOrigin2); |
| 711 | |
telsoa01 | 4fcda01 | 2018-03-09 14:13:49 +0000 | [diff] [blame] | 712 | std::unique_ptr<armnn::ITensorHandle> outputHandle = workloadFactory.CreateTensorHandle(outputTensorInfo); |
| 713 | |
| 714 | bool subTensorsSupported = workloadFactory.SupportsSubTensors(); |
| 715 | |
| 716 | std::unique_ptr<armnn::ITensorHandle> inputHandle1 = |
| 717 | subTensorsSupported ? |
| 718 | workloadFactory.CreateSubTensorHandle(*outputHandle, inputTensorInfo1.GetShape(), wOrigin1.data()) : |
| 719 | workloadFactory.CreateTensorHandle(inputTensorInfo1); |
| 720 | |
| 721 | std::unique_ptr<armnn::ITensorHandle> inputHandle2 = |
| 722 | subTensorsSupported ? |
| 723 | workloadFactory.CreateSubTensorHandle(*outputHandle, inputTensorInfo2.GetShape(), wOrigin2.data()) : |
| 724 | workloadFactory.CreateTensorHandle(inputTensorInfo2); |
| 725 | |
telsoa01 | 4fcda01 | 2018-03-09 14:13:49 +0000 | [diff] [blame] | 726 | armnn::MergerQueueDescriptor data; |
| 727 | armnn::WorkloadInfo info; |
| 728 | AddInputToWorkload(data, info, inputTensorInfo1, inputHandle1.get()); |
| 729 | AddInputToWorkload(data, info, inputTensorInfo2, inputHandle2.get()); |
telsoa01 | 4fcda01 | 2018-03-09 14:13:49 +0000 | [diff] [blame] | 730 | AddOutputToWorkload(data, info, outputTensorInfo, outputHandle.get()); |
| 731 | |
| 732 | data.m_ViewOrigins.push_back(window1); |
| 733 | data.m_ViewOrigins.push_back(window2); |
telsoa01 | 4fcda01 | 2018-03-09 14:13:49 +0000 | [diff] [blame] | 734 | |
| 735 | std::unique_ptr<armnn::IWorkload> workload = workloadFactory.CreateMerger(data, info); |
| 736 | |
| 737 | inputHandle1->Allocate(); |
| 738 | inputHandle2->Allocate(); |
telsoa01 | 4fcda01 | 2018-03-09 14:13:49 +0000 | [diff] [blame] | 739 | outputHandle->Allocate(); |
| 740 | |
| 741 | CopyDataToITensorHandle(inputHandle1.get(), &input1[0][0][0]); |
| 742 | CopyDataToITensorHandle(inputHandle2.get(), &input2[0][0][0]); |
telsoa01 | 4fcda01 | 2018-03-09 14:13:49 +0000 | [diff] [blame] | 743 | |
surmeh01 | 3537c2c | 2018-05-18 16:31:43 +0100 | [diff] [blame] | 744 | workloadFactory.Finalize(); |
telsoa01 | 4fcda01 | 2018-03-09 14:13:49 +0000 | [diff] [blame] | 745 | workload->Execute(); |
| 746 | |
| 747 | CopyDataFromITensorHandle(&ret.output[0][0][0], outputHandle.get()); |
| 748 | |
| 749 | return ret; |
| 750 | } |
| 751 | |
| 752 | LayerTestResult<float,4> AdditionTest(armnn::IWorkloadFactory& workloadFactory) |
| 753 | { |
| 754 | unsigned int batchSize = 2; |
| 755 | unsigned int channels = 2; |
| 756 | unsigned int height = 2; |
| 757 | unsigned int width = 3; |
| 758 | |
| 759 | armnn::TensorInfo inputTensorInfo1, inputTensorInfo2; |
| 760 | armnn::TensorInfo outputTensorInfo; |
| 761 | |
| 762 | unsigned int shape[] = {batchSize, channels, height, width}; |
| 763 | |
| 764 | inputTensorInfo1 = armnn::TensorInfo(4, shape, armnn::DataType::Float32); |
| 765 | inputTensorInfo2 = armnn::TensorInfo(4, shape, armnn::DataType::Float32); |
| 766 | outputTensorInfo = armnn::TensorInfo(4, shape, armnn::DataType::Float32); |
| 767 | |
| 768 | |
| 769 | auto input1 = MakeTensor<float, 4>(inputTensorInfo1, std::vector<float>( |
| 770 | { |
| 771 | 0.0f, 2.0f, 1.0f, |
| 772 | 0.2f, 1.0f, 2.0f, |
| 773 | |
| 774 | 1.0f, 2.0f, 1.0f, |
| 775 | 0.2f, 1.0f, 2.0f, |
| 776 | |
| 777 | 0.0f, 2.0f, 1.0f, |
| 778 | 4.2f, 1.0f, 2.0f, |
| 779 | |
| 780 | 0.0f, 0.0f, 1.0f, |
| 781 | 0.2f, 1.0f, 2.0f, |
| 782 | })); |
| 783 | |
| 784 | auto input2 = MakeTensor<float, 4>(inputTensorInfo2, std::vector<float>( |
| 785 | { |
| 786 | 1.0f, 2.0f, 1.0f, |
| 787 | 0.0f, 1.0f, 2.0f, |
| 788 | |
| 789 | 1.0f, 2.0f, -2.0f, |
| 790 | 0.2f, 1.0f, 2.0f, |
| 791 | |
| 792 | 0.0f, 2.0f, 1.0f, |
| 793 | 4.2f, 0.0f, -3.0f, |
| 794 | |
| 795 | 0.0f, 0.0f, 1.0f, |
| 796 | 0.7f, 1.0f, 5.0f, |
| 797 | })); |
| 798 | |
| 799 | LayerTestResult<float,4> ret(outputTensorInfo); |
| 800 | ret.outputExpected = MakeTensor<float, 4>(outputTensorInfo, std::vector<float>( |
| 801 | { |
| 802 | 1.0f, 4.0f, 2.0f, |
| 803 | 0.2f, 2.0f, 4.0f, |
| 804 | |
| 805 | 2.0f, 4.0f, -1.0f, |
| 806 | 0.4f, 2.0f, 4.0f, |
| 807 | |
| 808 | 0.0f, 4.0f, 2.0f, |
| 809 | 8.4f, 1.0f, -1.0f, |
| 810 | |
| 811 | 0.0f, 0.0f, 2.0f, |
| 812 | 0.9f, 2.0f, 7.0f, |
| 813 | })); |
| 814 | |
| 815 | std::unique_ptr<armnn::ITensorHandle> inputHandle1 = workloadFactory.CreateTensorHandle(inputTensorInfo1); |
| 816 | std::unique_ptr<armnn::ITensorHandle> inputHandle2 = workloadFactory.CreateTensorHandle(inputTensorInfo2); |
| 817 | std::unique_ptr<armnn::ITensorHandle> outputHandle = workloadFactory.CreateTensorHandle(outputTensorInfo); |
| 818 | |
| 819 | armnn::AdditionQueueDescriptor data; |
| 820 | armnn::WorkloadInfo info; |
| 821 | AddInputToWorkload(data, info, inputTensorInfo1, inputHandle1.get()); |
| 822 | AddInputToWorkload(data, info, inputTensorInfo2, inputHandle2.get()); |
| 823 | AddOutputToWorkload(data, info, outputTensorInfo, outputHandle.get()); |
| 824 | |
| 825 | std::unique_ptr<armnn::IWorkload> workload = workloadFactory.CreateAddition(data, info); |
| 826 | |
| 827 | inputHandle1->Allocate(); |
| 828 | inputHandle2->Allocate(); |
| 829 | outputHandle->Allocate(); |
| 830 | |
| 831 | CopyDataToITensorHandle(inputHandle1.get(), &input1[0][0][0][0]); |
| 832 | CopyDataToITensorHandle(inputHandle2.get(), &input2[0][0][0][0]); |
| 833 | |
surmeh01 | 3537c2c | 2018-05-18 16:31:43 +0100 | [diff] [blame] | 834 | workloadFactory.Finalize(); |
telsoa01 | 4fcda01 | 2018-03-09 14:13:49 +0000 | [diff] [blame] | 835 | workload->Execute(); |
| 836 | |
| 837 | CopyDataFromITensorHandle(&ret.output[0][0][0][0], outputHandle.get()); |
| 838 | |
| 839 | return ret; |
| 840 | } |
| 841 | |
| 842 | template <typename T> |
| 843 | LayerTestResult<T, 4> AdditionBroadcastTestImpl(armnn::IWorkloadFactory& workloadFactory, |
| 844 | float qScale, |
| 845 | int32_t qOffset) |
| 846 | { |
| 847 | armnn::TensorInfo inputTensorInfo1 = armnn::TensorInfo({1, 3, 2, 1}, armnn::GetDataType<T>()); |
| 848 | armnn::TensorInfo inputTensorInfo2 = armnn::TensorInfo({1, 1, 2, 3}, armnn::GetDataType<T>()); |
| 849 | armnn::TensorInfo outputTensorInfo = armnn::TensorInfo({1, 3, 2, 3}, armnn::GetDataType<T>()); |
| 850 | |
| 851 | if (armnn::IsQuantizedType<T>()) |
| 852 | { |
| 853 | inputTensorInfo1.SetQuantizationScale(qScale); |
| 854 | inputTensorInfo1.SetQuantizationOffset(qOffset); |
| 855 | inputTensorInfo2.SetQuantizationScale(qScale); |
| 856 | inputTensorInfo2.SetQuantizationOffset(qOffset); |
| 857 | outputTensorInfo.SetQuantizationScale(qScale); |
| 858 | outputTensorInfo.SetQuantizationOffset(qOffset); |
| 859 | } |
| 860 | |
| 861 | auto input1 = MakeTensor<T, 4>(inputTensorInfo1, QuantizedVector<T>(qScale, qOffset, |
| 862 | { |
| 863 | 0.0f, |
| 864 | 1.0f, |
| 865 | |
| 866 | 2.0f, |
| 867 | 3.0f, |
| 868 | |
| 869 | 4.0f, |
| 870 | 5.0f, |
| 871 | })); |
| 872 | |
| 873 | auto input2 = MakeTensor<T, 4>(inputTensorInfo2, QuantizedVector<T>(qScale, qOffset, |
| 874 | { |
| 875 | 0.5f, 1.5f, 2.5f, |
| 876 | 3.5f, 4.5f, 5.5f, |
| 877 | })); |
| 878 | |
| 879 | LayerTestResult<T,4> ret(outputTensorInfo); |
| 880 | ret.outputExpected = MakeTensor<T, 4>(outputTensorInfo, QuantizedVector<T>(qScale, qOffset, |
| 881 | { |
| 882 | 0.5f, 1.5f, 2.5f, |
| 883 | 4.5f, 5.5f, 6.5f, |
| 884 | |
| 885 | 2.5f, 3.5f, 4.5f, |
| 886 | 6.5f, 7.5f, 8.5f, |
| 887 | |
| 888 | 4.5f, 5.5f, 6.5f, |
| 889 | 8.5f, 9.5f, 10.5f, |
| 890 | })); |
| 891 | |
| 892 | std::unique_ptr<armnn::ITensorHandle> inputHandle1 = workloadFactory.CreateTensorHandle(inputTensorInfo1); |
| 893 | std::unique_ptr<armnn::ITensorHandle> inputHandle2 = workloadFactory.CreateTensorHandle(inputTensorInfo2); |
| 894 | std::unique_ptr<armnn::ITensorHandle> outputHandle = workloadFactory.CreateTensorHandle(outputTensorInfo); |
| 895 | |
| 896 | armnn::AdditionQueueDescriptor data; |
| 897 | armnn::WorkloadInfo info; |
| 898 | AddInputToWorkload(data, info, inputTensorInfo1, inputHandle1.get()); |
| 899 | AddInputToWorkload(data, info, inputTensorInfo2, inputHandle2.get()); |
| 900 | AddOutputToWorkload(data, info, outputTensorInfo, outputHandle.get()); |
| 901 | |
| 902 | std::unique_ptr<armnn::IWorkload> workload = workloadFactory.CreateAddition(data, info); |
| 903 | |
| 904 | inputHandle1->Allocate(); |
| 905 | inputHandle2->Allocate(); |
| 906 | outputHandle->Allocate(); |
| 907 | |
| 908 | CopyDataToITensorHandle(inputHandle1.get(), &input1[0][0][0][0]); |
| 909 | CopyDataToITensorHandle(inputHandle2.get(), &input2[0][0][0][0]); |
| 910 | |
surmeh01 | 3537c2c | 2018-05-18 16:31:43 +0100 | [diff] [blame] | 911 | workloadFactory.Finalize(); |
telsoa01 | 4fcda01 | 2018-03-09 14:13:49 +0000 | [diff] [blame] | 912 | workload->Execute(); |
| 913 | |
| 914 | CopyDataFromITensorHandle(&ret.output[0][0][0][0], outputHandle.get()); |
| 915 | |
| 916 | return ret; |
| 917 | } |
| 918 | |
| 919 | template <typename T> |
| 920 | LayerTestResult<T, 4> AdditionBroadcast1ElementTestImpl(armnn::IWorkloadFactory& workloadFactory, |
| 921 | float qScale, |
| 922 | int32_t qOffset) |
| 923 | { |
| 924 | armnn::TensorInfo inputTensorInfo1 = armnn::TensorInfo({1, 3, 2, 3}, armnn::GetDataType<T>()); |
| 925 | armnn::TensorInfo inputTensorInfo2 = armnn::TensorInfo({1, 1, 1, 1}, armnn::GetDataType<T>()); |
| 926 | armnn::TensorInfo outputTensorInfo = armnn::TensorInfo({1, 3, 2, 3}, armnn::GetDataType<T>()); |
| 927 | |
| 928 | if (armnn::IsQuantizedType<T>()) |
| 929 | { |
| 930 | inputTensorInfo1.SetQuantizationScale(qScale); |
| 931 | inputTensorInfo1.SetQuantizationOffset(qOffset); |
| 932 | inputTensorInfo2.SetQuantizationScale(qScale); |
| 933 | inputTensorInfo2.SetQuantizationOffset(qOffset); |
| 934 | outputTensorInfo.SetQuantizationScale(qScale); |
| 935 | outputTensorInfo.SetQuantizationOffset(qOffset); |
| 936 | } |
| 937 | |
| 938 | auto input1 = MakeTensor<T, 4>(inputTensorInfo1, QuantizedVector<T>(qScale, qOffset, |
| 939 | { |
| 940 | 0.0f, 1.0f, 2.0f, |
| 941 | 3.0f, 4.0f, 5.0f, |
| 942 | 6.0f, 7.0f, 8.0f, |
| 943 | 9.0f, 10.0f, 11.0f, |
| 944 | 12.0f, 13.0f, 14.0f, |
| 945 | 15.0f, 16.0f, 17.0f, |
| 946 | })); |
| 947 | |
| 948 | auto input2 = MakeTensor<T, 4>(inputTensorInfo2, QuantizedVector<T>(qScale, qOffset, |
| 949 | { |
| 950 | 0.5f, |
| 951 | })); |
| 952 | |
| 953 | LayerTestResult<T,4> ret(outputTensorInfo); |
| 954 | ret.outputExpected = MakeTensor<T, 4>(outputTensorInfo, QuantizedVector<T>(qScale, qOffset, |
| 955 | { |
| 956 | 0.5f, 1.5f, 2.5f, |
| 957 | 3.5f, 4.5f, 5.5f, |
| 958 | 6.5f, 7.5f, 8.5f, |
| 959 | 9.5f, 10.5f, 11.5f, |
| 960 | 12.5f, 13.5f, 14.5f, |
| 961 | 15.5f, 16.5f, 17.5f, |
| 962 | })); |
| 963 | |
| 964 | std::unique_ptr<armnn::ITensorHandle> inputHandle1 = workloadFactory.CreateTensorHandle(inputTensorInfo1); |
| 965 | std::unique_ptr<armnn::ITensorHandle> inputHandle2 = workloadFactory.CreateTensorHandle(inputTensorInfo2); |
| 966 | std::unique_ptr<armnn::ITensorHandle> outputHandle = workloadFactory.CreateTensorHandle(outputTensorInfo); |
| 967 | |
| 968 | armnn::AdditionQueueDescriptor data; |
| 969 | armnn::WorkloadInfo info; |
| 970 | AddInputToWorkload(data, info, inputTensorInfo1, inputHandle1.get()); |
| 971 | AddInputToWorkload(data, info, inputTensorInfo2, inputHandle2.get()); |
| 972 | AddOutputToWorkload(data, info, outputTensorInfo, outputHandle.get()); |
| 973 | |
| 974 | std::unique_ptr<armnn::IWorkload> workload = workloadFactory.CreateAddition(data, info); |
| 975 | |
| 976 | inputHandle1->Allocate(); |
| 977 | inputHandle2->Allocate(); |
| 978 | outputHandle->Allocate(); |
| 979 | |
| 980 | CopyDataToITensorHandle(inputHandle1.get(), &input1[0][0][0][0]); |
| 981 | CopyDataToITensorHandle(inputHandle2.get(), &input2[0][0][0][0]); |
| 982 | |
surmeh01 | 3537c2c | 2018-05-18 16:31:43 +0100 | [diff] [blame] | 983 | workloadFactory.Finalize(); |
telsoa01 | 4fcda01 | 2018-03-09 14:13:49 +0000 | [diff] [blame] | 984 | workload->Execute(); |
| 985 | |
| 986 | CopyDataFromITensorHandle(&ret.output[0][0][0][0], outputHandle.get()); |
| 987 | |
| 988 | return ret; |
| 989 | } |
| 990 | |
| 991 | LayerTestResult<float, 4> AdditionBroadcastTest(armnn::IWorkloadFactory& workloadFactory) |
| 992 | { |
| 993 | return AdditionBroadcastTestImpl<float>(workloadFactory, 0.0f, 0); |
| 994 | } |
| 995 | |
| 996 | LayerTestResult<uint8_t, 4> AdditionBroadcastUint8Test(armnn::IWorkloadFactory& workloadFactory) |
| 997 | { |
| 998 | return AdditionBroadcastTestImpl<uint8_t>(workloadFactory, 2.f, 0); |
| 999 | } |
| 1000 | |
| 1001 | LayerTestResult<float, 4> AdditionBroadcast1ElementTest(armnn::IWorkloadFactory& workloadFactory) |
| 1002 | { |
| 1003 | return AdditionBroadcast1ElementTestImpl<float>(workloadFactory, 0.0f, 0); |
| 1004 | } |
| 1005 | |
| 1006 | LayerTestResult<uint8_t, 4> AdditionBroadcast1ElementUint8Test(armnn::IWorkloadFactory& workloadFactory) |
| 1007 | { |
| 1008 | return AdditionBroadcast1ElementTestImpl<uint8_t>(workloadFactory, 0.1333333f, 128); |
| 1009 | } |
| 1010 | |
| 1011 | LayerTestResult<float,4> CompareAdditionTest(armnn::IWorkloadFactory& workloadFactory, |
David Beck | f195f03 | 2018-09-06 16:46:34 +0100 | [diff] [blame] | 1012 | armnn::IWorkloadFactory& refWorkloadFactory) |
telsoa01 | 4fcda01 | 2018-03-09 14:13:49 +0000 | [diff] [blame] | 1013 | { |
| 1014 | unsigned int batchSize = 4; |
| 1015 | unsigned int channels = 1; |
| 1016 | unsigned int height = 2; |
| 1017 | unsigned int width = 3; |
| 1018 | |
| 1019 | armnn::TensorInfo inputTensorInfo1, inputTensorInfo2; |
| 1020 | armnn::TensorInfo outputTensorInfo; |
| 1021 | |
| 1022 | unsigned int shape[] = {batchSize, channels, height, width}; |
| 1023 | |
| 1024 | inputTensorInfo1 = armnn::TensorInfo(4, shape, armnn::DataType::Float32); |
| 1025 | inputTensorInfo2 = armnn::TensorInfo(4, shape, armnn::DataType::Float32); |
| 1026 | outputTensorInfo = armnn::TensorInfo(4, shape, armnn::DataType::Float32); |
| 1027 | |
| 1028 | auto input1 = MakeRandomTensor<float, 4>(inputTensorInfo1, 1232); |
| 1029 | auto input2 = MakeRandomTensor<float, 4>(inputTensorInfo2, 456); |
| 1030 | |
| 1031 | LayerTestResult<float,4> ret(outputTensorInfo); |
| 1032 | |
| 1033 | std::unique_ptr<armnn::ITensorHandle> inputHandle1 = workloadFactory.CreateTensorHandle(inputTensorInfo1); |
| 1034 | std::unique_ptr<armnn::ITensorHandle> inputHandle2 = workloadFactory.CreateTensorHandle(inputTensorInfo2); |
| 1035 | std::unique_ptr<armnn::ITensorHandle> outputHandle = workloadFactory.CreateTensorHandle(outputTensorInfo); |
| 1036 | |
| 1037 | std::unique_ptr<armnn::ITensorHandle> inputHandle1Ref = refWorkloadFactory.CreateTensorHandle(inputTensorInfo1); |
| 1038 | std::unique_ptr<armnn::ITensorHandle> inputHandle2Ref = refWorkloadFactory.CreateTensorHandle(inputTensorInfo2); |
| 1039 | std::unique_ptr<armnn::ITensorHandle> outputHandleRef = refWorkloadFactory.CreateTensorHandle(outputTensorInfo); |
| 1040 | |
| 1041 | armnn::AdditionQueueDescriptor data; |
| 1042 | armnn::WorkloadInfo info; |
| 1043 | AddInputToWorkload(data, info, inputTensorInfo1, inputHandle1.get()); |
| 1044 | AddInputToWorkload(data, info, inputTensorInfo2, inputHandle2.get()); |
| 1045 | AddOutputToWorkload(data, info, outputTensorInfo, outputHandle.get()); |
| 1046 | |
| 1047 | armnn::AdditionQueueDescriptor refData = data; |
| 1048 | armnn::WorkloadInfo refInfo = info; |
| 1049 | SetWorkloadInput(refData, refInfo, 0, inputTensorInfo1, inputHandle1Ref.get()); |
| 1050 | SetWorkloadInput(refData, refInfo, 1, inputTensorInfo2, inputHandle2Ref.get()); |
| 1051 | SetWorkloadOutput(refData, refInfo, 0, outputTensorInfo, outputHandleRef.get()); |
| 1052 | |
| 1053 | std::unique_ptr<armnn::IWorkload> workload = workloadFactory.CreateAddition(data, info); |
| 1054 | std::unique_ptr<armnn::IWorkload> workloadRef = refWorkloadFactory.CreateAddition(refData, refInfo); |
| 1055 | |
| 1056 | inputHandle1->Allocate(); |
| 1057 | inputHandle2->Allocate(); |
| 1058 | outputHandle->Allocate(); |
| 1059 | inputHandle1Ref->Allocate(); |
| 1060 | inputHandle2Ref->Allocate(); |
| 1061 | outputHandleRef->Allocate(); |
| 1062 | |
| 1063 | CopyDataToITensorHandle(inputHandle1.get(), &input1[0][0][0][0]); |
| 1064 | CopyDataToITensorHandle(inputHandle2.get(), &input2[0][0][0][0]); |
| 1065 | CopyDataToITensorHandle(inputHandle1Ref.get(), &input1[0][0][0][0]); |
| 1066 | CopyDataToITensorHandle(inputHandle2Ref.get(), &input2[0][0][0][0]); |
| 1067 | |
surmeh01 | 3537c2c | 2018-05-18 16:31:43 +0100 | [diff] [blame] | 1068 | workloadFactory.Finalize(); |
telsoa01 | 4fcda01 | 2018-03-09 14:13:49 +0000 | [diff] [blame] | 1069 | workload->Execute(); |
surmeh01 | 3537c2c | 2018-05-18 16:31:43 +0100 | [diff] [blame] | 1070 | refWorkloadFactory.Finalize(); |
telsoa01 | 4fcda01 | 2018-03-09 14:13:49 +0000 | [diff] [blame] | 1071 | workloadRef->Execute(); |
| 1072 | |
| 1073 | CopyDataFromITensorHandle(&ret.output[0][0][0][0], outputHandle.get()); |
| 1074 | CopyDataFromITensorHandle(&ret.outputExpected[0][0][0][0], outputHandleRef.get()); |
| 1075 | |
| 1076 | return ret; |
| 1077 | } |
| 1078 | |
surmeh01 | bceff2f | 2018-03-29 16:29:27 +0100 | [diff] [blame] | 1079 | namespace { |
David Beck | 5cd01f3 | 2018-09-12 16:00:08 +0100 | [diff] [blame] | 1080 | template <typename T> |
| 1081 | LayerTestResult<T, 4> DivisionTestHelper(armnn::IWorkloadFactory& workloadFactory, |
| 1082 | const unsigned int shape0[4], |
| 1083 | const std::vector<T>& values0, |
| 1084 | float scale0, |
| 1085 | int32_t offset0, |
| 1086 | const unsigned int shape1[4], |
| 1087 | const std::vector<T> & values1, |
| 1088 | float scale1, |
| 1089 | int32_t offset1, |
| 1090 | const unsigned int outShape[4], |
| 1091 | const std::vector<T> & outValues, |
| 1092 | float outScale, |
| 1093 | int32_t outOffset) |
| 1094 | { |
| 1095 | auto dataType = (std::is_same<T, uint8_t>::value ? |
| 1096 | armnn::DataType::QuantisedAsymm8 : |
| 1097 | armnn::DataType::Float32); |
Francis Murtagh | e7a86a4 | 2018-08-29 12:42:10 +0100 | [diff] [blame] | 1098 | |
David Beck | 5cd01f3 | 2018-09-12 16:00:08 +0100 | [diff] [blame] | 1099 | armnn::TensorInfo inputTensorInfo0(4, shape0, dataType); |
| 1100 | armnn::TensorInfo inputTensorInfo1(4, shape1, dataType); |
| 1101 | armnn::TensorInfo outputTensorInfo(4, outShape, dataType); |
Francis Murtagh | e7a86a4 | 2018-08-29 12:42:10 +0100 | [diff] [blame] | 1102 | |
David Beck | 5cd01f3 | 2018-09-12 16:00:08 +0100 | [diff] [blame] | 1103 | inputTensorInfo0.SetQuantizationScale(scale0); |
| 1104 | inputTensorInfo0.SetQuantizationOffset(offset0); |
Francis Murtagh | e7a86a4 | 2018-08-29 12:42:10 +0100 | [diff] [blame] | 1105 | |
David Beck | 5cd01f3 | 2018-09-12 16:00:08 +0100 | [diff] [blame] | 1106 | inputTensorInfo1.SetQuantizationScale(scale1); |
| 1107 | inputTensorInfo1.SetQuantizationOffset(offset1); |
Francis Murtagh | e7a86a4 | 2018-08-29 12:42:10 +0100 | [diff] [blame] | 1108 | |
David Beck | 5cd01f3 | 2018-09-12 16:00:08 +0100 | [diff] [blame] | 1109 | outputTensorInfo.SetQuantizationScale(outScale); |
| 1110 | outputTensorInfo.SetQuantizationOffset(outOffset); |
Francis Murtagh | e7a86a4 | 2018-08-29 12:42:10 +0100 | [diff] [blame] | 1111 | |
David Beck | 5cd01f3 | 2018-09-12 16:00:08 +0100 | [diff] [blame] | 1112 | auto input0 = MakeTensor<T, 4>(inputTensorInfo0, values0); |
| 1113 | auto input1 = MakeTensor<T, 4>(inputTensorInfo1, values1); |
Francis Murtagh | e7a86a4 | 2018-08-29 12:42:10 +0100 | [diff] [blame] | 1114 | |
David Beck | 5cd01f3 | 2018-09-12 16:00:08 +0100 | [diff] [blame] | 1115 | LayerTestResult<T, 4> result(outputTensorInfo); |
| 1116 | result.outputExpected = MakeTensor<T, 4>(outputTensorInfo, outValues); |
Francis Murtagh | e7a86a4 | 2018-08-29 12:42:10 +0100 | [diff] [blame] | 1117 | |
David Beck | 5cd01f3 | 2018-09-12 16:00:08 +0100 | [diff] [blame] | 1118 | std::unique_ptr<armnn::ITensorHandle> inputHandle0 = workloadFactory.CreateTensorHandle(inputTensorInfo0); |
| 1119 | std::unique_ptr<armnn::ITensorHandle> inputHandle1 = workloadFactory.CreateTensorHandle(inputTensorInfo1); |
| 1120 | std::unique_ptr<armnn::ITensorHandle> outputHandle = workloadFactory.CreateTensorHandle(outputTensorInfo); |
Francis Murtagh | e7a86a4 | 2018-08-29 12:42:10 +0100 | [diff] [blame] | 1121 | |
David Beck | 5cd01f3 | 2018-09-12 16:00:08 +0100 | [diff] [blame] | 1122 | armnn::DivisionQueueDescriptor data; |
| 1123 | armnn::WorkloadInfo info; |
| 1124 | AddInputToWorkload(data, info, inputTensorInfo0, inputHandle0.get()); |
| 1125 | AddInputToWorkload(data, info, inputTensorInfo1, inputHandle1.get()); |
| 1126 | AddOutputToWorkload(data, info, outputTensorInfo, outputHandle.get()); |
Francis Murtagh | e7a86a4 | 2018-08-29 12:42:10 +0100 | [diff] [blame] | 1127 | |
David Beck | 5cd01f3 | 2018-09-12 16:00:08 +0100 | [diff] [blame] | 1128 | std::unique_ptr<armnn::IWorkload> workload = workloadFactory.CreateDivision(data, info); |
Francis Murtagh | e7a86a4 | 2018-08-29 12:42:10 +0100 | [diff] [blame] | 1129 | |
David Beck | 5cd01f3 | 2018-09-12 16:00:08 +0100 | [diff] [blame] | 1130 | inputHandle0->Allocate(); |
| 1131 | inputHandle1->Allocate(); |
| 1132 | outputHandle->Allocate(); |
| 1133 | |
| 1134 | CopyDataToITensorHandle(inputHandle0.get(), &input0[0][0][0][0]); |
| 1135 | CopyDataToITensorHandle(inputHandle1.get(), &input1[0][0][0][0]); |
| 1136 | |
| 1137 | workloadFactory.Finalize(); |
| 1138 | workload->Execute(); |
| 1139 | |
| 1140 | CopyDataFromITensorHandle(&result.output[0][0][0][0], outputHandle.get()); |
| 1141 | |
| 1142 | return result; |
| 1143 | } |
Francis Murtagh | e7a86a4 | 2018-08-29 12:42:10 +0100 | [diff] [blame] | 1144 | } // anonymous namespace |
| 1145 | |
Francis Murtagh | 8c5e3dc | 2018-08-30 17:18:37 +0100 | [diff] [blame] | 1146 | LayerTestResult<float,4> DivisionByZeroTest(armnn::IWorkloadFactory& workloadFactory) |
| 1147 | { |
| 1148 | const unsigned int width = 2; |
| 1149 | const unsigned int height = 2; |
| 1150 | const unsigned int channelCount = 2; |
| 1151 | const unsigned int batchSize = 2; |
| 1152 | |
| 1153 | unsigned int shape[] = { batchSize, channelCount, height, width }; |
| 1154 | |
| 1155 | std::vector<float> input0({ |
| 1156 | 1.f, 1.f, 1.f, 1.f, 0.f, 0.f, 0.f, 0.f, |
| 1157 | -1.f, -1.f, -1.f, -1.f, 5.f, 5.f, 5.f, 5.f }); |
| 1158 | |
| 1159 | std::vector<float> input1({ |
| 1160 | 0.f, 0.f, -0.f, -0.f, 0.f, 0.f, -0.f, -0.f, |
| 1161 | 0.f, 0.f, -0.f, -0.f, 5.f, 5.f, 5.f, 5.f }); |
| 1162 | |
| 1163 | std::vector<float> output({ |
| 1164 | INFINITY, INFINITY, -INFINITY, -INFINITY, NAN, NAN, -NAN, -NAN, |
| 1165 | -INFINITY, -INFINITY, INFINITY, INFINITY, 1, 1, 1, 1 }); |
| 1166 | |
David Beck | 5cd01f3 | 2018-09-12 16:00:08 +0100 | [diff] [blame] | 1167 | return DivisionTestHelper<float>(workloadFactory, |
| 1168 | shape, input0, 1.0f, 0, |
| 1169 | shape, input1, 1.0f, 0, |
| 1170 | shape, output, 1.0f, 0); |
Francis Murtagh | 8c5e3dc | 2018-08-30 17:18:37 +0100 | [diff] [blame] | 1171 | } |
| 1172 | |
Francis Murtagh | e7a86a4 | 2018-08-29 12:42:10 +0100 | [diff] [blame] | 1173 | LayerTestResult<float,4> DivisionTest(armnn::IWorkloadFactory& workloadFactory) |
| 1174 | { |
| 1175 | const unsigned int width = 2; |
| 1176 | const unsigned int height = 2; |
| 1177 | const unsigned int channelCount = 2; |
| 1178 | const unsigned int batchSize = 2; |
| 1179 | |
| 1180 | unsigned int shape[] = { batchSize, channelCount, height, width }; |
| 1181 | |
| 1182 | std::vector<float> input0({ |
| 1183 | 2, 2, 2, 2, 3, 3, 3, 3, |
| 1184 | 4, 4, 4, 4, 5, 5, 5, 5 }); |
| 1185 | |
| 1186 | std::vector<float> input1({ |
| 1187 | 1, 1, 1, 1, 2, 2, 2, 2, |
| 1188 | 4, 4, 4, 4, 4, 4, 4, 4 }); |
| 1189 | |
| 1190 | std::vector<float> output({ |
| 1191 | 2, 2, 2, 2, 1.5, 1.5, 1.5, 1.5, |
| 1192 | 1, 1, 1, 1, 1.25, 1.25, 1.25, 1.25 }); |
| 1193 | |
David Beck | 5cd01f3 | 2018-09-12 16:00:08 +0100 | [diff] [blame] | 1194 | |
| 1195 | return DivisionTestHelper<float>(workloadFactory, |
| 1196 | shape, input0, 1.0f, 0, |
| 1197 | shape, input1, 1.0f, 0, |
| 1198 | shape, output, 1.0f, 0); |
Francis Murtagh | e7a86a4 | 2018-08-29 12:42:10 +0100 | [diff] [blame] | 1199 | } |
| 1200 | |
| 1201 | LayerTestResult<float, 4> DivisionBroadcast1ElementTest(armnn::IWorkloadFactory& workloadFactory) |
| 1202 | { |
| 1203 | unsigned int shape0[] = { 1, 2, 2, 2 }; |
| 1204 | std::vector<float> input0({ 2, 4, 6, 8, 10, 12, 14, 16}); |
| 1205 | |
| 1206 | unsigned int shape1[] = { 1, 1, 1, 1 }; |
| 1207 | std::vector<float> input1({ 2 }); |
| 1208 | |
| 1209 | std::vector<float> output({ 1, 2, 3, 4, 5, 6, 7, 8}); |
| 1210 | |
David Beck | 5cd01f3 | 2018-09-12 16:00:08 +0100 | [diff] [blame] | 1211 | |
| 1212 | return DivisionTestHelper<float>(workloadFactory, |
| 1213 | shape0, input0, 1.0f, 0, |
| 1214 | shape1, input1, 1.0f, 0, |
| 1215 | shape0, output, 1.0f, 0); |
Francis Murtagh | e7a86a4 | 2018-08-29 12:42:10 +0100 | [diff] [blame] | 1216 | } |
| 1217 | |
| 1218 | LayerTestResult<float, 4> DivisionBroadcast1DVectorTest(armnn::IWorkloadFactory& workloadFactory) |
| 1219 | { |
| 1220 | unsigned int shape0[] = { 1, 3, 3, 2 }; |
| 1221 | std::vector<float> input0({ |
| 1222 | 1, 4, 3, 8, 5, 12, |
| 1223 | 7, 16, 9, 20, 11, 24, |
| 1224 | 13, 28, 15, 32, 17, 36}); |
| 1225 | |
| 1226 | unsigned int shape1[] = { 1, 1, 1, 2 }; |
| 1227 | std::vector<float> input1({ 1, 2 }); |
| 1228 | |
| 1229 | std::vector<float> output({ |
| 1230 | 1, 2, 3, 4, 5, 6, |
| 1231 | 7, 8, 9, 10, 11, 12, |
| 1232 | 13, 14, 15, 16, 17, 18}); |
| 1233 | |
David Beck | 5cd01f3 | 2018-09-12 16:00:08 +0100 | [diff] [blame] | 1234 | return DivisionTestHelper<float>(workloadFactory, |
| 1235 | shape0, input0, 1.0f, 0, |
| 1236 | shape1, input1, 1.0f, 0, |
| 1237 | shape0, output, 1.0f, 0); |
| 1238 | } |
| 1239 | |
| 1240 | |
| 1241 | LayerTestResult<uint8_t,4> DivisionUint8Test(armnn::IWorkloadFactory& workloadFactory) |
| 1242 | { |
| 1243 | const unsigned int width = 2; |
| 1244 | const unsigned int height = 2; |
| 1245 | const unsigned int channelCount = 2; |
| 1246 | const unsigned int batchSize = 2; |
| 1247 | |
| 1248 | unsigned int shape[] = { batchSize, channelCount, height, width }; |
| 1249 | |
| 1250 | std::vector<uint8_t> input0({2, 2, 2, 2, 3, 3, 3, 3, |
| 1251 | 4, 4, 4, 4, 5, 5, 5, 5 }); |
| 1252 | |
| 1253 | std::vector<uint8_t> input1({1, 1, 1, 1, 2, 2, 2, 2, |
| 1254 | 4, 4, 4, 4, 4, 4, 4, 4 }); |
| 1255 | |
| 1256 | std::vector<uint8_t> output({8, 8, 8, 8, 6, 6, 6, 6, |
| 1257 | 4, 4, 4, 4, 5, 5, 5, 5}); |
| 1258 | |
| 1259 | |
| 1260 | return DivisionTestHelper<uint8_t>(workloadFactory, |
| 1261 | shape, input0, 1.0f, 0, |
| 1262 | shape, input1, 1.0f, 0, |
| 1263 | shape, output, 0.25f, 0); |
| 1264 | } |
| 1265 | |
| 1266 | LayerTestResult<uint8_t, 4> DivisionBroadcast1ElementUint8Test(armnn::IWorkloadFactory& workloadFactory) |
| 1267 | { |
| 1268 | unsigned int shape0[] = { 1, 2, 2, 2 }; |
| 1269 | std::vector<uint8_t> input0({ 2, 4, 6, 8, 10, 12, 14, 16}); |
| 1270 | |
| 1271 | unsigned int shape1[] = { 1, 1, 1, 1 }; |
| 1272 | std::vector<uint8_t> input1({ 2 }); |
| 1273 | |
| 1274 | std::vector<uint8_t> output({ 1, 2, 3, 4, 5, 6, 7, 8}); |
| 1275 | |
| 1276 | return DivisionTestHelper<uint8_t>(workloadFactory, |
| 1277 | shape0, input0, 1.0f, 0, |
| 1278 | shape1, input1, 1.0f, 0, |
| 1279 | shape0, output, 1.0f, 0); |
| 1280 | } |
| 1281 | |
| 1282 | LayerTestResult<uint8_t, 4> DivisionBroadcast1DVectorUint8Test(armnn::IWorkloadFactory& workloadFactory) |
| 1283 | { |
| 1284 | unsigned int shape0[] = { 1, 3, 3, 2 }; |
| 1285 | std::vector<uint8_t> input0({1, 4, 3, 8, 5, 12, |
| 1286 | 7, 16, 9, 20, 11, 24, |
| 1287 | 13, 28, 15, 32, 17, 36}); |
| 1288 | |
| 1289 | unsigned int shape1[] = { 1, 1, 1, 2 }; |
| 1290 | std::vector<uint8_t> input1({ 1, 2 }); |
| 1291 | |
| 1292 | std::vector<uint8_t> output({1, 2, 3, 4, 5, 6, |
| 1293 | 7, 8, 9, 10, 11, 12, |
| 1294 | 13, 14, 15, 16, 17, 18}); |
| 1295 | |
| 1296 | return DivisionTestHelper<uint8_t>(workloadFactory, |
| 1297 | shape0, input0, 1.0f, 0, |
| 1298 | shape1, input1, 1.0f, 0, |
| 1299 | shape0, output, 1.0f, 0); |
Francis Murtagh | e7a86a4 | 2018-08-29 12:42:10 +0100 | [diff] [blame] | 1300 | } |
| 1301 | |
| 1302 | namespace { |
surmeh01 | bceff2f | 2018-03-29 16:29:27 +0100 | [diff] [blame] | 1303 | LayerTestResult<float,4> MultiplicationTestHelper(armnn::IWorkloadFactory& workloadFactory, |
| 1304 | const unsigned int shape0[4], |
| 1305 | const std::vector<float> & values0, |
| 1306 | const unsigned int shape1[4], |
| 1307 | const std::vector<float> & values1, |
| 1308 | const unsigned int outShape[4], |
| 1309 | const std::vector<float> & outValues) |
telsoa01 | 4fcda01 | 2018-03-09 14:13:49 +0000 | [diff] [blame] | 1310 | { |
surmeh01 | bceff2f | 2018-03-29 16:29:27 +0100 | [diff] [blame] | 1311 | const size_t dimensionCount = 4; |
| 1312 | armnn::TensorInfo inputTensorInfo0{dimensionCount, shape0, armnn::DataType::Float32}; |
| 1313 | armnn::TensorInfo inputTensorInfo1{dimensionCount, shape1, armnn::DataType::Float32}; |
| 1314 | armnn::TensorInfo outputTensorInfo{dimensionCount, outShape, armnn::DataType::Float32}; |
telsoa01 | 4fcda01 | 2018-03-09 14:13:49 +0000 | [diff] [blame] | 1315 | |
surmeh01 | bceff2f | 2018-03-29 16:29:27 +0100 | [diff] [blame] | 1316 | auto input0 = MakeTensor<float, 4>(inputTensorInfo0, values0); |
| 1317 | auto input1 = MakeTensor<float, 4>(inputTensorInfo1, values1); |
telsoa01 | 4fcda01 | 2018-03-09 14:13:49 +0000 | [diff] [blame] | 1318 | |
| 1319 | LayerTestResult<float,4> ret(outputTensorInfo); |
| 1320 | |
| 1321 | std::unique_ptr<armnn::ITensorHandle> inputHandle0 = workloadFactory.CreateTensorHandle(inputTensorInfo0); |
| 1322 | std::unique_ptr<armnn::ITensorHandle> inputHandle1 = workloadFactory.CreateTensorHandle(inputTensorInfo1); |
| 1323 | std::unique_ptr<armnn::ITensorHandle> outputHandle = workloadFactory.CreateTensorHandle(outputTensorInfo); |
| 1324 | |
| 1325 | armnn::MultiplicationQueueDescriptor data; |
| 1326 | armnn::WorkloadInfo info; |
| 1327 | AddInputToWorkload(data, info, inputTensorInfo0, inputHandle0.get()); |
| 1328 | AddInputToWorkload(data, info, inputTensorInfo1, inputHandle1.get()); |
| 1329 | AddOutputToWorkload(data, info, outputTensorInfo, outputHandle.get()); |
| 1330 | |
| 1331 | std::unique_ptr<armnn::IWorkload> workload = workloadFactory.CreateMultiplication(data, info); |
| 1332 | |
| 1333 | inputHandle0->Allocate(); |
| 1334 | inputHandle1->Allocate(); |
| 1335 | outputHandle->Allocate(); |
| 1336 | |
| 1337 | CopyDataToITensorHandle(inputHandle0.get(), &input0[0][0][0][0]); |
| 1338 | CopyDataToITensorHandle(inputHandle1.get(), &input1[0][0][0][0]); |
| 1339 | |
surmeh01 | 3537c2c | 2018-05-18 16:31:43 +0100 | [diff] [blame] | 1340 | workloadFactory.Finalize(); |
telsoa01 | 4fcda01 | 2018-03-09 14:13:49 +0000 | [diff] [blame] | 1341 | workload->Execute(); |
| 1342 | |
| 1343 | CopyDataFromITensorHandle(&ret.output[0][0][0][0], outputHandle.get()); |
| 1344 | |
surmeh01 | bceff2f | 2018-03-29 16:29:27 +0100 | [diff] [blame] | 1345 | ret.outputExpected = MakeTensor<float, 4>(outputTensorInfo, outValues); |
telsoa01 | 4fcda01 | 2018-03-09 14:13:49 +0000 | [diff] [blame] | 1346 | return ret; |
| 1347 | } |
surmeh01 | bceff2f | 2018-03-29 16:29:27 +0100 | [diff] [blame] | 1348 | } // anonymous namespace |
| 1349 | |
| 1350 | |
| 1351 | LayerTestResult<float,4> MultiplicationTest(armnn::IWorkloadFactory& workloadFactory) |
| 1352 | { |
| 1353 | const unsigned int width = 2; |
| 1354 | const unsigned int height = 2; |
| 1355 | const unsigned int channelCount = 2; |
| 1356 | const unsigned int batchSize = 2; |
| 1357 | |
| 1358 | unsigned int shape[] = { batchSize, channelCount, height, width }; |
| 1359 | |
| 1360 | std::vector<float> input0({ |
| 1361 | 1, 1, 1, 1, 2, 2, 2, 2, |
| 1362 | 3, 3, 3, 3, 4, 4, 4, 4 }); |
| 1363 | |
| 1364 | std::vector<float> input1({ |
| 1365 | 2, 2, 2, 2, 3, 3, 3, 3, |
| 1366 | 4, 4, 4, 4, 5, 5, 5, 5 }); |
| 1367 | |
| 1368 | std::vector<float> output({ |
| 1369 | 2, 2, 2, 2, 6, 6, 6, 6, |
| 1370 | 12, 12, 12, 12, 20, 20, 20, 20 }); |
| 1371 | |
| 1372 | return MultiplicationTestHelper(workloadFactory, |
| 1373 | shape, |
| 1374 | input0, |
| 1375 | shape, |
| 1376 | input1, |
| 1377 | shape, |
| 1378 | output); |
| 1379 | } |
| 1380 | |
| 1381 | LayerTestResult<float, 4> MultiplicationBroadcast1ElementTest(armnn::IWorkloadFactory& workloadFactory) |
| 1382 | { |
| 1383 | unsigned int shape0[] = { 1, 2, 2, 2 }; |
| 1384 | std::vector<float> input0({ 1, 2, 3, 4, 5, 6, 7, 8}); |
| 1385 | |
| 1386 | unsigned int shape1[] = { 1, 1, 1, 1 }; |
| 1387 | std::vector<float> input1({ 2 }); |
| 1388 | |
| 1389 | std::vector<float> output({ 2, 4, 6, 8, 10, 12, 14, 16}); |
| 1390 | |
| 1391 | return MultiplicationTestHelper(workloadFactory, |
| 1392 | shape0, |
| 1393 | input0, |
| 1394 | shape1, |
| 1395 | input1, |
| 1396 | shape0, |
| 1397 | output); |
| 1398 | } |
| 1399 | |
| 1400 | LayerTestResult<float, 4> MultiplicationBroadcast1DVectorTest(armnn::IWorkloadFactory& workloadFactory) |
| 1401 | { |
| 1402 | unsigned int shape0[] = { 1, 3, 3, 2 }; |
| 1403 | std::vector<float> input0({ |
| 1404 | 1, 2, 3, 4, 5, 6, |
| 1405 | 7, 8, 9, 10, 11, 12, |
| 1406 | 13, 14, 15, 16, 17, 18}); |
| 1407 | |
| 1408 | unsigned int shape1[] = { 1, 1, 1, 2 }; |
| 1409 | std::vector<float> input1({ 1, 2 }); |
| 1410 | |
| 1411 | std::vector<float> output({ |
| 1412 | 1, 4, 3, 8, 5, 12, |
| 1413 | 7, 16, 9, 20, 11, 24, |
| 1414 | 13, 28, 15, 32, 17, 36}); |
| 1415 | |
| 1416 | return MultiplicationTestHelper(workloadFactory, |
| 1417 | shape0, |
| 1418 | input0, |
| 1419 | shape1, |
| 1420 | input1, |
| 1421 | shape0, |
| 1422 | output); |
| 1423 | } |
telsoa01 | 4fcda01 | 2018-03-09 14:13:49 +0000 | [diff] [blame] | 1424 | |
| 1425 | LayerTestResult<float,4> CompareMultiplicationTest(armnn::IWorkloadFactory& workloadFactory, |
| 1426 | armnn::IWorkloadFactory& refWorkloadFactory) |
| 1427 | { |
| 1428 | const unsigned int width = 16; |
| 1429 | const unsigned int height = 32; |
| 1430 | const unsigned int channelCount = 2; |
| 1431 | const unsigned int batchSize = 5; |
| 1432 | |
| 1433 | armnn::TensorInfo inputTensorInfo0; |
| 1434 | armnn::TensorInfo inputTensorInfo1; |
| 1435 | armnn::TensorInfo outputTensorInfo; |
| 1436 | |
| 1437 | constexpr unsigned int shape[] = { batchSize, channelCount, height, width }; |
| 1438 | |
| 1439 | inputTensorInfo0 = armnn::TensorInfo(4, shape, armnn::DataType::Float32); |
| 1440 | inputTensorInfo1 = armnn::TensorInfo(4, shape, armnn::DataType::Float32); |
| 1441 | outputTensorInfo = armnn::TensorInfo(4, shape, armnn::DataType::Float32); |
| 1442 | |
| 1443 | LayerTestResult<float,4> comparisonResult(outputTensorInfo); |
| 1444 | |
| 1445 | auto input0 = MakeRandomTensor<float, 4>(inputTensorInfo0, 803506992); |
| 1446 | auto input1 = MakeRandomTensor<float, 4>(inputTensorInfo1, 54902257); |
| 1447 | |
| 1448 | std::unique_ptr<armnn::ITensorHandle> inputHandle0 = workloadFactory.CreateTensorHandle(inputTensorInfo0); |
| 1449 | std::unique_ptr<armnn::ITensorHandle> inputHandle1 = workloadFactory.CreateTensorHandle(inputTensorInfo1); |
| 1450 | std::unique_ptr<armnn::ITensorHandle> outputHandle = workloadFactory.CreateTensorHandle(outputTensorInfo); |
| 1451 | |
| 1452 | std::unique_ptr<armnn::ITensorHandle> inputHandle0Ref = refWorkloadFactory.CreateTensorHandle(inputTensorInfo0); |
| 1453 | std::unique_ptr<armnn::ITensorHandle> inputHandle1Ref = refWorkloadFactory.CreateTensorHandle(inputTensorInfo1); |
| 1454 | std::unique_ptr<armnn::ITensorHandle> outputHandleRef = refWorkloadFactory.CreateTensorHandle(outputTensorInfo); |
| 1455 | |
| 1456 | armnn::MultiplicationQueueDescriptor data; |
| 1457 | armnn::WorkloadInfo info; |
| 1458 | AddInputToWorkload(data, info, inputTensorInfo0, inputHandle0.get()); |
| 1459 | AddInputToWorkload(data, info, inputTensorInfo1, inputHandle1.get()); |
| 1460 | AddOutputToWorkload(data, info, outputTensorInfo, outputHandle.get()); |
| 1461 | |
| 1462 | armnn::MultiplicationQueueDescriptor refData = data; |
| 1463 | armnn::WorkloadInfo refInfo = info; |
| 1464 | SetWorkloadInput(refData, refInfo, 0, inputTensorInfo0, inputHandle0Ref.get()); |
| 1465 | SetWorkloadInput(refData, refInfo, 1, inputTensorInfo1, inputHandle1Ref.get()); |
| 1466 | SetWorkloadOutput(refData, refInfo, 0, outputTensorInfo, outputHandleRef.get()); |
| 1467 | |
| 1468 | std::unique_ptr<armnn::IWorkload> workload = workloadFactory.CreateMultiplication(data, info); |
| 1469 | std::unique_ptr<armnn::IWorkload> workloadRef = refWorkloadFactory.CreateMultiplication(refData, refInfo); |
| 1470 | |
| 1471 | inputHandle0->Allocate(); |
| 1472 | inputHandle1->Allocate(); |
| 1473 | outputHandle->Allocate(); |
| 1474 | inputHandle0Ref->Allocate(); |
| 1475 | inputHandle1Ref->Allocate(); |
| 1476 | outputHandleRef->Allocate(); |
| 1477 | |
| 1478 | CopyDataToITensorHandle(inputHandle0.get(), &input0[0][0][0][0]); |
| 1479 | CopyDataToITensorHandle(inputHandle1.get(), &input1[0][0][0][0]); |
| 1480 | CopyDataToITensorHandle(inputHandle0Ref.get(), &input0[0][0][0][0]); |
| 1481 | CopyDataToITensorHandle(inputHandle1Ref.get(), &input1[0][0][0][0]); |
| 1482 | |
surmeh01 | 3537c2c | 2018-05-18 16:31:43 +0100 | [diff] [blame] | 1483 | workloadFactory.Finalize(); |
telsoa01 | 4fcda01 | 2018-03-09 14:13:49 +0000 | [diff] [blame] | 1484 | workload->Execute(); |
surmeh01 | 3537c2c | 2018-05-18 16:31:43 +0100 | [diff] [blame] | 1485 | refWorkloadFactory.Finalize(); |
telsoa01 | 4fcda01 | 2018-03-09 14:13:49 +0000 | [diff] [blame] | 1486 | workloadRef->Execute(); |
| 1487 | |
| 1488 | CopyDataFromITensorHandle(&comparisonResult.output[0][0][0][0], outputHandle.get()); |
| 1489 | CopyDataFromITensorHandle(&comparisonResult.outputExpected[0][0][0][0], outputHandleRef.get()); |
| 1490 | |
| 1491 | return comparisonResult; |
| 1492 | } |
| 1493 | |
| 1494 | LayerTestResult<float,4> CompareBatchNormTest(armnn::IWorkloadFactory& workloadFactory, |
| 1495 | armnn::IWorkloadFactory& refWorkloadFactory) |
| 1496 | { |
| 1497 | const unsigned int width = 2; |
| 1498 | const unsigned int height = 3; |
| 1499 | const unsigned int channels = 5; |
| 1500 | const unsigned int batchSize = 3; |
| 1501 | |
| 1502 | armnn::TensorInfo inputTensorInfo; |
| 1503 | armnn::TensorInfo outputTensorInfo; |
| 1504 | armnn::TensorInfo tensorInfo; |
| 1505 | |
| 1506 | constexpr unsigned int shape[] = {batchSize, channels, height, width}; |
| 1507 | constexpr unsigned int tensorShape[] = {channels}; |
| 1508 | |
| 1509 | inputTensorInfo = armnn::TensorInfo(4, shape, armnn::DataType::Float32); |
| 1510 | outputTensorInfo = armnn::TensorInfo(4, shape, armnn::DataType::Float32); |
| 1511 | tensorInfo = armnn::TensorInfo(1, tensorShape, armnn::DataType::Float32); |
| 1512 | |
| 1513 | auto input = MakeRandomTensor<float, 4>(inputTensorInfo, 21312); |
| 1514 | |
| 1515 | auto mean = MakeRandomTensor<float, 1>(tensorInfo, 123); |
| 1516 | auto variance = MakeRandomTensor<float, 1>(tensorInfo, 234, 0.0f); |
| 1517 | auto beta = MakeRandomTensor<float, 1>(tensorInfo, 123); |
| 1518 | auto gamma = MakeRandomTensor<float, 1>(tensorInfo, 345); |
| 1519 | |
| 1520 | LayerTestResult<float,4> ret(outputTensorInfo); |
| 1521 | |
| 1522 | std::unique_ptr<armnn::ITensorHandle> inputHandle = workloadFactory.CreateTensorHandle(inputTensorInfo); |
| 1523 | std::unique_ptr<armnn::ITensorHandle> outputHandle = workloadFactory.CreateTensorHandle(outputTensorInfo); |
| 1524 | |
| 1525 | std::unique_ptr<armnn::ITensorHandle> inputHandleRef = refWorkloadFactory.CreateTensorHandle(inputTensorInfo); |
| 1526 | std::unique_ptr<armnn::ITensorHandle> outputHandleRef = refWorkloadFactory.CreateTensorHandle(outputTensorInfo); |
| 1527 | |
| 1528 | armnn::BatchNormalizationQueueDescriptor data; |
| 1529 | armnn::WorkloadInfo info; |
| 1530 | armnn::ScopedCpuTensorHandle meanTensor(tensorInfo); |
| 1531 | armnn::ScopedCpuTensorHandle varianceTensor(tensorInfo); |
| 1532 | armnn::ScopedCpuTensorHandle betaTensor(tensorInfo); |
| 1533 | armnn::ScopedCpuTensorHandle gammaTensor(tensorInfo); |
| 1534 | |
| 1535 | AllocateAndCopyDataToITensorHandle(&meanTensor, &mean[0]); |
| 1536 | AllocateAndCopyDataToITensorHandle(&varianceTensor, &variance[0]); |
| 1537 | AllocateAndCopyDataToITensorHandle(&betaTensor, &beta[0]); |
| 1538 | AllocateAndCopyDataToITensorHandle(&gammaTensor, &gamma[0]); |
| 1539 | |
| 1540 | AddInputToWorkload(data, info, inputTensorInfo, inputHandle.get()); |
| 1541 | AddOutputToWorkload(data, info, outputTensorInfo, outputHandle.get()); |
| 1542 | data.m_Mean = &meanTensor; |
| 1543 | data.m_Variance = &varianceTensor; |
| 1544 | data.m_Beta = &betaTensor; |
| 1545 | data.m_Gamma = &gammaTensor; |
| 1546 | data.m_Parameters.m_Eps = 0.01f; |
| 1547 | |
| 1548 | armnn::BatchNormalizationQueueDescriptor refData = data; |
| 1549 | armnn::WorkloadInfo refInfo = info; |
| 1550 | SetWorkloadInput(refData, refInfo, 0, inputTensorInfo, inputHandleRef.get()); |
| 1551 | SetWorkloadOutput(refData, refInfo, 0, outputTensorInfo, outputHandleRef.get()); |
| 1552 | |
| 1553 | std::unique_ptr<armnn::IWorkload> workload = workloadFactory.CreateBatchNormalization(data, info); |
| 1554 | std::unique_ptr<armnn::IWorkload> workloadRef = refWorkloadFactory.CreateBatchNormalization(refData, refInfo); |
| 1555 | |
| 1556 | inputHandle->Allocate(); |
| 1557 | outputHandle->Allocate(); |
| 1558 | inputHandleRef->Allocate(); |
| 1559 | outputHandleRef->Allocate(); |
| 1560 | |
| 1561 | CopyDataToITensorHandle(inputHandle.get(), &input[0][0][0][0]); |
| 1562 | CopyDataToITensorHandle(inputHandleRef.get(), &input[0][0][0][0]); |
| 1563 | |
surmeh01 | 3537c2c | 2018-05-18 16:31:43 +0100 | [diff] [blame] | 1564 | workloadFactory.Finalize(); |
telsoa01 | 4fcda01 | 2018-03-09 14:13:49 +0000 | [diff] [blame] | 1565 | workload->Execute(); |
surmeh01 | 3537c2c | 2018-05-18 16:31:43 +0100 | [diff] [blame] | 1566 | refWorkloadFactory.Finalize(); |
telsoa01 | 4fcda01 | 2018-03-09 14:13:49 +0000 | [diff] [blame] | 1567 | workloadRef->Execute(); |
| 1568 | |
| 1569 | CopyDataFromITensorHandle(&ret.output[0][0][0][0], outputHandle.get()); |
| 1570 | CopyDataFromITensorHandle(&ret.outputExpected[0][0][0][0], outputHandleRef.get()); |
| 1571 | |
| 1572 | return ret; |
| 1573 | } |
| 1574 | |
surmeh01 | 3537c2c | 2018-05-18 16:31:43 +0100 | [diff] [blame] | 1575 | template<typename T> |
| 1576 | void PermuteTensorData( |
| 1577 | armnn::IWorkloadFactory& workloadFactory, |
| 1578 | const armnn::PermutationVector& mappings, |
| 1579 | armnn::TensorInfo & inputTensorInfo, |
| 1580 | const T * inputData, |
| 1581 | std::vector<T>& outputData) |
telsoa01 | 4fcda01 | 2018-03-09 14:13:49 +0000 | [diff] [blame] | 1582 | { |
surmeh01 | 3537c2c | 2018-05-18 16:31:43 +0100 | [diff] [blame] | 1583 | BOOST_ASSERT_MSG(inputData != nullptr, "inputData must not be null"); |
| 1584 | if (inputData == nullptr) |
| 1585 | { |
| 1586 | // Nullptr is an error in the test. By returning without doing the concatenation |
| 1587 | // I expect the caller to fail the test. It still makes sense to report this as |
| 1588 | // an assert for Debug builds. |
| 1589 | return; |
| 1590 | } |
telsoa01 | 4fcda01 | 2018-03-09 14:13:49 +0000 | [diff] [blame] | 1591 | |
surmeh01 | 3537c2c | 2018-05-18 16:31:43 +0100 | [diff] [blame] | 1592 | armnn::TensorInfo outputTensorInfo = armnnUtils::Permuted(inputTensorInfo, mappings); |
| 1593 | |
| 1594 | std::unique_ptr<armnn::ITensorHandle> inputHandle = workloadFactory.CreateTensorHandle(inputTensorInfo); |
| 1595 | std::unique_ptr<armnn::ITensorHandle> outputHandle = workloadFactory.CreateTensorHandle(outputTensorInfo); |
| 1596 | |
| 1597 | armnn::PermuteQueueDescriptor queueDescriptor; |
| 1598 | queueDescriptor.m_Parameters = armnn::PermuteDescriptor{mappings}; |
| 1599 | armnn::WorkloadInfo workloadInfo; |
| 1600 | AddInputToWorkload(queueDescriptor, workloadInfo, inputTensorInfo, inputHandle.get()); |
| 1601 | AddOutputToWorkload(queueDescriptor, workloadInfo, outputTensorInfo, outputHandle.get()); |
| 1602 | |
| 1603 | std::unique_ptr<armnn::IWorkload> workload = workloadFactory.CreatePermute(queueDescriptor, workloadInfo); |
| 1604 | |
| 1605 | inputHandle->Allocate(); |
| 1606 | outputHandle->Allocate(); |
| 1607 | |
| 1608 | CopyDataToITensorHandle(inputHandle.get(), inputData); |
| 1609 | |
| 1610 | workload->Execute(); |
| 1611 | |
| 1612 | outputData.resize(outputTensorInfo.GetNumElements()); |
| 1613 | CopyDataFromITensorHandle(&outputData[0], outputHandle.get()); |
| 1614 | inputTensorInfo = outputTensorInfo; |
| 1615 | } |
| 1616 | |
| 1617 | armnn::OriginsDescriptor CreateMergerDescriptorForConcatenation( |
| 1618 | const std::vector<armnn::TensorInfo> & inputTensorInfos, |
| 1619 | unsigned int concatDim) |
| 1620 | { |
telsoa01 | 4fcda01 | 2018-03-09 14:13:49 +0000 | [diff] [blame] | 1621 | std::vector<armnn::TensorShape> shapes; |
| 1622 | shapes.reserve(inputTensorInfos.size()); |
| 1623 | for (const armnn::TensorInfo& it: inputTensorInfos) |
| 1624 | { |
| 1625 | shapes.push_back(it.GetShape()); |
| 1626 | } |
surmeh01 | 3537c2c | 2018-05-18 16:31:43 +0100 | [diff] [blame] | 1627 | |
| 1628 | return armnn::CreateMergerDescriptorForConcatenation(shapes.begin(), |
| 1629 | shapes.end(), |
| 1630 | concatDim); |
| 1631 | } |
| 1632 | |
| 1633 | // |
| 1634 | // Concatenation is only supported for N and C dimensions for NCHW. In case of |
telsoa01 | c577f2c | 2018-08-31 09:22:23 +0100 | [diff] [blame] | 1635 | // <4 dimensions we need to make sure that the concat dimensions are at least |
surmeh01 | 3537c2c | 2018-05-18 16:31:43 +0100 | [diff] [blame] | 1636 | // the 3rd slowest iterating one. |
| 1637 | // |
| 1638 | |
| 1639 | bool NeedPermuteForConcat( |
| 1640 | const std::vector<armnn::TensorInfo> & inputTensorInfos, |
| 1641 | unsigned int concatDim) |
| 1642 | { |
| 1643 | // See note above. Additionally we expect the input shapes to have the |
| 1644 | // same number of dimensions. |
| 1645 | unsigned int nDimensions = 0; |
| 1646 | |
telsoa01 | c577f2c | 2018-08-31 09:22:23 +0100 | [diff] [blame] | 1647 | // Determine the number of dimensions as well as sanity check them |
| 1648 | // agains test implementation issues. |
surmeh01 | 3537c2c | 2018-05-18 16:31:43 +0100 | [diff] [blame] | 1649 | for (auto && tensorInfo : inputTensorInfos) |
| 1650 | { |
| 1651 | if (!nDimensions) |
| 1652 | { |
| 1653 | nDimensions = tensorInfo.GetShape().GetNumDimensions(); |
| 1654 | } |
| 1655 | else |
| 1656 | { |
| 1657 | BOOST_ASSERT_MSG(nDimensions == tensorInfo.GetShape().GetNumDimensions(), |
| 1658 | "Input shapes must have the same number of dimensions"); |
| 1659 | } |
| 1660 | } |
| 1661 | |
| 1662 | return (nDimensions-concatDim) < 3; |
| 1663 | } |
| 1664 | |
| 1665 | armnn::TensorShape ExpandTensorShapeTo3dForPermute(const armnn::TensorShape & inputShape) |
| 1666 | { |
| 1667 | unsigned int numDims = inputShape.GetNumDimensions(); |
| 1668 | if (numDims >= 3) |
| 1669 | { |
| 1670 | // Nothing to do if the inputShape has at least 3 dimensions. |
| 1671 | return inputShape; |
| 1672 | } |
| 1673 | |
| 1674 | std::vector<unsigned int> newDims(size_t(3), 1u); |
| 1675 | unsigned int expandedBy = 3 - numDims; |
| 1676 | for (unsigned int i=0; i<numDims; ++i) |
| 1677 | { |
| 1678 | newDims[expandedBy+i] = inputShape[i]; |
| 1679 | } |
| 1680 | return armnn::TensorShape(3u, &newDims[0]); |
| 1681 | } |
| 1682 | |
| 1683 | void Generate3dPermuteVectorForConcat( |
| 1684 | unsigned int numDimensions, |
| 1685 | unsigned int & concatDim, |
| 1686 | std::pair<armnn::PermutationVector, armnn::PermutationVector> & permutations) |
| 1687 | { |
| 1688 | BOOST_ASSERT_MSG(numDimensions <= 3, |
| 1689 | "Only dimensions 1,2 and 3 are supported by this helper"); |
| 1690 | |
| 1691 | unsigned int expandedBy = 3 - numDimensions; |
| 1692 | unsigned int expandedConcatAxis = concatDim + expandedBy; |
| 1693 | |
| 1694 | if (expandedConcatAxis == 2) |
| 1695 | { |
| 1696 | concatDim = 0; |
| 1697 | armnn::PermutationVector forwardPermutation({1, 2, 0}); |
| 1698 | armnn::PermutationVector reversePermutation({2, 0, 1}); |
| 1699 | permutations = std::make_pair(forwardPermutation, reversePermutation); |
| 1700 | } |
| 1701 | else if (expandedConcatAxis == 1) |
| 1702 | { |
| 1703 | concatDim = 0; |
| 1704 | armnn::PermutationVector forwardPermutation({2, 0, 1}); |
| 1705 | armnn::PermutationVector reversePermutation({1, 2, 0}); |
| 1706 | permutations = std::make_pair(forwardPermutation, reversePermutation); |
| 1707 | } |
| 1708 | else |
| 1709 | { |
| 1710 | BOOST_ASSERT(expandedConcatAxis == 0); |
| 1711 | concatDim = 0; |
| 1712 | } |
| 1713 | } |
| 1714 | |
| 1715 | // |
| 1716 | // Permute the input tensors so we can do a supported concatenation. |
| 1717 | // Also treat lower than 3d tensors as 3d by adding dummy 1 dimensions |
| 1718 | // at the front. Finally this function tells what the output shape |
| 1719 | // of the permuted concatenated tensor is going to be. |
| 1720 | // |
| 1721 | template <typename T> |
| 1722 | void PermuteInputsForConcat( |
| 1723 | armnn::IWorkloadFactory& workloadFactory, |
| 1724 | std::vector<armnn::TensorInfo> & inputTensorInfos, |
| 1725 | std::vector<T *> & inputData, |
| 1726 | std::vector<std::vector<T>> & inputDataStorage, |
| 1727 | armnn::PermutationVector & permuteVector, |
| 1728 | unsigned int & concatDim, |
| 1729 | armnn::TensorInfo & outputTensorInfo) |
| 1730 | { |
| 1731 | BOOST_ASSERT_MSG(inputTensorInfos.size() > 1, |
| 1732 | "Expecting more than one tensor to be concatenated here"); |
| 1733 | |
| 1734 | unsigned int numDims = 0; |
| 1735 | unsigned int nthInput = 0; |
| 1736 | const armnn::PermutationVector identity({0, 1, 2}); |
| 1737 | |
| 1738 | std::pair<armnn::PermutationVector, armnn::PermutationVector> permutations = |
| 1739 | std::make_pair(identity, identity); |
| 1740 | |
| 1741 | inputDataStorage.resize(inputData.size()); |
| 1742 | |
| 1743 | for (auto && tensorInfo : inputTensorInfos) |
| 1744 | { |
| 1745 | if (numDims == 0) |
| 1746 | { |
| 1747 | numDims = tensorInfo.GetShape().GetNumDimensions(); |
| 1748 | Generate3dPermuteVectorForConcat(numDims, concatDim, permutations); |
telsoa01 | c577f2c | 2018-08-31 09:22:23 +0100 | [diff] [blame] | 1749 | // Store the reverese permutation. |
surmeh01 | 3537c2c | 2018-05-18 16:31:43 +0100 | [diff] [blame] | 1750 | permuteVector = permutations.second; |
| 1751 | BOOST_ASSERT_MSG(!permuteVector.IsEqual(identity), |
| 1752 | "Test logic error, we don't need permutation, so we shouldn't arrive here"); |
| 1753 | } |
| 1754 | else |
| 1755 | { |
| 1756 | BOOST_ASSERT_MSG(numDims == tensorInfo.GetShape().GetNumDimensions(), |
| 1757 | "All inputs must have the same number of dimensions"); |
| 1758 | } |
| 1759 | |
| 1760 | armnn::TensorInfo newTensorInfo = tensorInfo; |
| 1761 | newTensorInfo.SetShape(ExpandTensorShapeTo3dForPermute(tensorInfo.GetShape())); |
| 1762 | |
| 1763 | PermuteTensorData<T>(workloadFactory, |
| 1764 | permutations.first, |
| 1765 | newTensorInfo, |
| 1766 | inputData[nthInput], |
| 1767 | inputDataStorage[nthInput]); |
| 1768 | |
| 1769 | inputData[nthInput] = inputDataStorage[nthInput].data(); |
| 1770 | inputTensorInfos[nthInput] = newTensorInfo; |
| 1771 | |
| 1772 | ++nthInput; |
| 1773 | } |
| 1774 | |
| 1775 | outputTensorInfo.SetShape( |
| 1776 | armnnUtils::Permuted( |
| 1777 | ExpandTensorShapeTo3dForPermute(outputTensorInfo.GetShape()), |
| 1778 | permutations.first)); |
| 1779 | } |
| 1780 | |
| 1781 | |
| 1782 | // |
| 1783 | // This is the pair of PermuteInputsForConcat(...) which permutes back |
telsoa01 | c577f2c | 2018-08-31 09:22:23 +0100 | [diff] [blame] | 1784 | // the output of the concatenation so we can check it against an expected |
surmeh01 | 3537c2c | 2018-05-18 16:31:43 +0100 | [diff] [blame] | 1785 | // output. |
| 1786 | // |
| 1787 | template <typename T> |
| 1788 | void PermuteOutputForConcat( |
| 1789 | armnn::IWorkloadFactory& workloadFactory, |
| 1790 | const armnn::TensorInfo & tensorInfo, |
| 1791 | const armnn::PermutationVector & permuteVector, |
| 1792 | std::unique_ptr<armnn::ITensorHandle> && inputDataHandle, |
| 1793 | T * data) |
| 1794 | { |
| 1795 | BOOST_ASSERT_MSG(data != nullptr, "data must not be null"); |
| 1796 | if (data == nullptr) |
| 1797 | { |
| 1798 | // Nullptr is an error in the test. By returning without doing the permutation |
| 1799 | // I expect the caller to fail the test. It still makes sense to report this as |
| 1800 | // an assert for Debug builds. |
| 1801 | return; |
| 1802 | } |
| 1803 | |
| 1804 | armnn::TensorInfo resultTensorInfo = tensorInfo; |
| 1805 | std::vector<T> inputData(tensorInfo.GetNumElements()); |
| 1806 | std::vector<T> outputData; |
| 1807 | |
| 1808 | CopyDataFromITensorHandle(&inputData[0], inputDataHandle.get()); |
| 1809 | |
| 1810 | PermuteTensorData<T>(workloadFactory, |
| 1811 | permuteVector, |
| 1812 | resultTensorInfo, |
| 1813 | &inputData[0], |
| 1814 | outputData); |
| 1815 | |
| 1816 | ::memcpy(data, &outputData[0], sizeof(T)*outputData.size()); |
| 1817 | } |
| 1818 | |
| 1819 | template <typename T> |
| 1820 | void Concatenate(armnn::IWorkloadFactory& workloadFactory, |
| 1821 | std::initializer_list<const armnn::TensorInfo> inputTensorInfosOrig, |
| 1822 | std::initializer_list<T *> inputsOrig, |
| 1823 | const armnn::TensorInfo& outputTensorInfoOrig, |
| 1824 | T * output, |
| 1825 | unsigned int concatDim) |
| 1826 | { |
| 1827 | BOOST_ASSERT_MSG(output != nullptr, "output must not be null"); |
| 1828 | if (output == nullptr) |
| 1829 | { |
| 1830 | // Nullptr is an error in the test. By returning without doing the permutation |
| 1831 | // I expect the caller to fail the test. It still makes sense to report this as |
| 1832 | // an assert for Debug builds. |
| 1833 | return; |
| 1834 | } |
| 1835 | |
| 1836 | armnn::MergerQueueDescriptor queueDescriptor; |
| 1837 | |
telsoa01 | c577f2c | 2018-08-31 09:22:23 +0100 | [diff] [blame] | 1838 | // Saves a copy of the parameters which we might need to change. |
surmeh01 | 3537c2c | 2018-05-18 16:31:43 +0100 | [diff] [blame] | 1839 | std::vector<armnn::TensorInfo> inputTensorInfos(inputTensorInfosOrig.begin(), inputTensorInfosOrig.end()); |
| 1840 | std::vector<T *> inputs = inputsOrig; |
| 1841 | armnn::TensorInfo outputTensorInfo = outputTensorInfoOrig; |
| 1842 | |
| 1843 | armnn::PermutationVector permuteVector{0, 1, 2}; |
| 1844 | |
telsoa01 | c577f2c | 2018-08-31 09:22:23 +0100 | [diff] [blame] | 1845 | // Holds and automatically releases memory for the reshaped input data. |
surmeh01 | 3537c2c | 2018-05-18 16:31:43 +0100 | [diff] [blame] | 1846 | std::vector<std::vector<T>> tmpInputDataStorage; |
| 1847 | |
| 1848 | const size_t inputCount = inputTensorInfos.size(); |
| 1849 | |
| 1850 | bool needPermuteForConcat = NeedPermuteForConcat(inputTensorInfos, concatDim); |
| 1851 | |
| 1852 | if (needPermuteForConcat) |
| 1853 | { |
| 1854 | // |
| 1855 | // We need to permute the inputs, because concatenation along |
telsoa01 | c577f2c | 2018-08-31 09:22:23 +0100 | [diff] [blame] | 1856 | // the requested axis is not supported. |
surmeh01 | 3537c2c | 2018-05-18 16:31:43 +0100 | [diff] [blame] | 1857 | // |
| 1858 | PermuteInputsForConcat<T>(workloadFactory, |
| 1859 | inputTensorInfos, |
| 1860 | inputs, |
| 1861 | tmpInputDataStorage, |
| 1862 | permuteVector, |
| 1863 | concatDim, |
| 1864 | outputTensorInfo); |
| 1865 | } |
| 1866 | |
| 1867 | armnn::OriginsDescriptor viewsDescriptor = CreateMergerDescriptorForConcatenation(inputTensorInfos, concatDim); |
telsoa01 | 4fcda01 | 2018-03-09 14:13:49 +0000 | [diff] [blame] | 1868 | |
| 1869 | queueDescriptor.m_ViewOrigins.reserve(viewsDescriptor.GetNumViews()); |
| 1870 | for (unsigned int i = 0; i < viewsDescriptor.GetNumViews(); ++i) |
| 1871 | { |
| 1872 | queueDescriptor.m_ViewOrigins.emplace_back(std::vector<unsigned int>(viewsDescriptor.GetViewOrigin(i), |
| 1873 | viewsDescriptor.GetViewOrigin(i) + viewsDescriptor.GetNumDimensions())); |
| 1874 | } |
| 1875 | |
telsoa01 | 4fcda01 | 2018-03-09 14:13:49 +0000 | [diff] [blame] | 1876 | std::unique_ptr<armnn::ITensorHandle> outputHandle = workloadFactory.CreateTensorHandle(outputTensorInfo); |
| 1877 | |
| 1878 | std::vector<std::unique_ptr<armnn::ITensorHandle>> inputHandles; |
| 1879 | inputHandles.reserve(inputCount); |
| 1880 | |
| 1881 | const bool subTensorsSupported = workloadFactory.SupportsSubTensors(); |
| 1882 | for (unsigned int i = 0; i < inputCount; ++i) |
| 1883 | { |
surmeh01 | 3537c2c | 2018-05-18 16:31:43 +0100 | [diff] [blame] | 1884 | const armnn::TensorInfo& inputTensorInfo = inputTensorInfos[i]; |
telsoa01 | 4fcda01 | 2018-03-09 14:13:49 +0000 | [diff] [blame] | 1885 | |
| 1886 | std::unique_ptr<armnn::ITensorHandle> inputHandle = subTensorsSupported ? |
| 1887 | workloadFactory.CreateSubTensorHandle(*outputHandle, inputTensorInfo.GetShape(), |
| 1888 | queueDescriptor.m_ViewOrigins[i].m_Origin.data()) |
| 1889 | : workloadFactory.CreateTensorHandle(inputTensorInfo); |
| 1890 | |
| 1891 | inputHandles.emplace_back(std::move(inputHandle)); |
| 1892 | } |
| 1893 | |
| 1894 | armnn::WorkloadInfo workloadInfo; |
| 1895 | |
| 1896 | for (unsigned int i = 0; i < inputCount; ++i) |
| 1897 | { |
surmeh01 | 3537c2c | 2018-05-18 16:31:43 +0100 | [diff] [blame] | 1898 | AddInputToWorkload(queueDescriptor, workloadInfo, inputTensorInfos[i], inputHandles[i].get()); |
telsoa01 | 4fcda01 | 2018-03-09 14:13:49 +0000 | [diff] [blame] | 1899 | } |
| 1900 | |
| 1901 | AddOutputToWorkload(queueDescriptor, workloadInfo, outputTensorInfo, outputHandle.get()); |
| 1902 | |
| 1903 | std::unique_ptr<armnn::IWorkload> workload = workloadFactory.CreateMerger(queueDescriptor, workloadInfo); |
| 1904 | |
| 1905 | for (auto& inputHandle : inputHandles) |
| 1906 | { |
| 1907 | inputHandle->Allocate(); |
| 1908 | } |
| 1909 | |
| 1910 | outputHandle->Allocate(); |
| 1911 | |
| 1912 | unsigned int nextInputId = 0; |
| 1913 | for (auto& inputHandle : inputHandles) |
| 1914 | { |
surmeh01 | 3537c2c | 2018-05-18 16:31:43 +0100 | [diff] [blame] | 1915 | CopyDataToITensorHandle(inputHandle.get(), inputs[nextInputId]); |
| 1916 | ++nextInputId; |
telsoa01 | 4fcda01 | 2018-03-09 14:13:49 +0000 | [diff] [blame] | 1917 | } |
| 1918 | |
surmeh01 | 3537c2c | 2018-05-18 16:31:43 +0100 | [diff] [blame] | 1919 | workloadFactory.Finalize(); |
telsoa01 | 4fcda01 | 2018-03-09 14:13:49 +0000 | [diff] [blame] | 1920 | workload->Execute(); |
| 1921 | |
surmeh01 | 3537c2c | 2018-05-18 16:31:43 +0100 | [diff] [blame] | 1922 | if (needPermuteForConcat) |
| 1923 | { |
| 1924 | PermuteOutputForConcat<T>(workloadFactory, |
| 1925 | outputTensorInfo, |
| 1926 | permuteVector, |
| 1927 | std::move(outputHandle), |
| 1928 | output); |
| 1929 | } |
| 1930 | else |
| 1931 | { |
| 1932 | CopyDataFromITensorHandle(output, outputHandle.get()); |
| 1933 | } |
telsoa01 | 4fcda01 | 2018-03-09 14:13:49 +0000 | [diff] [blame] | 1934 | } |
| 1935 | |
| 1936 | template <typename T> |
| 1937 | LayerTestResult<T, 1> Concatenation1dTestImpl(armnn::IWorkloadFactory& workloadFactory, float qScale, int32_t qOffset) |
| 1938 | { |
| 1939 | armnn::TensorInfo inputTensorInfo({ 3 }, armnn::GetDataType<T>()); |
| 1940 | |
| 1941 | auto input0 = MakeTensor<T, 1>(inputTensorInfo, QuantizedVector<T>(qScale, qOffset, { 1.0f, 2.0f, 3.0f })); |
| 1942 | auto input1 = MakeTensor<T, 1>(inputTensorInfo, QuantizedVector<T>(qScale, qOffset, { 4.0f, 5.0f, 6.0f })); |
| 1943 | auto input2 = MakeTensor<T, 1>(inputTensorInfo, QuantizedVector<T>(qScale, qOffset, { 7.0f, 8.0f, 9.0f })); |
| 1944 | |
| 1945 | armnn::TensorInfo outputTensorInfo({ 9 }, armnn::GetDataType<T>()); |
| 1946 | |
| 1947 | LayerTestResult<T, 1> result(outputTensorInfo); |
| 1948 | |
| 1949 | std::vector<T> output; |
| 1950 | output.resize(outputTensorInfo.GetNumElements()); |
surmeh01 | 3537c2c | 2018-05-18 16:31:43 +0100 | [diff] [blame] | 1951 | Concatenate<T>(workloadFactory, |
telsoa01 | 4fcda01 | 2018-03-09 14:13:49 +0000 | [diff] [blame] | 1952 | { inputTensorInfo, inputTensorInfo, inputTensorInfo }, |
| 1953 | { input0.data(), input1.data(), input2.data() }, |
| 1954 | outputTensorInfo, |
| 1955 | output.data(), |
| 1956 | 0); |
| 1957 | |
| 1958 | result.output = MakeTensor<T, 1>(outputTensorInfo, output); |
| 1959 | result.outputExpected = MakeTensor<T, 1>(outputTensorInfo, QuantizedVector<T>(qScale, qOffset, { |
| 1960 | 1.0f, 2.0f, 3.0f, 4.0f, 5.0f, 6.0f, 7.0f, 8.0f, 9.0f |
| 1961 | })); |
| 1962 | |
| 1963 | return result; |
| 1964 | } |
| 1965 | |
| 1966 | LayerTestResult<float, 1> Concatenation1dTest(armnn::IWorkloadFactory& workloadFactory) |
| 1967 | { |
| 1968 | return Concatenation1dTestImpl<float>(workloadFactory, 0.0f, 0); |
| 1969 | } |
| 1970 | |
| 1971 | template <typename T> |
| 1972 | LayerTestResult<T, 2> Concatenation2dTestImpl(armnn::IWorkloadFactory& workloadFactory, |
| 1973 | const armnn::TensorInfo& outputTensorInfo, |
| 1974 | unsigned int dimension, |
| 1975 | const float qScale, |
| 1976 | const int32_t qOffset) |
| 1977 | { |
| 1978 | armnn::TensorInfo inputTensorInfo({ 2, 3 }, armnn::GetDataType<T>()); |
| 1979 | |
| 1980 | auto input0 = MakeTensor<T, 2>(inputTensorInfo, QuantizedVector<T>(qScale, qOffset, { |
| 1981 | // Batch 0 |
| 1982 | 1.0f, 2.0f, 3.0f, |
| 1983 | |
| 1984 | // Batch 1 |
| 1985 | 10.0f, 11.0f, 12.0f, |
| 1986 | })); |
| 1987 | |
| 1988 | auto input1 = MakeTensor<T, 2>(inputTensorInfo, QuantizedVector<T>(qScale, qOffset, { |
| 1989 | // Batch 0 |
| 1990 | 4.0f, 5.0f, 6.0f, |
| 1991 | |
| 1992 | // Batch 1 |
| 1993 | 13.0f, 14.0f, 15.0f, |
| 1994 | })); |
| 1995 | |
| 1996 | auto input2 = MakeTensor<T, 2>(inputTensorInfo, QuantizedVector<T>(qScale, qOffset, { |
| 1997 | // Batch 0 |
| 1998 | 7.0f, 8.0f, 9.0f, |
| 1999 | |
| 2000 | // Batch 1 |
| 2001 | 16.0f, 17.0f, 18.0f, |
| 2002 | })); |
| 2003 | |
| 2004 | LayerTestResult<T, 2> result(outputTensorInfo); |
| 2005 | |
| 2006 | std::vector<T> output; |
| 2007 | output.resize(outputTensorInfo.GetNumElements()); |
surmeh01 | 3537c2c | 2018-05-18 16:31:43 +0100 | [diff] [blame] | 2008 | Concatenate<T>(workloadFactory, |
telsoa01 | 4fcda01 | 2018-03-09 14:13:49 +0000 | [diff] [blame] | 2009 | { inputTensorInfo, inputTensorInfo, inputTensorInfo }, |
| 2010 | { input0.data(), input1.data(), input2.data() }, |
| 2011 | outputTensorInfo, |
| 2012 | output.data(), |
| 2013 | dimension); |
| 2014 | |
| 2015 | result.output = MakeTensor<T, 2>(outputTensorInfo, output); |
| 2016 | return result; |
| 2017 | } |
| 2018 | |
| 2019 | template <typename T> |
| 2020 | LayerTestResult<T, 2> Concatenation2dDim0TestImpl(armnn::IWorkloadFactory& workloadFactory, |
| 2021 | float qScale, int32_t qOffset) |
| 2022 | { |
| 2023 | armnn::TensorInfo outputTensorInfo({ 6, 3 }, armnn::GetDataType<T>()); |
| 2024 | |
| 2025 | LayerTestResult<T, 2> result = Concatenation2dTestImpl<T>(workloadFactory, outputTensorInfo, 0, qScale, qOffset); |
| 2026 | result.outputExpected = MakeTensor<T, 2>(outputTensorInfo, QuantizedVector<T>(qScale, qOffset, { |
| 2027 | // Batch 0 |
| 2028 | 1.0f, 2.0f, 3.0f, |
| 2029 | |
| 2030 | // Batch 1 |
| 2031 | 10.0f, 11.0f, 12.0f, |
| 2032 | |
| 2033 | // Batch 2 |
| 2034 | 4.0f, 5.0f, 6.0f, |
| 2035 | |
| 2036 | // Batch 3 |
| 2037 | 13.0f, 14.0f, 15.0f, |
| 2038 | |
| 2039 | // Batch 4 |
| 2040 | 7.0f, 8.0f, 9.0f, |
| 2041 | |
| 2042 | // Batch 5 |
| 2043 | 16.0f, 17.0f, 18.0f, |
| 2044 | })); |
| 2045 | |
| 2046 | return result; |
| 2047 | } |
| 2048 | |
| 2049 | LayerTestResult<float, 2> Concatenation2dDim0Test(armnn::IWorkloadFactory& workloadFactory) |
| 2050 | { |
| 2051 | return Concatenation2dDim0TestImpl<float>(workloadFactory, 0.0f, 0); |
| 2052 | } |
| 2053 | |
| 2054 | template <typename T> |
| 2055 | LayerTestResult<T, 2> Concatenation2dDim1TestImpl(armnn::IWorkloadFactory& workloadFactory, |
| 2056 | float qScale, int32_t qOffset) |
| 2057 | { |
| 2058 | armnn::TensorInfo outputTensorInfo({ 2, 9 }, armnn::GetDataType<T>()); |
| 2059 | |
| 2060 | LayerTestResult<T, 2> result = Concatenation2dTestImpl<T>(workloadFactory, outputTensorInfo, 1, qScale, qOffset); |
| 2061 | result.outputExpected = MakeTensor<T, 2>(outputTensorInfo, QuantizedVector<T>(qScale, qOffset, { |
| 2062 | // Batch 0 |
| 2063 | 1.0f, 2.0f, 3.0f, 4.0f, 5.0f, 6.0f, 7.0f, 8.0f, 9.0f, |
| 2064 | |
| 2065 | // Batch 1 |
| 2066 | 10.0f, 11.0f, 12.0f, 13.0f, 14.0f, 15.0f, 16.0f, 17.0f, 18.0f |
| 2067 | })); |
| 2068 | |
| 2069 | return result; |
| 2070 | } |
| 2071 | |
| 2072 | LayerTestResult<float, 2> Concatenation2dDim1Test(armnn::IWorkloadFactory& workloadFactory) |
| 2073 | { |
| 2074 | return Concatenation2dDim1TestImpl<float>(workloadFactory, 0.0f, 0); |
| 2075 | } |
| 2076 | |
| 2077 | template <typename T> |
| 2078 | LayerTestResult<T, 2> Concatenation2dDim0DiffInputDimsTestImpl(armnn::IWorkloadFactory& workloadFactory, float qScale, |
| 2079 | int32_t qOffset) |
| 2080 | { |
| 2081 | armnn::TensorInfo input0TensorInfo({ 2, 3 }, armnn::GetDataType<T>()); |
| 2082 | auto input0 = MakeTensor<T, 2>(input0TensorInfo, QuantizedVector<T>(qScale, qOffset, { |
| 2083 | // Batch 0 |
| 2084 | 1.0f, 2.0f, 3.0f, |
| 2085 | |
| 2086 | // Batch 1 |
| 2087 | 10.0f, 11.0f, 12.0f, |
| 2088 | })); |
| 2089 | |
| 2090 | armnn::TensorInfo input1TensorInfo({ 3, 3 }, armnn::GetDataType<T>()); |
| 2091 | auto input1 = MakeTensor<T, 2>(input1TensorInfo, QuantizedVector<T>(qScale, qOffset, { |
| 2092 | // Batch 0 |
| 2093 | 4.0f, 5.0f, 6.0f, |
| 2094 | |
| 2095 | // Batch 1 |
| 2096 | 13.0f, 14.0f, 15.0f, |
| 2097 | |
| 2098 | // Batch 0 |
| 2099 | 7.0f, 8.0f, 9.0f, |
| 2100 | })); |
| 2101 | |
| 2102 | armnn::TensorInfo input2TensorInfo({ 1, 3 }, armnn::GetDataType<T>()); |
| 2103 | auto input2 = MakeTensor<T, 2>(input2TensorInfo, QuantizedVector<T>(qScale, qOffset, { |
| 2104 | // Batch 1 |
| 2105 | 16.0f, 17.0f, 18.0f, |
| 2106 | })); |
| 2107 | |
| 2108 | armnn::TensorInfo outputTensorInfo({ 6, 3 }, armnn::GetDataType<T>()); |
| 2109 | LayerTestResult<T, 2> result(outputTensorInfo); |
| 2110 | |
| 2111 | std::vector<T> output; |
| 2112 | output.resize(outputTensorInfo.GetNumElements()); |
surmeh01 | 3537c2c | 2018-05-18 16:31:43 +0100 | [diff] [blame] | 2113 | Concatenate<T>(workloadFactory, |
telsoa01 | 4fcda01 | 2018-03-09 14:13:49 +0000 | [diff] [blame] | 2114 | { input0TensorInfo, input1TensorInfo, input2TensorInfo }, |
| 2115 | { input0.data(), input1.data(), input2.data() }, |
| 2116 | outputTensorInfo, |
| 2117 | output.data(), |
| 2118 | 0); |
| 2119 | |
| 2120 | result.output = MakeTensor<T, 2>(outputTensorInfo, output); |
| 2121 | result.outputExpected = MakeTensor<T, 2>(outputTensorInfo, QuantizedVector<T>(qScale, qOffset, { |
| 2122 | // Batch 0 |
| 2123 | 1.0f, 2.0f, 3.0f, |
| 2124 | |
| 2125 | // Batch 1 |
| 2126 | 10.0f, 11.0f, 12.0f, |
| 2127 | |
| 2128 | // Batch 2 |
| 2129 | 4.0f, 5.0f, 6.0f, |
| 2130 | |
| 2131 | // Batch 3 |
| 2132 | 13.0f, 14.0f, 15.0f, |
| 2133 | |
| 2134 | // Batch 4 |
| 2135 | 7.0f, 8.0f, 9.0f, |
| 2136 | |
| 2137 | // Batch 5 |
| 2138 | 16.0f, 17.0f, 18.0f, |
| 2139 | })); |
| 2140 | |
| 2141 | return result; |
| 2142 | } |
| 2143 | |
| 2144 | LayerTestResult<float, 2> Concatenation2dDim0DiffInputDimsTest(armnn::IWorkloadFactory& workloadFactory) |
| 2145 | { |
| 2146 | return Concatenation2dDim0DiffInputDimsTestImpl<float>(workloadFactory, 0.0f, 0); |
| 2147 | } |
| 2148 | |
| 2149 | template <typename T> |
| 2150 | LayerTestResult<T, 2> Concatenation2dDim1DiffInputDimsTestImpl(armnn::IWorkloadFactory& workloadFactory, float qScale, |
| 2151 | int32_t qOffset) |
| 2152 | { |
| 2153 | armnn::TensorInfo input0TensorInfo({ 2, 3 }, armnn::GetDataType<T>()); |
| 2154 | auto input0 = MakeTensor<T, 2>(input0TensorInfo, QuantizedVector<T>(qScale, qOffset, { |
| 2155 | // Batch 0 |
| 2156 | 1.0f, 2.0f, 3.0f, |
| 2157 | |
| 2158 | // Batch 1 |
| 2159 | 10.0f, 11.0f, 12.0f, |
| 2160 | })); |
| 2161 | |
| 2162 | armnn::TensorInfo input1TensorInfo({ 2, 5 }, armnn::GetDataType<T>()); |
| 2163 | auto input1 = MakeTensor<T, 2>(input1TensorInfo, QuantizedVector<T>(qScale, qOffset, { |
| 2164 | // Batch 0 |
| 2165 | 4.0f, 5.0f, 6.0f, 7.0f, 8.0f, |
| 2166 | |
| 2167 | // Batch 1 |
| 2168 | 13.0f, 14.0f, 15.0f, 16.0f, 17.0f, |
| 2169 | })); |
| 2170 | |
| 2171 | armnn::TensorInfo input2TensorInfo({ 2, 1 }, armnn::GetDataType<T>()); |
| 2172 | auto input2 = MakeTensor<T, 2>(input2TensorInfo, QuantizedVector<T>(qScale, qOffset, { |
| 2173 | // Batch 0 |
| 2174 | 9.0f, |
| 2175 | |
| 2176 | // Batch 1 |
| 2177 | 18.0f |
| 2178 | })); |
| 2179 | |
| 2180 | armnn::TensorInfo outputTensorInfo({ 2, 9 }, armnn::GetDataType<T>()); |
| 2181 | LayerTestResult<T, 2> result(outputTensorInfo); |
| 2182 | |
| 2183 | std::vector<T> output; |
| 2184 | output.resize(outputTensorInfo.GetNumElements()); |
surmeh01 | 3537c2c | 2018-05-18 16:31:43 +0100 | [diff] [blame] | 2185 | Concatenate<T>(workloadFactory, |
telsoa01 | 4fcda01 | 2018-03-09 14:13:49 +0000 | [diff] [blame] | 2186 | { input0TensorInfo, input1TensorInfo, input2TensorInfo }, |
| 2187 | { input0.data(), input1.data(), input2.data() }, |
| 2188 | outputTensorInfo, |
| 2189 | output.data(), |
| 2190 | 1); |
| 2191 | |
| 2192 | result.output = MakeTensor<T, 2>(outputTensorInfo, output); |
| 2193 | result.outputExpected = MakeTensor<T, 2>(outputTensorInfo, QuantizedVector<T>(qScale, qOffset, { |
| 2194 | // Batch 0 |
| 2195 | 1.0f, 2.0f, 3.0f, 4.0f, 5.0f, 6.0f, 7.0f, 8.0f, 9.0f, |
| 2196 | |
| 2197 | // Batch 1 |
| 2198 | 10.0f, 11.0f, 12.0f, 13.0f, 14.0f, 15.0f, 16.0f, 17.0f, 18.0f, |
| 2199 | })); |
| 2200 | |
| 2201 | return result; |
| 2202 | } |
| 2203 | |
| 2204 | LayerTestResult<float, 2> Concatenation2dDim1DiffInputDimsTest(armnn::IWorkloadFactory& workloadFactory) |
| 2205 | { |
| 2206 | return Concatenation2dDim1DiffInputDimsTestImpl<float>(workloadFactory, 0.0f, 0); |
| 2207 | } |
| 2208 | |
| 2209 | template <typename T> |
| 2210 | LayerTestResult<T, 3> Concatenation3dTestImpl(armnn::IWorkloadFactory& workloadFactory, |
| 2211 | const armnn::TensorInfo& outputTensorInfo, |
| 2212 | unsigned int dimension, |
| 2213 | float qScale, |
| 2214 | int32_t qOffset) |
| 2215 | { |
| 2216 | armnn::TensorInfo inputTensorInfo({ 2, 3, 2 }, armnn::GetDataType<T>()); |
| 2217 | |
| 2218 | auto input0 = MakeTensor<T, 3>(inputTensorInfo, QuantizedVector<T>(qScale, qOffset, { |
| 2219 | // Batch 0, Channel 0 |
| 2220 | 1.0f, 2.0f, |
| 2221 | |
| 2222 | // Batch 0, Channel 1 |
| 2223 | 3.0f, 4.0f, |
| 2224 | |
| 2225 | // Batch 0, Channel 2 |
| 2226 | 5.0f, 6.0f, |
| 2227 | |
| 2228 | // Batch 1, Channel 0 |
| 2229 | 19.0f, 20.0f, |
| 2230 | |
| 2231 | // Batch 1, Channel 1 |
| 2232 | 21.0f, 22.0f, |
| 2233 | |
| 2234 | // Batch 1, Channel 2 |
| 2235 | 23.0f, 24.0f |
| 2236 | })); |
| 2237 | |
| 2238 | auto input1 = MakeTensor<T, 3>(inputTensorInfo, QuantizedVector<T>(qScale, qOffset, { |
| 2239 | // Batch 0, Channel 0 |
| 2240 | 7.0f, 8.0f, |
| 2241 | |
| 2242 | // Batch 0, Channel 1 |
| 2243 | 9.0f, 10.0f, |
| 2244 | |
| 2245 | // Batch 0, Channel 2 |
| 2246 | 11.0f, 12.0f, |
| 2247 | |
| 2248 | // Batch 1, Channel 0 |
| 2249 | 25.0f, 26.0f, |
| 2250 | |
| 2251 | // Batch 1, Channel 1 |
| 2252 | 27.0f, 28.0f, |
| 2253 | |
| 2254 | // Batch 1, Channel 2 |
| 2255 | 29.0f, 30.0f |
| 2256 | })); |
| 2257 | |
| 2258 | auto input2 = MakeTensor<T, 3>(inputTensorInfo, QuantizedVector<T>(qScale, qOffset, { |
| 2259 | // Batch 0, Channel 0 |
| 2260 | 13.0f, 14.0f, |
| 2261 | |
| 2262 | // Batch 0, Channel 1 |
| 2263 | 15.0f, 16.0f, |
| 2264 | |
| 2265 | // Batch 0, Channel 2 |
| 2266 | 17.0f, 18.0f, |
| 2267 | |
| 2268 | // Batch 1, Channel 0 |
| 2269 | 31.0f, 32.0f, |
| 2270 | |
| 2271 | // Batch 1, Channel 1 |
| 2272 | 33.0f, 34.0f, |
| 2273 | |
| 2274 | // Batch 1, Channel 2 |
| 2275 | 35.0f, 36.0f |
| 2276 | })); |
| 2277 | |
| 2278 | LayerTestResult<T, 3> result(outputTensorInfo); |
| 2279 | |
| 2280 | std::vector<T> output; |
| 2281 | output.resize(outputTensorInfo.GetNumElements()); |
surmeh01 | 3537c2c | 2018-05-18 16:31:43 +0100 | [diff] [blame] | 2282 | Concatenate<T>(workloadFactory, |
telsoa01 | 4fcda01 | 2018-03-09 14:13:49 +0000 | [diff] [blame] | 2283 | { inputTensorInfo, inputTensorInfo, inputTensorInfo }, |
| 2284 | { input0.data(), input1.data(), input2.data() }, |
| 2285 | outputTensorInfo, |
| 2286 | output.data(), |
| 2287 | dimension); |
| 2288 | |
| 2289 | result.output = MakeTensor<T, 3>(outputTensorInfo, output); |
| 2290 | return result; |
| 2291 | } |
| 2292 | |
| 2293 | template <typename T> |
| 2294 | LayerTestResult<T, 3> Concatenation3dDim0TestImpl(armnn::IWorkloadFactory& workloadFactory, float qScale, |
| 2295 | int32_t qOffset) |
| 2296 | { |
| 2297 | armnn::TensorInfo outputTensorInfo({ 6, 3, 2 }, armnn::GetDataType<T>()); |
| 2298 | |
| 2299 | LayerTestResult<T, 3> result = Concatenation3dTestImpl<T>(workloadFactory, outputTensorInfo, 0, |
| 2300 | qScale, qOffset); |
| 2301 | result.outputExpected = MakeTensor<T, 3>(outputTensorInfo, QuantizedVector<T>(qScale, qOffset, { |
| 2302 | // Batch 0, Channel 0 |
| 2303 | 1.0f, 2.0f, |
| 2304 | |
| 2305 | // Batch 0, Channel 1 |
| 2306 | 3.0f, 4.0f, |
| 2307 | |
| 2308 | // Batch 0, Channel 2 |
| 2309 | 5.0f, 6.0f, |
| 2310 | |
| 2311 | // Batch 1, Channel 0 |
| 2312 | 19.0f, 20.0f, |
| 2313 | |
| 2314 | // Batch 1, Channel 1 |
| 2315 | 21.0f, 22.0f, |
| 2316 | |
| 2317 | // Batch 1, Channel 2 |
| 2318 | 23.0f, 24.0f, |
| 2319 | |
| 2320 | // Batch 2, Channel 0 |
| 2321 | 7.0f, 8.0f, |
| 2322 | |
| 2323 | // Batch 2, Channel 1 |
| 2324 | 9.0f, 10.0f, |
| 2325 | |
| 2326 | // Batch 2, Channel 2 |
| 2327 | 11.0f, 12.0f, |
| 2328 | |
| 2329 | // Batch 3, Channel 0 |
| 2330 | 25.0f, 26.0f, |
| 2331 | |
| 2332 | // Batch 3, Channel 1 |
| 2333 | 27.0f, 28.0f, |
| 2334 | |
| 2335 | // Batch 3, Channel 2 |
| 2336 | 29.0f, 30.0f, |
| 2337 | |
| 2338 | // Batch 4, Channel 0 |
| 2339 | 13.0f, 14.0f, |
| 2340 | |
| 2341 | // Batch 4, Channel 1 |
| 2342 | 15.0f, 16.0f, |
| 2343 | |
| 2344 | // Batch 4, Channel 2 |
| 2345 | 17.0f, 18.0f, |
| 2346 | |
| 2347 | // Batch 5, Channel 0 |
| 2348 | 31.0f, 32.0f, |
| 2349 | |
| 2350 | // Batch 5, Channel 1 |
| 2351 | 33.0f, 34.0f, |
| 2352 | |
| 2353 | // Batch 5, Channel 2 |
| 2354 | 35.0f, 36.0f |
| 2355 | })); |
| 2356 | return result; |
| 2357 | } |
| 2358 | |
| 2359 | LayerTestResult<float, 3> Concatenation3dDim0Test(armnn::IWorkloadFactory& workloadFactory) |
| 2360 | { |
| 2361 | return Concatenation3dDim0TestImpl<float>(workloadFactory, 0.0f, 0); |
| 2362 | } |
| 2363 | |
| 2364 | template <typename T> |
| 2365 | LayerTestResult<T, 3> Concatenation3dDim1TestImpl(armnn::IWorkloadFactory& workloadFactory, |
| 2366 | float qScale, int32_t qOffset) |
| 2367 | { |
| 2368 | armnn::TensorInfo outputTensorInfo({ 2, 9, 2 }, armnn::GetDataType<T>()); |
| 2369 | |
| 2370 | LayerTestResult<T, 3> result = Concatenation3dTestImpl<T>(workloadFactory, outputTensorInfo, 1, qScale, qOffset); |
| 2371 | result.outputExpected = MakeTensor<T, 3>(outputTensorInfo, QuantizedVector<T>(qScale, qOffset, { |
| 2372 | // Batch 0, Channel 0 |
| 2373 | 1.0f, 2.0f, |
| 2374 | |
| 2375 | // Batch 0, Channel 1 |
| 2376 | 3.0f, 4.0f, |
| 2377 | |
| 2378 | // Batch 0, Channel 2 |
| 2379 | 5.0f, 6.0f, |
| 2380 | |
| 2381 | // Batch 0, Channel 3 |
| 2382 | 7.0f, 8.0f, |
| 2383 | |
| 2384 | // Batch 0, Channel 4 |
| 2385 | 9.0f, 10.0f, |
| 2386 | |
| 2387 | // Batch 0, Channel 5 |
| 2388 | 11.0f, 12.0f, |
| 2389 | |
| 2390 | // Batch 0, Channel 6 |
| 2391 | 13.0f, 14.0f, |
| 2392 | |
| 2393 | // Batch 0, Channel 7 |
| 2394 | 15.0f, 16.0f, |
| 2395 | |
| 2396 | // Batch 0, Channel 8 |
| 2397 | 17.0f, 18.0f, |
| 2398 | |
| 2399 | // Batch 1, Channel 0 |
| 2400 | 19.0f, 20.0f, |
| 2401 | |
| 2402 | // Batch 1, Channel 1 |
| 2403 | 21.0f, 22.0f, |
| 2404 | |
| 2405 | // Batch 1, Channel 2 |
| 2406 | 23.0f, 24.0f, |
| 2407 | |
| 2408 | // Batch 1, Channel 3 |
| 2409 | 25.0f, 26.0f, |
| 2410 | |
| 2411 | // Batch 1, Channel 4 |
| 2412 | 27.0f, 28.0f, |
| 2413 | |
| 2414 | // Batch 1, Channel 5 |
| 2415 | 29.0f, 30.0f, |
| 2416 | |
| 2417 | // Batch 1, Channel 6 |
| 2418 | 31.0f, 32.0f, |
| 2419 | |
| 2420 | // Batch 1, Channel 7 |
| 2421 | 33.0f, 34.0f, |
| 2422 | |
| 2423 | // Batch 1, Channel 8 |
| 2424 | 35.0f, 36.0f |
| 2425 | })); |
| 2426 | |
| 2427 | return result; |
| 2428 | } |
| 2429 | |
| 2430 | LayerTestResult<float, 3> Concatenation3dDim1Test(armnn::IWorkloadFactory& workloadFactory) |
| 2431 | { |
| 2432 | return Concatenation3dDim1TestImpl<float>(workloadFactory, 0.0f, 0); |
| 2433 | } |
| 2434 | |
| 2435 | template <typename T> |
| 2436 | LayerTestResult<T, 3> Concatenation3dDim2TestImpl(armnn::IWorkloadFactory& workloadFactory, |
| 2437 | float qScale, int32_t qOffset) |
| 2438 | { |
| 2439 | armnn::TensorInfo outputTensorInfo({ 2, 3, 6 }, armnn::GetDataType<T>()); |
| 2440 | |
| 2441 | LayerTestResult<T, 3> result = Concatenation3dTestImpl<T>(workloadFactory, outputTensorInfo, 2, qScale, qOffset); |
| 2442 | result.outputExpected = MakeTensor<T, 3>(outputTensorInfo, QuantizedVector<T>(qScale, qOffset, { |
| 2443 | // Batch 0, Channel 0 |
| 2444 | 1.0f, 2.0f, 7.0f, 8.0f, 13.0f, 14.0f, |
| 2445 | |
| 2446 | // Batch 0, Channel 1 |
| 2447 | 3.0f, 4.0f, 9.0f, 10.0f, 15.0f, 16.0f, |
| 2448 | |
| 2449 | // Batch 0, Channel 2 |
| 2450 | 5.0f, 6.0f, 11.0f, 12.0f, 17.0f, 18.0f, |
| 2451 | |
| 2452 | // Batch 1, Channel 0 |
| 2453 | 19.0f, 20.0f, 25.0f, 26.0f, 31.0f, 32.0f, |
| 2454 | |
| 2455 | // Batch 1, Channel 1 |
| 2456 | 21.0f, 22.0f, 27.0f, 28.0f, 33.0f, 34.0f, |
| 2457 | |
| 2458 | // Batch 1, Channel 2 |
| 2459 | 23.0f, 24.0f, 29.0f, 30.0f, 35.0f, 36.0f, |
| 2460 | })); |
| 2461 | |
| 2462 | return result; |
| 2463 | } |
| 2464 | |
| 2465 | LayerTestResult<float, 3> Concatenation3dDim2Test(armnn::IWorkloadFactory& workloadFactory) |
| 2466 | { |
| 2467 | return Concatenation3dDim2TestImpl<float>(workloadFactory, 0.0f, 0); |
| 2468 | } |
| 2469 | |
| 2470 | template <typename T> |
| 2471 | LayerTestResult<T, 3> Concatenation3dDim0DiffInputDimsTestImpl(armnn::IWorkloadFactory& workloadFactory, float qScale, |
| 2472 | int32_t qOffset) |
| 2473 | { |
| 2474 | armnn::TensorInfo input0TensorInfo({ 2, 3, 2 }, armnn::GetDataType<T>()); |
| 2475 | auto input0 = MakeTensor<T, 3>(input0TensorInfo, QuantizedVector<T>(qScale, qOffset, { |
| 2476 | // Batch 0, Channel 0 |
| 2477 | 1.0f, 2.0f, |
| 2478 | |
| 2479 | // Batch 0, Channel 1 |
| 2480 | 3.0f, 4.0f, |
| 2481 | |
| 2482 | // Batch 0, Channel 2 |
| 2483 | 5.0f, 6.0f, |
| 2484 | |
| 2485 | // Batch 1, Channel 0 |
| 2486 | 19.0f, 20.0f, |
| 2487 | |
| 2488 | // Batch 1, Channel 1 |
| 2489 | 21.0f, 22.0f, |
| 2490 | |
| 2491 | // Batch 1, Channel 2 |
| 2492 | 23.0f, 24.0f |
| 2493 | })); |
| 2494 | |
| 2495 | armnn::TensorInfo input1TensorInfo({ 1, 3, 2 }, armnn::GetDataType<T>()); |
| 2496 | auto input1 = MakeTensor<T, 3>(input1TensorInfo, QuantizedVector<T>(qScale, qOffset, { |
| 2497 | // Batch 0, Channel 0 |
| 2498 | 7.0f, 8.0f, |
| 2499 | |
| 2500 | // Batch 0, Channel 1 |
| 2501 | 9.0f, 10.0f, |
| 2502 | |
| 2503 | // Batch 0, Channel 2 |
| 2504 | 11.0f, 12.0f, |
| 2505 | })); |
| 2506 | |
| 2507 | armnn::TensorInfo input2TensorInfo({ 3, 3, 2 }, armnn::GetDataType<T>()); |
| 2508 | auto input2 = MakeTensor<T, 3>(input2TensorInfo, QuantizedVector<T>(qScale, qOffset, { |
| 2509 | // Batch 0, Channel 0 |
| 2510 | 25.0f, 26.0f, |
| 2511 | |
| 2512 | // Batch 0, Channel 1 |
| 2513 | 27.0f, 28.0f, |
| 2514 | |
| 2515 | // Batch 0, Channel 2 |
| 2516 | 29.0f, 30.0f, |
| 2517 | |
| 2518 | // Batch 1, Channel 0 |
| 2519 | 13.0f, 14.0f, |
| 2520 | |
| 2521 | // Batch 1, Channel 1 |
| 2522 | 15.0f, 16.0f, |
| 2523 | |
| 2524 | // Batch 1, Channel 2 |
| 2525 | 17.0f, 18.0f, |
| 2526 | |
| 2527 | // Batch 2, Channel 0 |
| 2528 | 31.0f, 32.0f, |
| 2529 | |
| 2530 | // Batch 2, Channel 1 |
| 2531 | 33.0f, 34.0f, |
| 2532 | |
| 2533 | // Batch 2, Channel 2 |
| 2534 | 35.0f, 36.0f |
| 2535 | })); |
| 2536 | |
| 2537 | armnn::TensorInfo outputTensorInfo({ 6, 3, 2 }, armnn::GetDataType<T>()); |
| 2538 | LayerTestResult<T, 3> result(outputTensorInfo); |
| 2539 | |
| 2540 | std::vector<T> output; |
| 2541 | output.resize(outputTensorInfo.GetNumElements()); |
surmeh01 | 3537c2c | 2018-05-18 16:31:43 +0100 | [diff] [blame] | 2542 | Concatenate<T>(workloadFactory, |
telsoa01 | 4fcda01 | 2018-03-09 14:13:49 +0000 | [diff] [blame] | 2543 | { input0TensorInfo, input1TensorInfo, input2TensorInfo }, |
| 2544 | { input0.data(), input1.data(), input2.data() }, |
| 2545 | outputTensorInfo, |
| 2546 | output.data(), |
| 2547 | 0); |
| 2548 | |
| 2549 | result.output = MakeTensor<T, 3>(outputTensorInfo, output); |
| 2550 | result.outputExpected = MakeTensor<T, 3>(outputTensorInfo, QuantizedVector<T>(qScale, qOffset, { |
| 2551 | // Batch 0, Channel 0 |
| 2552 | 1.0f, 2.0f, |
| 2553 | |
| 2554 | // Batch 0, Channel 1 |
| 2555 | 3.0f, 4.0f, |
| 2556 | |
| 2557 | // Batch 0, Channel 2 |
| 2558 | 5.0f, 6.0f, |
| 2559 | |
| 2560 | // Batch 1, Channel 0 |
| 2561 | 19.0f, 20.0f, |
| 2562 | |
| 2563 | // Batch 1, Channel 1 |
| 2564 | 21.0f, 22.0f, |
| 2565 | |
| 2566 | // Batch 1, Channel 2 |
| 2567 | 23.0f, 24.0f, |
| 2568 | |
| 2569 | // Batch 2, Channel 0 |
| 2570 | 7.0f, 8.0f, |
| 2571 | |
| 2572 | // Batch 2, Channel 1 |
| 2573 | 9.0f, 10.0f, |
| 2574 | |
| 2575 | // Batch 2, Channel 2 |
| 2576 | 11.0f, 12.0f, |
| 2577 | |
| 2578 | // Batch 3, Channel 0 |
| 2579 | 25.0f, 26.0f, |
| 2580 | |
| 2581 | // Batch 3, Channel 1 |
| 2582 | 27.0f, 28.0f, |
| 2583 | |
| 2584 | // Batch 3, Channel 2 |
| 2585 | 29.0f, 30.0f, |
| 2586 | |
| 2587 | // Batch 4, Channel 0 |
| 2588 | 13.0f, 14.0f, |
| 2589 | |
| 2590 | // Batch 4, Channel 1 |
| 2591 | 15.0f, 16.0f, |
| 2592 | |
| 2593 | // Batch 4, Channel 2 |
| 2594 | 17.0f, 18.0f, |
| 2595 | |
| 2596 | // Batch 5, Channel 0 |
| 2597 | 31.0f, 32.0f, |
| 2598 | |
| 2599 | // Batch 5, Channel 1 |
| 2600 | 33.0f, 34.0f, |
| 2601 | |
| 2602 | // Batch 5, Channel 2 |
| 2603 | 35.0f, 36.0f |
| 2604 | })); |
| 2605 | |
| 2606 | return result; |
| 2607 | } |
| 2608 | |
| 2609 | LayerTestResult<float, 3> Concatenation3dDim0DiffInputDimsTest(armnn::IWorkloadFactory& workloadFactory) |
| 2610 | { |
| 2611 | return Concatenation3dDim0DiffInputDimsTestImpl<float>(workloadFactory, 0.0f, 0); |
| 2612 | } |
| 2613 | |
| 2614 | template <typename T> |
| 2615 | LayerTestResult<T, 3> Concatenation3dDim1DiffInputDimsTestImpl(armnn::IWorkloadFactory& workloadFactory, float qScale, |
| 2616 | int32_t qOffset) |
| 2617 | { |
| 2618 | armnn::TensorInfo input0TensorInfo({ 2, 3, 2 }, armnn::GetDataType<T>()); |
| 2619 | auto input0 = MakeTensor<T, 3>(input0TensorInfo, QuantizedVector<T>(qScale, qOffset, { |
| 2620 | // Batch 0, Channel 0 |
| 2621 | 1.0f, 2.0f, |
| 2622 | |
| 2623 | // Batch 0, Channel 1 |
| 2624 | 3.0f, 4.0f, |
| 2625 | |
| 2626 | // Batch 0, Channel 2 |
| 2627 | 5.0f, 6.0f, |
| 2628 | |
| 2629 | // Batch 1, Channel 0 |
| 2630 | 19.0f, 20.0f, |
| 2631 | |
| 2632 | // Batch 1, Channel 1 |
| 2633 | 21.0f, 22.0f, |
| 2634 | |
| 2635 | // Batch 1, Channel 2 |
| 2636 | 23.0f, 24.0f |
| 2637 | })); |
| 2638 | |
| 2639 | armnn::TensorInfo input1TensorInfo({ 2, 4, 2 }, armnn::GetDataType<T>()); |
| 2640 | auto input1 = MakeTensor<T, 3>(input1TensorInfo, QuantizedVector<T>(qScale, qOffset, { |
| 2641 | // Batch 0, Channel 0 |
| 2642 | 7.0f, 8.0f, |
| 2643 | |
| 2644 | // Batch 0, Channel 1 |
| 2645 | 9.0f, 10.0f, |
| 2646 | |
| 2647 | // Batch 0, Channel 2 |
| 2648 | 11.0f, 12.0f, |
| 2649 | |
| 2650 | // Batch 0, Channel 3 |
| 2651 | 25.0f, 26.0f, |
| 2652 | |
| 2653 | // Batch 1, Channel 0 |
| 2654 | 27.0f, 28.0f, |
| 2655 | |
| 2656 | // Batch 1, Channel 1 |
| 2657 | 29.0f, 30.0f, |
| 2658 | |
| 2659 | // Batch 1, Channel 2 |
| 2660 | 13.0f, 14.0f, |
| 2661 | |
| 2662 | // Batch 1, Channel 3 |
| 2663 | 15.0f, 16.0f, |
| 2664 | })); |
| 2665 | |
| 2666 | armnn::TensorInfo input2TensorInfo({ 2, 1, 2 }, armnn::GetDataType<T>()); |
| 2667 | auto input2 = MakeTensor<T, 3>(input2TensorInfo, QuantizedVector<T>(qScale, qOffset, { |
| 2668 | // Batch 0, Channel 0 |
| 2669 | 17.0f, 18.0f, |
| 2670 | |
| 2671 | // Batch 1, Channel 0 |
| 2672 | 31.0f, 32.0f, |
| 2673 | })); |
| 2674 | |
| 2675 | armnn::TensorInfo outputTensorInfo({ 2, 8, 2 }, armnn::GetDataType<T>()); |
| 2676 | LayerTestResult<T, 3> result(outputTensorInfo); |
| 2677 | |
| 2678 | std::vector<T> output; |
| 2679 | output.resize(outputTensorInfo.GetNumElements()); |
surmeh01 | 3537c2c | 2018-05-18 16:31:43 +0100 | [diff] [blame] | 2680 | Concatenate<T>(workloadFactory, |
telsoa01 | 4fcda01 | 2018-03-09 14:13:49 +0000 | [diff] [blame] | 2681 | { input0TensorInfo, input1TensorInfo, input2TensorInfo }, |
| 2682 | { input0.data(), input1.data(), input2.data() }, |
| 2683 | outputTensorInfo, |
| 2684 | output.data(), |
| 2685 | 1); |
| 2686 | |
| 2687 | result.output = MakeTensor<T, 3>(outputTensorInfo, output); |
| 2688 | result.outputExpected = MakeTensor<T, 3>(outputTensorInfo, QuantizedVector<T>(qScale, qOffset, { |
| 2689 | // Batch 0, Channel 0 |
| 2690 | 1.0f, 2.0f, |
| 2691 | |
| 2692 | // Batch 0, Channel 1 |
| 2693 | 3.0f, 4.0f, |
| 2694 | |
| 2695 | // Batch 0, Channel 2 |
| 2696 | 5.0f, 6.0f, |
| 2697 | |
| 2698 | // Batch 0, Channel 3 |
| 2699 | 7.0f, 8.0f, |
| 2700 | |
| 2701 | // Batch 0, Channel 4 |
| 2702 | 9.0f, 10.0f, |
| 2703 | |
| 2704 | // Batch 0, Channel 5 |
| 2705 | 11.0f, 12.0f, |
| 2706 | |
| 2707 | // Batch 0, Channel 6 |
| 2708 | 25.0f, 26.0f, |
| 2709 | |
| 2710 | // Batch 0, Channel 7 |
| 2711 | 17.0f, 18.0f, |
| 2712 | |
| 2713 | // Batch 1, Channel 0 |
| 2714 | 19.0f, 20.0f, |
| 2715 | |
| 2716 | // Batch 1, Channel 1 |
| 2717 | 21.0f, 22.0f, |
| 2718 | |
| 2719 | // Batch 1, Channel 2 |
| 2720 | 23.0f, 24.0f, |
| 2721 | |
| 2722 | // Batch 1, Channel 3 |
| 2723 | 27.0f, 28.0f, |
| 2724 | |
| 2725 | // Batch 1, Channel 4 |
| 2726 | 29.0f, 30.0f, |
| 2727 | |
| 2728 | // Batch 1, Channel 5 |
| 2729 | 13.0f, 14.0f, |
| 2730 | |
| 2731 | // Batch 1, Channel 6 |
| 2732 | 15.0f, 16.0f, |
| 2733 | |
| 2734 | // Batch 1, Channel 7 |
| 2735 | 31.0f, 32.0f, |
| 2736 | })); |
| 2737 | |
| 2738 | return result; |
| 2739 | } |
| 2740 | |
| 2741 | LayerTestResult<float, 3> Concatenation3dDim1DiffInputDimsTest(armnn::IWorkloadFactory& workloadFactory) |
| 2742 | { |
| 2743 | return Concatenation3dDim1DiffInputDimsTestImpl<float>(workloadFactory, 0.0f, 0); |
| 2744 | } |
| 2745 | |
| 2746 | template <typename T> |
| 2747 | LayerTestResult<T, 3> Concatenation3dDim2DiffInputDimsTestImpl(armnn::IWorkloadFactory& workloadFactory, float qScale, |
| 2748 | int32_t qOffset) |
| 2749 | { |
| 2750 | armnn::TensorInfo input0TensorInfo({ 2, 3, 2 }, armnn::GetDataType<T>()); |
| 2751 | auto input0 = MakeTensor<T, 3>(input0TensorInfo, QuantizedVector<T>(qScale, qOffset, { |
| 2752 | // Batch 0, Channel 0 |
| 2753 | 1.0f, 2.0f, |
| 2754 | |
| 2755 | // Batch 0, Channel 1 |
| 2756 | 3.0f, 4.0f, |
| 2757 | |
| 2758 | // Batch 0, Channel 2 |
| 2759 | 5.0f, 6.0f, |
| 2760 | |
| 2761 | // Batch 1, Channel 0 |
| 2762 | 19.0f, 20.0f, |
| 2763 | |
| 2764 | // Batch 1, Channel 1 |
| 2765 | 21.0f, 22.0f, |
| 2766 | |
| 2767 | // Batch 1, Channel 2 |
| 2768 | 23.0f, 24.0f |
| 2769 | })); |
| 2770 | |
| 2771 | armnn::TensorInfo input1TensorInfo({ 2, 3, 1 }, armnn::GetDataType<T>()); |
| 2772 | auto input1 = MakeTensor<T, 3>(input1TensorInfo, QuantizedVector<T>(qScale, qOffset, { |
| 2773 | // Batch 0, Channel 0 |
| 2774 | 7.0f, |
| 2775 | |
| 2776 | // Batch 0, Channel 1 |
| 2777 | 9.0f, |
| 2778 | |
| 2779 | // Batch 0, Channel 2 |
| 2780 | 11.0f, |
| 2781 | |
| 2782 | // Batch 1, Channel 0 |
| 2783 | 25.0f, |
| 2784 | |
| 2785 | // Batch 1, Channel 1 |
| 2786 | 27.0f, |
| 2787 | |
| 2788 | // Batch 1, Channel 2 |
| 2789 | 29.0f |
| 2790 | })); |
| 2791 | |
| 2792 | armnn::TensorInfo input2TensorInfo({ 2, 3, 3 }, armnn::GetDataType<T>()); |
| 2793 | auto input2 = MakeTensor<T, 3>(input2TensorInfo, QuantizedVector<T>(qScale, qOffset, { |
| 2794 | // Batch 0, Channel 0 |
| 2795 | 13.0f, 14.0f, 50.0f, |
| 2796 | |
| 2797 | // Batch 0, Channel 1 |
| 2798 | 15.0f, 16.0f, 51.0f, |
| 2799 | |
| 2800 | // Batch 0, Channel 2 |
| 2801 | 17.0f, 18.0f, 52.0f, |
| 2802 | |
| 2803 | // Batch 1, Channel 0 |
| 2804 | 31.0f, 32.0f, 53.0f, |
| 2805 | |
| 2806 | // Batch 1, Channel 1 |
| 2807 | 33.0f, 34.0f, 54.0f, |
| 2808 | |
| 2809 | // Batch 1, Channel 2 |
| 2810 | 35.0f, 36.0f, 55.0f, |
| 2811 | })); |
| 2812 | |
| 2813 | armnn::TensorInfo outputTensorInfo({ 2, 3, 6 }, armnn::GetDataType<T>()); |
| 2814 | LayerTestResult<T, 3> result(outputTensorInfo); |
| 2815 | |
| 2816 | std::vector<T> output; |
| 2817 | output.resize(outputTensorInfo.GetNumElements()); |
surmeh01 | 3537c2c | 2018-05-18 16:31:43 +0100 | [diff] [blame] | 2818 | Concatenate<T>(workloadFactory, |
telsoa01 | 4fcda01 | 2018-03-09 14:13:49 +0000 | [diff] [blame] | 2819 | { input0TensorInfo, input1TensorInfo, input2TensorInfo }, |
| 2820 | { input0.data(), input1.data(), input2.data() }, |
| 2821 | outputTensorInfo, |
| 2822 | output.data(), |
| 2823 | 2); |
| 2824 | |
| 2825 | result.output = MakeTensor<T, 3>(outputTensorInfo, output); |
| 2826 | result.outputExpected = MakeTensor<T, 3>(outputTensorInfo, QuantizedVector<T>(qScale, qOffset, { |
| 2827 | // Batch 0, Channel 0 |
| 2828 | 1.0f, 2.0f, 7.0f, 13.0f, 14.0f, 50.0f, |
| 2829 | |
| 2830 | // Batch 0, Channel 1 |
| 2831 | 3.0f, 4.0f, 9.0f, 15.0f, 16.0f, 51.0f, |
| 2832 | |
| 2833 | // Batch 0, Channel 2 |
| 2834 | 5.0f, 6.0f, 11.0f, 17.0f, 18.0f, 52.0f, |
| 2835 | |
| 2836 | // Batch 1, Channel 0 |
| 2837 | 19.0f, 20.0f, 25.0f, 31.0f, 32.0f, 53.0f, |
| 2838 | |
| 2839 | // Batch 1, Channel 1 |
| 2840 | 21.0f, 22.0f, 27.0f, 33.0f, 34.0f, 54.0f, |
| 2841 | |
| 2842 | // Batch 1, Channel 2 |
| 2843 | 23.0f, 24.0f, 29.0f, 35.0f, 36.0f, 55.0f, |
| 2844 | })); |
| 2845 | |
| 2846 | return result; |
| 2847 | } |
| 2848 | |
| 2849 | LayerTestResult<float, 3> Concatenation3dDim2DiffInputDimsTest(armnn::IWorkloadFactory& workloadFactory) |
| 2850 | { |
| 2851 | return Concatenation3dDim2DiffInputDimsTestImpl<float>(workloadFactory, 0.0f, 0); |
| 2852 | } |
| 2853 | |
| 2854 | LayerTestResult<float, 4> ResizeBilinearNopTest(armnn::IWorkloadFactory& workloadFactory) |
| 2855 | { |
| 2856 | constexpr unsigned int inputWidth = 4; |
| 2857 | constexpr unsigned int inputHeight = 4; |
| 2858 | constexpr unsigned int inputChannels = 1; |
| 2859 | constexpr unsigned int inputBatchSize = 1; |
| 2860 | |
| 2861 | constexpr unsigned int outputWidth = inputWidth; |
| 2862 | constexpr unsigned int outputHeight = inputHeight; |
| 2863 | constexpr unsigned int outputChannels = inputChannels; |
| 2864 | constexpr unsigned int outputBatchSize = inputBatchSize; |
| 2865 | |
| 2866 | const armnn::TensorInfo inputTensorInfo({ inputBatchSize, inputChannels, inputHeight, inputWidth }, |
| 2867 | armnn::DataType::Float32); |
| 2868 | const armnn::TensorInfo outputTensorInfo({ outputBatchSize, outputChannels, outputHeight, outputWidth }, |
| 2869 | armnn::DataType::Float32); |
| 2870 | |
| 2871 | auto input = MakeTensor<float, 4>(inputTensorInfo, std::vector<float>({ |
| 2872 | 1.0f, 2.0f, 3.0f, 4.0f, |
| 2873 | 2.0f, 3.0f, 4.0f, 5.0f, |
| 2874 | 3.0f, 4.0f, 5.0f, 6.0f, |
| 2875 | 4.0f, 5.0f, 6.0f, 7.0f |
| 2876 | })); |
| 2877 | |
| 2878 | LayerTestResult<float, 4> result(outputTensorInfo); |
| 2879 | result.outputExpected = input; |
| 2880 | |
| 2881 | std::unique_ptr<armnn::ITensorHandle> inputHandle = workloadFactory.CreateTensorHandle(inputTensorInfo); |
| 2882 | std::unique_ptr<armnn::ITensorHandle> outputHandle = workloadFactory.CreateTensorHandle(outputTensorInfo); |
| 2883 | |
| 2884 | armnn::ResizeBilinearQueueDescriptor descriptor; |
| 2885 | armnn::WorkloadInfo info; |
| 2886 | AddInputToWorkload(descriptor, info, inputTensorInfo, inputHandle.get()); |
| 2887 | AddOutputToWorkload(descriptor, info, outputTensorInfo, outputHandle.get()); |
| 2888 | |
| 2889 | std::unique_ptr<armnn::IWorkload> workload = workloadFactory.CreateResizeBilinear(descriptor, info); |
| 2890 | |
| 2891 | inputHandle->Allocate(); |
| 2892 | outputHandle->Allocate(); |
| 2893 | CopyDataToITensorHandle(inputHandle.get(), &input[0][0][0][0]); |
| 2894 | |
surmeh01 | 3537c2c | 2018-05-18 16:31:43 +0100 | [diff] [blame] | 2895 | workloadFactory.Finalize(); |
telsoa01 | 4fcda01 | 2018-03-09 14:13:49 +0000 | [diff] [blame] | 2896 | workload->Execute(); |
| 2897 | |
| 2898 | CopyDataFromITensorHandle(&result.output[0][0][0][0], outputHandle.get()); |
| 2899 | return result; |
| 2900 | } |
| 2901 | |
| 2902 | LayerTestResult<float, 4> SimpleResizeBilinearTest(armnn::IWorkloadFactory& workloadFactory) |
| 2903 | { |
| 2904 | constexpr unsigned int inputWidth = 2; |
| 2905 | constexpr unsigned int inputHeight = 2; |
| 2906 | constexpr unsigned int inputChannels = 1; |
| 2907 | constexpr unsigned int inputBatchSize = 1; |
| 2908 | |
| 2909 | constexpr unsigned int outputWidth = inputWidth / 2; |
| 2910 | constexpr unsigned int outputHeight = inputHeight / 2; |
| 2911 | constexpr unsigned int outputChannels = inputChannels; |
| 2912 | constexpr unsigned int outputBatchSize = inputBatchSize; |
| 2913 | |
| 2914 | const armnn::TensorInfo inputTensorInfo({ inputBatchSize, inputChannels, inputHeight, inputWidth }, |
| 2915 | armnn::DataType::Float32); |
| 2916 | const armnn::TensorInfo outputTensorInfo({ outputBatchSize, outputChannels, outputHeight, outputWidth }, |
| 2917 | armnn::DataType::Float32); |
| 2918 | |
| 2919 | auto input = MakeTensor<float, 4>(inputTensorInfo, std::vector<float>({ |
| 2920 | 1.0f, 255.0f, |
| 2921 | 200.0f, 250.f, |
| 2922 | })); |
| 2923 | |
| 2924 | // The 'resize bilinear' operation projects the top-left corner of output texels into the input image, |
| 2925 | // then figures out the interpolants and weights. Note this is different to projecting the centre of the |
telsoa01 | c577f2c | 2018-08-31 09:22:23 +0100 | [diff] [blame] | 2926 | // output texel - and thus we'll expect the output 1x1 matrix to contain, as its single element, the value |
telsoa01 | 4fcda01 | 2018-03-09 14:13:49 +0000 | [diff] [blame] | 2927 | // that was at position (0,0) of the input matrix (rather than an average, which we would expect if projecting |
| 2928 | // the centre). |
| 2929 | LayerTestResult<float, 4> result(outputTensorInfo); |
| 2930 | result.outputExpected = MakeTensor<float, 4>(outputTensorInfo, std::vector<float>({ |
| 2931 | 1.0f |
| 2932 | })); |
| 2933 | |
| 2934 | std::unique_ptr<armnn::ITensorHandle> inputHandle = workloadFactory.CreateTensorHandle(inputTensorInfo); |
| 2935 | std::unique_ptr<armnn::ITensorHandle> outputHandle = workloadFactory.CreateTensorHandle(outputTensorInfo); |
| 2936 | |
| 2937 | armnn::ResizeBilinearQueueDescriptor descriptor; |
| 2938 | armnn::WorkloadInfo info; |
| 2939 | AddInputToWorkload(descriptor, info, inputTensorInfo, inputHandle.get()); |
| 2940 | AddOutputToWorkload(descriptor, info, outputTensorInfo, outputHandle.get()); |
| 2941 | |
| 2942 | std::unique_ptr<armnn::IWorkload> workload = workloadFactory.CreateResizeBilinear(descriptor, info); |
| 2943 | |
| 2944 | inputHandle->Allocate(); |
| 2945 | outputHandle->Allocate(); |
| 2946 | CopyDataToITensorHandle(inputHandle.get(), &input[0][0][0][0]); |
| 2947 | |
surmeh01 | 3537c2c | 2018-05-18 16:31:43 +0100 | [diff] [blame] | 2948 | workloadFactory.Finalize(); |
telsoa01 | 4fcda01 | 2018-03-09 14:13:49 +0000 | [diff] [blame] | 2949 | workload->Execute(); |
| 2950 | |
| 2951 | CopyDataFromITensorHandle(&result.output[0][0][0][0], outputHandle.get()); |
| 2952 | return result; |
| 2953 | } |
| 2954 | |
| 2955 | LayerTestResult<float, 4> ResizeBilinearSqMinTest(armnn::IWorkloadFactory& workloadFactory) |
| 2956 | { |
| 2957 | constexpr unsigned int inputWidth = 4; |
| 2958 | constexpr unsigned int inputHeight = 4; |
| 2959 | constexpr unsigned int inputChannels = 1; |
| 2960 | constexpr unsigned int inputBatchSize = 1; |
| 2961 | |
| 2962 | constexpr unsigned int outputWidth = inputWidth / 2; |
| 2963 | constexpr unsigned int outputHeight = inputHeight / 2; |
| 2964 | constexpr unsigned int outputChannels = inputChannels; |
| 2965 | constexpr unsigned int outputBatchSize = inputBatchSize; |
| 2966 | |
| 2967 | const armnn::TensorInfo inputTensorInfo({ inputBatchSize, inputChannels, inputHeight, inputWidth }, |
| 2968 | armnn::DataType::Float32); |
| 2969 | const armnn::TensorInfo outputTensorInfo({ outputBatchSize, outputChannels, outputHeight, outputWidth }, |
| 2970 | armnn::DataType::Float32); |
| 2971 | |
| 2972 | auto input = MakeTensor<float, 4>(inputTensorInfo, std::vector<float>({ |
| 2973 | 1.0f, 2.0f, 3.0f, 4.0f, |
| 2974 | 2.0f, 3.0f, 4.0f, 5.0f, |
| 2975 | 3.0f, 4.0f, 5.0f, 6.0f, |
| 2976 | 4.0f, 5.0f, 6.0f, 7.0f |
| 2977 | })); |
| 2978 | |
| 2979 | LayerTestResult<float, 4> result(outputTensorInfo); |
| 2980 | result.outputExpected = MakeTensor<float, 4>(outputTensorInfo, std::vector<float>({ |
| 2981 | 1.f, 3.f, |
| 2982 | 3.f, 5.f |
| 2983 | })); |
| 2984 | |
| 2985 | std::unique_ptr<armnn::ITensorHandle> inputHandle = workloadFactory.CreateTensorHandle(inputTensorInfo); |
| 2986 | std::unique_ptr<armnn::ITensorHandle> outputHandle = workloadFactory.CreateTensorHandle(outputTensorInfo); |
| 2987 | |
| 2988 | armnn::ResizeBilinearQueueDescriptor descriptor; |
| 2989 | armnn::WorkloadInfo info; |
| 2990 | AddInputToWorkload(descriptor, info, inputTensorInfo, inputHandle.get()); |
| 2991 | AddOutputToWorkload(descriptor, info, outputTensorInfo, outputHandle.get()); |
| 2992 | |
| 2993 | std::unique_ptr<armnn::IWorkload> workload = workloadFactory.CreateResizeBilinear(descriptor, info); |
| 2994 | |
| 2995 | inputHandle->Allocate(); |
| 2996 | outputHandle->Allocate(); |
| 2997 | CopyDataToITensorHandle(inputHandle.get(), &input[0][0][0][0]); |
| 2998 | |
surmeh01 | 3537c2c | 2018-05-18 16:31:43 +0100 | [diff] [blame] | 2999 | workloadFactory.Finalize(); |
telsoa01 | 4fcda01 | 2018-03-09 14:13:49 +0000 | [diff] [blame] | 3000 | workload->Execute(); |
| 3001 | |
| 3002 | CopyDataFromITensorHandle(&result.output[0][0][0][0], outputHandle.get()); |
| 3003 | return result; |
| 3004 | } |
| 3005 | |
| 3006 | LayerTestResult<float, 4> ResizeBilinearMinTest(armnn::IWorkloadFactory& workloadFactory) |
| 3007 | { |
| 3008 | constexpr unsigned int inputWidth = 5; |
| 3009 | constexpr unsigned int inputHeight = 3; |
| 3010 | constexpr unsigned int inputChannels = 1; |
| 3011 | constexpr unsigned int inputBatchSize = 1; |
| 3012 | |
| 3013 | constexpr unsigned int outputWidth = 3; |
| 3014 | constexpr unsigned int outputHeight = 2; |
| 3015 | constexpr unsigned int outputChannels = inputChannels; |
| 3016 | constexpr unsigned int outputBatchSize = inputBatchSize; |
| 3017 | |
| 3018 | const armnn::TensorInfo inputTensorInfo({ inputBatchSize, inputChannels, inputHeight, inputWidth }, |
| 3019 | armnn::DataType::Float32); |
| 3020 | const armnn::TensorInfo outputTensorInfo({ outputBatchSize, outputChannels, outputHeight, outputWidth }, |
| 3021 | armnn::DataType::Float32); |
| 3022 | |
| 3023 | auto input = MakeTensor<float, 4>(inputTensorInfo, std::vector<float>({ |
| 3024 | 1.0f, 2.0f, 3.0f, 5.0f, 8.0f, |
| 3025 | 13.0f, 21.0f, 34.0f, 55.0f, 89.0f, |
| 3026 | 144.0f, 233.0f, 377.0f, 610.0f, 987.0f |
| 3027 | })); |
| 3028 | |
| 3029 | LayerTestResult<float, 4> result(outputTensorInfo); |
| 3030 | result.outputExpected = MakeTensor<float, 4>(outputTensorInfo, std::vector<float>({ |
| 3031 | 1.0f, 2.6666f, 6.0f, |
| 3032 | 78.5f, 179.3333f, 401.f |
| 3033 | })); |
| 3034 | |
| 3035 | std::unique_ptr<armnn::ITensorHandle> inputHandle = workloadFactory.CreateTensorHandle(inputTensorInfo); |
| 3036 | std::unique_ptr<armnn::ITensorHandle> outputHandle = workloadFactory.CreateTensorHandle(outputTensorInfo); |
| 3037 | |
| 3038 | armnn::ResizeBilinearQueueDescriptor descriptor; |
| 3039 | armnn::WorkloadInfo info; |
| 3040 | AddInputToWorkload(descriptor, info, inputTensorInfo, inputHandle.get()); |
| 3041 | AddOutputToWorkload(descriptor, info, outputTensorInfo, outputHandle.get()); |
| 3042 | |
| 3043 | std::unique_ptr<armnn::IWorkload> workload = workloadFactory.CreateResizeBilinear(descriptor, info); |
| 3044 | |
| 3045 | inputHandle->Allocate(); |
| 3046 | outputHandle->Allocate(); |
| 3047 | CopyDataToITensorHandle(inputHandle.get(), &input[0][0][0][0]); |
| 3048 | |
surmeh01 | 3537c2c | 2018-05-18 16:31:43 +0100 | [diff] [blame] | 3049 | workloadFactory.Finalize(); |
telsoa01 | 4fcda01 | 2018-03-09 14:13:49 +0000 | [diff] [blame] | 3050 | workload->Execute(); |
| 3051 | |
| 3052 | CopyDataFromITensorHandle(&result.output[0][0][0][0], outputHandle.get()); |
| 3053 | return result; |
| 3054 | } |
| 3055 | |
| 3056 | LayerTestResult<float, 4> ResizeBilinearMagTest(armnn::IWorkloadFactory& workloadFactory) |
| 3057 | { |
| 3058 | constexpr unsigned int inputWidth = 2; |
| 3059 | constexpr unsigned int inputHeight = 3; |
| 3060 | constexpr unsigned int inputChannels = 1; |
| 3061 | constexpr unsigned int inputBatchSize = 1; |
| 3062 | |
| 3063 | constexpr unsigned int outputWidth = 5; |
| 3064 | constexpr unsigned int outputHeight = 3; |
| 3065 | constexpr unsigned int outputChannels = inputChannels; |
| 3066 | constexpr unsigned int outputBatchSize = inputBatchSize; |
| 3067 | |
| 3068 | const armnn::TensorInfo inputTensorInfo({ inputBatchSize, inputChannels, inputHeight, inputWidth }, |
| 3069 | armnn::DataType::Float32); |
| 3070 | const armnn::TensorInfo outputTensorInfo({ outputBatchSize, outputChannels, outputHeight, outputWidth }, |
| 3071 | armnn::DataType::Float32); |
| 3072 | |
| 3073 | auto input = MakeTensor<float, 4>(inputTensorInfo, std::vector<float>({ |
| 3074 | 1.0f, 2.0f, |
| 3075 | 13.0f, 21.0f, |
| 3076 | 144.0f, 233.0f |
| 3077 | })); |
| 3078 | |
| 3079 | LayerTestResult<float, 4> result(outputTensorInfo); |
| 3080 | result.outputExpected = MakeTensor<float, 4>(outputTensorInfo, std::vector<float>({ |
| 3081 | 1.0f, 1.4f, 1.8f, 2.f, 2.f, |
| 3082 | 13.f, 16.2f, 19.4f, 21.f, 21.f, |
| 3083 | 144.f, 179.6f, 215.2f, 233.f, 233.f |
| 3084 | })); |
| 3085 | |
| 3086 | std::unique_ptr<armnn::ITensorHandle> inputHandle = workloadFactory.CreateTensorHandle(inputTensorInfo); |
| 3087 | std::unique_ptr<armnn::ITensorHandle> outputHandle = workloadFactory.CreateTensorHandle(outputTensorInfo); |
| 3088 | |
| 3089 | armnn::ResizeBilinearQueueDescriptor descriptor; |
| 3090 | armnn::WorkloadInfo info; |
| 3091 | AddInputToWorkload(descriptor, info, inputTensorInfo, inputHandle.get()); |
| 3092 | AddOutputToWorkload(descriptor, info, outputTensorInfo, outputHandle.get()); |
| 3093 | |
| 3094 | std::unique_ptr<armnn::IWorkload> workload = workloadFactory.CreateResizeBilinear(descriptor, info); |
| 3095 | |
| 3096 | inputHandle->Allocate(); |
| 3097 | outputHandle->Allocate(); |
| 3098 | CopyDataToITensorHandle(inputHandle.get(), &input[0][0][0][0]); |
| 3099 | |
surmeh01 | 3537c2c | 2018-05-18 16:31:43 +0100 | [diff] [blame] | 3100 | workloadFactory.Finalize(); |
telsoa01 | 4fcda01 | 2018-03-09 14:13:49 +0000 | [diff] [blame] | 3101 | workload->Execute(); |
| 3102 | |
| 3103 | CopyDataFromITensorHandle(&result.output[0][0][0][0], outputHandle.get()); |
| 3104 | return result; |
| 3105 | } |
| 3106 | |
| 3107 | LayerTestResult<float, 2> FakeQuantizationTest(armnn::IWorkloadFactory& workloadFactory) |
| 3108 | { |
| 3109 | constexpr unsigned int width = 2; |
| 3110 | constexpr unsigned int height = 3; |
| 3111 | |
| 3112 | const armnn::TensorInfo tensorInfo({height, width }, |
| 3113 | armnn::DataType::Float32); |
| 3114 | auto input = MakeTensor<float, 2>(tensorInfo, std::vector<float>({ |
| 3115 | -10.0f, -5.0f, |
| 3116 | 0.0f, 5.0f, |
| 3117 | 10.0f, 10.0f |
| 3118 | })); |
| 3119 | |
| 3120 | LayerTestResult<float, 2> ret(tensorInfo); |
| 3121 | |
| 3122 | std::unique_ptr<armnn::ITensorHandle> inputHandle = workloadFactory.CreateTensorHandle(tensorInfo); |
| 3123 | |
| 3124 | std::unique_ptr<armnn::ITensorHandle> outputHandle = workloadFactory.CreateTensorHandle(tensorInfo); |
| 3125 | |
| 3126 | armnn::FakeQuantizationQueueDescriptor data; |
| 3127 | armnn::WorkloadInfo info; |
| 3128 | |
| 3129 | AddInputToWorkload(data, info, tensorInfo, inputHandle.get()); |
| 3130 | AddOutputToWorkload(data, info, tensorInfo, outputHandle.get()); |
| 3131 | float min = -10.f; |
| 3132 | float max = 10.f; |
| 3133 | |
| 3134 | data.m_Parameters.m_Min = min; |
| 3135 | data.m_Parameters.m_Max = max; |
| 3136 | |
| 3137 | armnn::PassthroughCpuTensorHandle refHandle(tensorInfo, &ret.outputExpected[0][0]); |
| 3138 | armnn::FakeQuantizationQueueDescriptor refData = data; |
| 3139 | armnn::WorkloadInfo refInfo = info; |
| 3140 | SetWorkloadOutput(refData, refInfo, 0, tensorInfo, &refHandle); |
| 3141 | |
| 3142 | std::unique_ptr<armnn::IWorkload> workload = workloadFactory.CreateFakeQuantization(data, info); |
| 3143 | |
| 3144 | inputHandle->Allocate(); |
| 3145 | outputHandle->Allocate(); |
| 3146 | |
| 3147 | CopyDataToITensorHandle(inputHandle.get(), &input[0][0]); |
| 3148 | |
surmeh01 | 3537c2c | 2018-05-18 16:31:43 +0100 | [diff] [blame] | 3149 | workloadFactory.Finalize(); |
telsoa01 | 4fcda01 | 2018-03-09 14:13:49 +0000 | [diff] [blame] | 3150 | workload->Execute(); |
| 3151 | |
| 3152 | CopyDataFromITensorHandle(&ret.output[0][0], outputHandle.get()); |
| 3153 | |
| 3154 | ret.outputExpected = MakeTensor<float, 2>(tensorInfo, std::vector<float>({ |
| 3155 | 0.0f, 63.0f, |
| 3156 | 128.0f, 191.0f, |
| 3157 | 255.0f, 255.0f |
| 3158 | })); |
| 3159 | return ret; |
| 3160 | } |
| 3161 | |
| 3162 | LayerTestResult<float, 4> L2Normalization1dTest(armnn::IWorkloadFactory& workloadFactory) |
| 3163 | { |
| 3164 | constexpr unsigned int inputWidth = 1; |
| 3165 | constexpr unsigned int inputHeight = 1; |
| 3166 | constexpr unsigned int inputChannels = 10; |
| 3167 | constexpr unsigned int inputBatchSize = 1; |
| 3168 | |
| 3169 | constexpr unsigned int outputWidth = inputWidth; |
| 3170 | constexpr unsigned int outputHeight = inputHeight; |
| 3171 | constexpr unsigned int outputChannels = inputChannels; |
| 3172 | constexpr unsigned int outputBatchSize = inputBatchSize; |
| 3173 | |
| 3174 | const armnn::TensorInfo inputTensorInfo({ inputBatchSize, inputChannels, inputHeight, inputWidth }, |
| 3175 | armnn::DataType::Float32); |
| 3176 | const armnn::TensorInfo outputTensorInfo({ outputBatchSize, outputChannels, outputHeight, outputWidth }, |
| 3177 | armnn::DataType::Float32); |
| 3178 | |
| 3179 | auto input = MakeTensor<float, 4>(inputTensorInfo, std::vector<float>({ |
| 3180 | 1.0f, 2.0f, 3.0f, 4.0f, 5.0f, 6.0f, 7.0f, 8.0f, 9.0f, 10.0f |
| 3181 | })); |
| 3182 | |
| 3183 | const float approxInvL2Norm = 0.050964719f; |
| 3184 | LayerTestResult<float, 4> result(outputTensorInfo); |
| 3185 | result.outputExpected = MakeTensor<float, 4>(inputTensorInfo, std::vector<float>({ |
| 3186 | 1.0f * approxInvL2Norm, |
| 3187 | 2.0f * approxInvL2Norm, |
| 3188 | 3.0f * approxInvL2Norm, |
| 3189 | 4.0f * approxInvL2Norm, |
| 3190 | 5.0f * approxInvL2Norm, |
| 3191 | 6.0f * approxInvL2Norm, |
| 3192 | 7.0f * approxInvL2Norm, |
| 3193 | 8.0f * approxInvL2Norm, |
| 3194 | 9.0f * approxInvL2Norm, |
| 3195 | 10.0f * approxInvL2Norm |
| 3196 | })); |
| 3197 | |
| 3198 | std::unique_ptr<armnn::ITensorHandle> inputHandle = workloadFactory.CreateTensorHandle(inputTensorInfo); |
| 3199 | std::unique_ptr<armnn::ITensorHandle> outputHandle = workloadFactory.CreateTensorHandle(outputTensorInfo); |
| 3200 | |
| 3201 | armnn::L2NormalizationQueueDescriptor descriptor; |
| 3202 | armnn::WorkloadInfo info; |
| 3203 | AddInputToWorkload(descriptor, info, inputTensorInfo, inputHandle.get()); |
| 3204 | AddOutputToWorkload(descriptor, info, outputTensorInfo, outputHandle.get()); |
| 3205 | |
| 3206 | std::unique_ptr<armnn::IWorkload> workload = workloadFactory.CreateL2Normalization(descriptor, info); |
| 3207 | |
| 3208 | inputHandle->Allocate(); |
| 3209 | outputHandle->Allocate(); |
| 3210 | CopyDataToITensorHandle(inputHandle.get(), &input[0][0][0][0]); |
| 3211 | |
surmeh01 | 3537c2c | 2018-05-18 16:31:43 +0100 | [diff] [blame] | 3212 | workloadFactory.Finalize(); |
telsoa01 | 4fcda01 | 2018-03-09 14:13:49 +0000 | [diff] [blame] | 3213 | workload->Execute(); |
| 3214 | |
| 3215 | CopyDataFromITensorHandle(&result.output[0][0][0][0], outputHandle.get()); |
| 3216 | return result; |
| 3217 | } |
| 3218 | |
| 3219 | namespace |
| 3220 | { |
| 3221 | |
| 3222 | float CalcInvL2Norm(std::initializer_list<float> elements) |
| 3223 | { |
| 3224 | const float reduction = std::accumulate(elements.begin(), elements.end(), 0.0f, |
| 3225 | [](float acc, float element) { return acc + element * element; }); |
| 3226 | return 1.0f / sqrtf(reduction); |
| 3227 | } |
| 3228 | |
| 3229 | } |
| 3230 | |
| 3231 | LayerTestResult<float, 4> L2Normalization2dTest(armnn::IWorkloadFactory& workloadFactory) |
| 3232 | { |
| 3233 | constexpr unsigned int inputWidth = 5; |
| 3234 | constexpr unsigned int inputHeight = 1; |
| 3235 | constexpr unsigned int inputChannels = 2; |
| 3236 | constexpr unsigned int inputBatchSize = 1; |
| 3237 | |
| 3238 | constexpr unsigned int outputWidth = inputWidth; |
| 3239 | constexpr unsigned int outputHeight = inputHeight; |
| 3240 | constexpr unsigned int outputChannels = inputChannels; |
| 3241 | constexpr unsigned int outputBatchSize = inputBatchSize; |
| 3242 | |
| 3243 | const armnn::TensorInfo inputTensorInfo({ inputBatchSize, inputChannels, inputHeight, inputWidth }, |
| 3244 | armnn::DataType::Float32); |
| 3245 | const armnn::TensorInfo outputTensorInfo({ outputBatchSize, outputChannels, outputHeight, outputWidth }, |
| 3246 | armnn::DataType::Float32); |
| 3247 | |
| 3248 | auto input = MakeTensor<float, 4>(inputTensorInfo, std::vector<float>({ |
| 3249 | 1.0f, 3.0f, 5.0f, 7.0f, 9.0f, |
| 3250 | 2.0f, 4.0f, 6.0f, 8.0f, 10.0f |
| 3251 | })); |
| 3252 | |
| 3253 | LayerTestResult<float, 4> result(outputTensorInfo); |
| 3254 | result.outputExpected = MakeTensor<float, 4>(inputTensorInfo, std::vector<float>({ |
| 3255 | 1.0f * CalcInvL2Norm({ 1.0f, 2.0f }), |
| 3256 | 3.0f * CalcInvL2Norm({ 3.0f, 4.0f }), |
| 3257 | 5.0f * CalcInvL2Norm({ 5.0f, 6.0f }), |
| 3258 | 7.0f * CalcInvL2Norm({ 7.0f, 8.0f }), |
| 3259 | 9.0f * CalcInvL2Norm({ 9.0f, 10.0f }), |
| 3260 | |
| 3261 | 2.0f * CalcInvL2Norm({ 1.0f, 2.0f }), |
| 3262 | 4.0f * CalcInvL2Norm({ 3.0f, 4.0f }), |
| 3263 | 6.0f * CalcInvL2Norm({ 5.0f, 6.0f }), |
| 3264 | 8.0f * CalcInvL2Norm({ 7.0f, 8.0f }), |
| 3265 | 10.0f * CalcInvL2Norm({ 9.0f, 10.0f }) |
| 3266 | })); |
| 3267 | |
| 3268 | std::unique_ptr<armnn::ITensorHandle> inputHandle = workloadFactory.CreateTensorHandle(inputTensorInfo); |
| 3269 | std::unique_ptr<armnn::ITensorHandle> outputHandle = workloadFactory.CreateTensorHandle(outputTensorInfo); |
| 3270 | |
| 3271 | armnn::L2NormalizationQueueDescriptor descriptor; |
| 3272 | armnn::WorkloadInfo info; |
| 3273 | AddInputToWorkload(descriptor, info, inputTensorInfo, inputHandle.get()); |
| 3274 | AddOutputToWorkload(descriptor, info, outputTensorInfo, outputHandle.get()); |
| 3275 | |
| 3276 | std::unique_ptr<armnn::IWorkload> workload = workloadFactory.CreateL2Normalization(descriptor, info); |
| 3277 | |
| 3278 | inputHandle->Allocate(); |
| 3279 | outputHandle->Allocate(); |
| 3280 | CopyDataToITensorHandle(inputHandle.get(), &input[0][0][0][0]); |
| 3281 | |
surmeh01 | 3537c2c | 2018-05-18 16:31:43 +0100 | [diff] [blame] | 3282 | workloadFactory.Finalize(); |
telsoa01 | 4fcda01 | 2018-03-09 14:13:49 +0000 | [diff] [blame] | 3283 | workload->Execute(); |
| 3284 | |
| 3285 | CopyDataFromITensorHandle(&result.output[0][0][0][0], outputHandle.get()); |
| 3286 | return result; |
| 3287 | } |
| 3288 | |
| 3289 | LayerTestResult<float, 4> L2Normalization3dTest(armnn::IWorkloadFactory& workloadFactory) |
| 3290 | { |
| 3291 | constexpr unsigned int inputWidth = 3; |
| 3292 | constexpr unsigned int inputHeight = 4; |
| 3293 | constexpr unsigned int inputChannels = 2; |
| 3294 | constexpr unsigned int inputBatchSize = 1; |
| 3295 | |
| 3296 | constexpr unsigned int outputWidth = inputWidth; |
| 3297 | constexpr unsigned int outputHeight = inputHeight; |
| 3298 | constexpr unsigned int outputChannels = inputChannels; |
| 3299 | constexpr unsigned int outputBatchSize = inputBatchSize; |
| 3300 | |
| 3301 | const armnn::TensorInfo inputTensorInfo({ inputBatchSize, inputChannels, inputHeight, inputWidth }, |
| 3302 | armnn::DataType::Float32); |
| 3303 | const armnn::TensorInfo outputTensorInfo({ outputBatchSize, outputChannels, outputHeight, outputWidth }, |
| 3304 | armnn::DataType::Float32); |
| 3305 | |
| 3306 | auto input = MakeTensor<float, 4>(inputTensorInfo, std::vector<float>({ |
| 3307 | // Channel 0 |
| 3308 | 119.0f, 21.0f, 150.0f, |
| 3309 | 149.0f, 32.0f, 179.0f, |
| 3310 | 15.0f, 227.0f, 141.0f, |
| 3311 | 147.0f, 199.0f, 220.0f, |
| 3312 | |
| 3313 | // Channel 1 |
| 3314 | 110.0f, 140.0f, 73.0f, |
| 3315 | 211.0f, 212.0f, 89.0f, |
| 3316 | 24.0f, 138.0f, 188.0f, |
| 3317 | 162.0f, 12.0f, 161.0f, |
| 3318 | })); |
| 3319 | |
| 3320 | LayerTestResult<float, 4> result(outputTensorInfo); |
| 3321 | result.outputExpected = MakeTensor<float, 4>(inputTensorInfo, std::vector<float>({ |
| 3322 | 119.0f * CalcInvL2Norm({ 119.0f, 110.0f }), |
| 3323 | 21.0f * CalcInvL2Norm({ 21.0f, 140.0f }), |
| 3324 | 150.0f * CalcInvL2Norm({ 150.0f, 73.0f }), |
| 3325 | 149.0f * CalcInvL2Norm({ 149.0f, 211.0f }), |
| 3326 | 32.0f * CalcInvL2Norm({ 32.0f, 212.0f }), |
| 3327 | 179.0f * CalcInvL2Norm({ 179.0f, 89.0f }), |
| 3328 | 15.0f * CalcInvL2Norm({ 15.0f, 24.0f }), |
| 3329 | 227.0f * CalcInvL2Norm({ 227.0f, 138.0f }), |
| 3330 | 141.0f * CalcInvL2Norm({ 141.0f, 188.0f }), |
| 3331 | 147.0f * CalcInvL2Norm({ 147.0f, 162.0f }), |
| 3332 | 199.0f * CalcInvL2Norm({ 199.0f, 12.0f }), |
| 3333 | 220.0f * CalcInvL2Norm({ 220.0f, 161.0f }), |
| 3334 | |
| 3335 | 110.0f * CalcInvL2Norm({ 119.0f, 110.0f }), |
| 3336 | 140.0f * CalcInvL2Norm({ 21.0f, 140.0f }), |
| 3337 | 73.0f * CalcInvL2Norm({ 150.0f, 73.0f }), |
| 3338 | 211.0f * CalcInvL2Norm({ 149.0f, 211.0f }), |
| 3339 | 212.0f * CalcInvL2Norm({ 32.0f, 212.0f }), |
| 3340 | 89.0f * CalcInvL2Norm({ 179.0f, 89.0f }), |
| 3341 | 24.0f * CalcInvL2Norm({ 15.0f, 24.0f }), |
| 3342 | 138.0f * CalcInvL2Norm({ 227.0f, 138.0f }), |
| 3343 | 188.0f * CalcInvL2Norm({ 141.0f, 188.0f }), |
| 3344 | 162.0f * CalcInvL2Norm({ 147.0f, 162.0f }), |
| 3345 | 12.0f * CalcInvL2Norm({ 199.0f, 12.0f }), |
| 3346 | 161.0f * CalcInvL2Norm({ 220.0f, 161.0f }), |
| 3347 | })); |
| 3348 | |
| 3349 | std::unique_ptr<armnn::ITensorHandle> inputHandle = workloadFactory.CreateTensorHandle(inputTensorInfo); |
| 3350 | std::unique_ptr<armnn::ITensorHandle> outputHandle = workloadFactory.CreateTensorHandle(outputTensorInfo); |
| 3351 | |
| 3352 | armnn::L2NormalizationQueueDescriptor descriptor; |
| 3353 | armnn::WorkloadInfo info; |
| 3354 | AddInputToWorkload(descriptor, info, inputTensorInfo, inputHandle.get()); |
| 3355 | AddOutputToWorkload(descriptor, info, outputTensorInfo, outputHandle.get()); |
| 3356 | |
| 3357 | std::unique_ptr<armnn::IWorkload> workload = workloadFactory.CreateL2Normalization(descriptor, info); |
| 3358 | |
| 3359 | inputHandle->Allocate(); |
| 3360 | outputHandle->Allocate(); |
| 3361 | CopyDataToITensorHandle(inputHandle.get(), &input[0][0][0][0]); |
| 3362 | |
surmeh01 | 3537c2c | 2018-05-18 16:31:43 +0100 | [diff] [blame] | 3363 | workloadFactory.Finalize(); |
telsoa01 | 4fcda01 | 2018-03-09 14:13:49 +0000 | [diff] [blame] | 3364 | workload->Execute(); |
| 3365 | |
| 3366 | CopyDataFromITensorHandle(&result.output[0][0][0][0], outputHandle.get()); |
| 3367 | return result; |
| 3368 | } |
| 3369 | |
| 3370 | LayerTestResult<float, 4> L2Normalization4dTest(armnn::IWorkloadFactory& workloadFactory) |
| 3371 | { |
| 3372 | constexpr unsigned int inputWidth = 3; |
| 3373 | constexpr unsigned int inputHeight = 4; |
| 3374 | constexpr unsigned int inputChannels = 3; |
| 3375 | constexpr unsigned int inputBatchSize = 2; |
| 3376 | |
| 3377 | constexpr unsigned int outputWidth = inputWidth; |
| 3378 | constexpr unsigned int outputHeight = inputHeight; |
| 3379 | constexpr unsigned int outputChannels = inputChannels; |
| 3380 | constexpr unsigned int outputBatchSize = inputBatchSize; |
| 3381 | |
| 3382 | const armnn::TensorInfo inputTensorInfo({ inputBatchSize, inputChannels, inputHeight, inputWidth }, |
| 3383 | armnn::DataType::Float32); |
| 3384 | const armnn::TensorInfo outputTensorInfo({ outputBatchSize, outputChannels, outputHeight, outputWidth }, |
| 3385 | armnn::DataType::Float32); |
| 3386 | |
| 3387 | auto input = MakeTensor<float, 4>(inputTensorInfo, std::vector<float>({ |
| 3388 | // Batch 0, Channel 0 |
| 3389 | 235.0f, 46.0f, 178.0f, |
| 3390 | 100.0f, 123.0f, 19.0f, |
| 3391 | 172.0f, 74.0f, 250.0f, |
| 3392 | 6.0f, 195.0f, 80.0f, |
| 3393 | |
| 3394 | // Batch 0, Channel 1 |
| 3395 | 113.0f, 95.0f, 202.0f, |
| 3396 | 77.0f, 114.0f, 71.0f, |
| 3397 | 122.0f, 246.0f, 166.0f, |
| 3398 | 82.0f, 28.0f, 37.0f, |
| 3399 | |
| 3400 | // Batch 0, Channel 2 |
| 3401 | 56.0f, 170.0f, 162.0f, |
| 3402 | 194.0f, 89.0f, 254.0f, |
| 3403 | 12.0f, 209.0f, 200.0f, |
| 3404 | 1.0f, 64.0f, 54.0f, |
| 3405 | |
| 3406 | // Batch 1, Channel 0 |
| 3407 | 67.0f, 90.0f, 49.0f, |
| 3408 | 7.0f, 163.0f, 18.0f, |
| 3409 | 25.0f, 117.0f, 103.0f, |
| 3410 | 247.0f, 59.0f, 189.0f, |
| 3411 | |
| 3412 | // Batch 1, Channel 1 |
| 3413 | 239.0f, 104.0f, 199.0f, |
| 3414 | 17.0f, 124.0f, 153.0f, |
| 3415 | 222.0f, 217.0f, 75.0f, |
| 3416 | 32.0f, 126.0f, 21.0f, |
| 3417 | |
| 3418 | // Batch 1, Channel 2 |
| 3419 | 97.0f, 145.0f, 215.0f, |
| 3420 | 115.0f, 116.0f, 238.0f, |
| 3421 | 226.0f, 16.0f, 132.0f, |
| 3422 | 92.0f, 125.0f, 88.0f, |
| 3423 | })); |
| 3424 | |
| 3425 | LayerTestResult<float, 4> result(outputTensorInfo); |
| 3426 | result.outputExpected = MakeTensor<float, 4>(inputTensorInfo, std::vector<float>({ |
| 3427 | |
| 3428 | // Batch 0, Channel 0 |
| 3429 | 235.0f * CalcInvL2Norm({ 235.0f, 113.0f, 56.0f }), |
| 3430 | 46.0f * CalcInvL2Norm({ 46.0f, 95.0f, 170.0f }), |
| 3431 | 178.0f * CalcInvL2Norm({ 178.0f, 202.0F, 162.0f }), |
| 3432 | 100.0f * CalcInvL2Norm({ 100.0f, 77.0f, 194.0f }), |
| 3433 | 123.0f * CalcInvL2Norm({ 123.0f, 114.0f, 89.0f }), |
| 3434 | 19.0f * CalcInvL2Norm({ 19.0f, 71.0f, 254.0f }), |
| 3435 | 172.0f * CalcInvL2Norm({ 172.0f, 122.0f, 12.0f }), |
| 3436 | 74.0f * CalcInvL2Norm({ 74.0f, 246.0f, 209.0f }), |
| 3437 | 250.0f * CalcInvL2Norm({ 250.0f, 166.0f, 200.0f }), |
| 3438 | 6.0f * CalcInvL2Norm({ 6.0f, 82.0f, 1.0f }), |
| 3439 | 195.0f * CalcInvL2Norm({ 195.0f, 28.0f, 64.0f }), |
| 3440 | 80.0f * CalcInvL2Norm({ 80.0f, 37.0f, 54.0f }), |
| 3441 | |
| 3442 | // Batch 0, Channel 1 |
| 3443 | 113.0f * CalcInvL2Norm({ 235.0f, 113.0f, 56.0f }), |
| 3444 | 95.0f * CalcInvL2Norm({ 46.0f, 95.0f, 170.0f }), |
| 3445 | 202.0f * CalcInvL2Norm({ 178.0f, 202.0F, 162.0f }), |
| 3446 | 77.0f * CalcInvL2Norm({ 100.0f, 77.0f, 194.0f }), |
| 3447 | 114.0f * CalcInvL2Norm({ 123.0f, 114.0f, 89.0f }), |
| 3448 | 71.0f * CalcInvL2Norm({ 19.0f, 71.0f, 254.0f }), |
| 3449 | 122.0f * CalcInvL2Norm({ 172.0f, 122.0f, 12.0f }), |
| 3450 | 246.0f * CalcInvL2Norm({ 74.0f, 246.0f, 209.0f }), |
| 3451 | 166.0f * CalcInvL2Norm({ 250.0f, 166.0f, 200.0f }), |
| 3452 | 82.0f * CalcInvL2Norm({ 6.0f, 82.0f, 1.0f }), |
| 3453 | 28.0f * CalcInvL2Norm({ 195.0f, 28.0f, 64.0f }), |
| 3454 | 37.0f * CalcInvL2Norm({ 80.0f, 37.0f, 54.0f }), |
| 3455 | |
| 3456 | // Batch 0, Channel 2 |
| 3457 | 56.0f * CalcInvL2Norm({ 235.0f, 113.0f, 56.0f }), |
| 3458 | 170.0f * CalcInvL2Norm({ 46.0f, 95.0f, 170.0f }), |
| 3459 | 162.0f * CalcInvL2Norm({ 178.0f, 202.0F, 162.0f }), |
| 3460 | 194.0f * CalcInvL2Norm({ 100.0f, 77.0f, 194.0f }), |
| 3461 | 89.0f * CalcInvL2Norm({ 123.0f, 114.0f, 89.0f }), |
| 3462 | 254.0f * CalcInvL2Norm({ 19.0f, 71.0f, 254.0f }), |
| 3463 | 12.0f * CalcInvL2Norm({ 172.0f, 122.0f, 12.0f }), |
| 3464 | 209.0f * CalcInvL2Norm({ 74.0f, 246.0f, 209.0f }), |
| 3465 | 200.0f * CalcInvL2Norm({ 250.0f, 166.0f, 200.0f }), |
| 3466 | 1.0f * CalcInvL2Norm({ 6.0f, 82.0f, 1.0f }), |
| 3467 | 64.0f * CalcInvL2Norm({ 195.0f, 28.0f, 64.0f }), |
| 3468 | 54.0f * CalcInvL2Norm({ 80.0f, 37.0f, 54.0f }), |
| 3469 | |
| 3470 | // Batch 1, Channel 0 |
| 3471 | 67.0f * CalcInvL2Norm({ 67.0f, 239.0f, 97.0f }), |
| 3472 | 90.0f * CalcInvL2Norm({ 90.0f, 104.0f, 145.0f }), |
| 3473 | 49.0f * CalcInvL2Norm({ 49.0f, 199.0f, 215.0f }), |
| 3474 | 7.0f * CalcInvL2Norm({ 7.0f, 17.0f, 115.0f }), |
| 3475 | 163.0f * CalcInvL2Norm({ 163.0f, 124.0f, 116.0f }), |
| 3476 | 18.0f * CalcInvL2Norm({ 18.0f, 153.0f, 238.0f }), |
| 3477 | 25.0f * CalcInvL2Norm({ 25.0f, 222.0f, 226.0f }), |
| 3478 | 117.0f * CalcInvL2Norm({ 117.0f, 217.0f, 16.0f }), |
| 3479 | 103.0f * CalcInvL2Norm({ 103.0f, 75.0f, 132.0f }), |
| 3480 | 247.0f * CalcInvL2Norm({ 247.0f, 32.0f, 92.0f }), |
| 3481 | 59.0f * CalcInvL2Norm({ 59.0f, 126.0f, 125.0f }), |
| 3482 | 189.0f * CalcInvL2Norm({ 189.0f, 21.0f, 88.0f }), |
| 3483 | |
| 3484 | // Batch 1, Channel 1 |
| 3485 | 239.0f * CalcInvL2Norm({ 67.0f, 239.0f, 97.0f }), |
| 3486 | 104.0f * CalcInvL2Norm({ 90.0f, 104.0f, 145.0f }), |
| 3487 | 199.0f * CalcInvL2Norm({ 49.0f, 199.0f, 215.0f }), |
| 3488 | 17.0f * CalcInvL2Norm({ 7.0f, 17.0f, 115.0f }), |
| 3489 | 124.0f * CalcInvL2Norm({ 163.0f, 124.0f, 116.0f }), |
| 3490 | 153.0f * CalcInvL2Norm({ 18.0f, 153.0f, 238.0f }), |
| 3491 | 222.0f * CalcInvL2Norm({ 25.0f, 222.0f, 226.0f }), |
| 3492 | 217.0f * CalcInvL2Norm({ 117.0f, 217.0f, 16.0f }), |
| 3493 | 75.0f * CalcInvL2Norm({ 103.0f, 75.0f, 132.0f }), |
| 3494 | 32.0f * CalcInvL2Norm({ 247.0f, 32.0f, 92.0f }), |
| 3495 | 126.0f * CalcInvL2Norm({ 59.0f, 126.0f, 125.0f }), |
| 3496 | 21.0f * CalcInvL2Norm({ 189.0f, 21.0f, 88.0f }), |
| 3497 | |
| 3498 | // Batch 1, Channel 2 |
| 3499 | 97.0f * CalcInvL2Norm({ 67.0f, 239.0f, 97.0f }), |
| 3500 | 145.0f * CalcInvL2Norm({ 90.0f, 104.0f, 145.0f }), |
| 3501 | 215.0f * CalcInvL2Norm({ 49.0f, 199.0f, 215.0f }), |
| 3502 | 115.0f * CalcInvL2Norm({ 7.0f, 17.0f, 115.0f }), |
| 3503 | 116.0f * CalcInvL2Norm({ 163.0f, 124.0f, 116.0f }), |
| 3504 | 238.0f * CalcInvL2Norm({ 18.0f, 153.0f, 238.0f }), |
| 3505 | 226.0f * CalcInvL2Norm({ 25.0f, 222.0f, 226.0f }), |
| 3506 | 16.0f * CalcInvL2Norm({ 117.0f, 217.0f, 16.0f }), |
| 3507 | 132.0f * CalcInvL2Norm({ 103.0f, 75.0f, 132.0f }), |
| 3508 | 92.0f * CalcInvL2Norm({ 247.0f, 32.0f, 92.0f }), |
| 3509 | 125.0f * CalcInvL2Norm({ 59.0f, 126.0f, 125.0f }), |
| 3510 | 88.0f * CalcInvL2Norm({ 189.0f, 21.0f, 88.0f }), |
| 3511 | })); |
| 3512 | |
| 3513 | std::unique_ptr<armnn::ITensorHandle> inputHandle = workloadFactory.CreateTensorHandle(inputTensorInfo); |
| 3514 | std::unique_ptr<armnn::ITensorHandle> outputHandle = workloadFactory.CreateTensorHandle(outputTensorInfo); |
| 3515 | |
| 3516 | armnn::L2NormalizationQueueDescriptor descriptor; |
| 3517 | armnn::WorkloadInfo info; |
| 3518 | AddInputToWorkload(descriptor, info, inputTensorInfo, inputHandle.get()); |
| 3519 | AddOutputToWorkload(descriptor, info, outputTensorInfo, outputHandle.get()); |
| 3520 | |
| 3521 | std::unique_ptr<armnn::IWorkload> workload = workloadFactory.CreateL2Normalization(descriptor, info); |
| 3522 | |
| 3523 | inputHandle->Allocate(); |
| 3524 | outputHandle->Allocate(); |
| 3525 | CopyDataToITensorHandle(inputHandle.get(), &input[0][0][0][0]); |
| 3526 | |
surmeh01 | 3537c2c | 2018-05-18 16:31:43 +0100 | [diff] [blame] | 3527 | workloadFactory.Finalize(); |
telsoa01 | 4fcda01 | 2018-03-09 14:13:49 +0000 | [diff] [blame] | 3528 | workload->Execute(); |
| 3529 | |
| 3530 | CopyDataFromITensorHandle(&result.output[0][0][0][0], outputHandle.get()); |
| 3531 | return result; |
| 3532 | } |
| 3533 | |
| 3534 | template <typename T> |
| 3535 | LayerTestResult<T, 4> ConstantTestImpl(armnn::IWorkloadFactory& workloadFactory, |
| 3536 | float qScale, |
| 3537 | int32_t qOffset) |
| 3538 | { |
| 3539 | constexpr unsigned int inputWidth = 3; |
| 3540 | constexpr unsigned int inputHeight = 4; |
| 3541 | constexpr unsigned int inputChannels = 3; |
| 3542 | constexpr unsigned int inputBatchSize = 2; |
| 3543 | |
| 3544 | constexpr unsigned int outputWidth = inputWidth; |
| 3545 | constexpr unsigned int outputHeight = inputHeight; |
| 3546 | constexpr unsigned int outputChannels = inputChannels; |
| 3547 | constexpr unsigned int outputBatchSize = inputBatchSize; |
| 3548 | |
| 3549 | armnn::TensorInfo inputTensorInfo({ inputBatchSize, inputChannels, inputHeight, inputWidth }, |
| 3550 | armnn::GetDataType<T>()); |
| 3551 | |
| 3552 | armnn::TensorInfo outputTensorInfo({ outputBatchSize, outputChannels, outputHeight, outputWidth }, |
| 3553 | armnn::GetDataType<T>()); |
| 3554 | |
| 3555 | // Set quantization parameters if the requested type is a quantized type. |
| 3556 | if(armnn::IsQuantizedType<T>()) |
| 3557 | { |
| 3558 | inputTensorInfo.SetQuantizationScale(qScale); |
| 3559 | inputTensorInfo.SetQuantizationOffset(qOffset); |
| 3560 | outputTensorInfo.SetQuantizationScale(qScale); |
| 3561 | outputTensorInfo.SetQuantizationOffset(qOffset); |
| 3562 | } |
| 3563 | |
| 3564 | auto input = MakeTensor<T, 4>(inputTensorInfo, std::vector<T>( |
| 3565 | QuantizedVector<T>(qScale, qOffset, { |
| 3566 | // Batch 0, Channel 0 |
| 3567 | 235.0f, 46.0f, 178.0f, |
| 3568 | 100.0f, 123.0f, 19.0f, |
| 3569 | 172.0f, 74.0f, 250.0f, |
| 3570 | 6.0f, 195.0f, 80.0f, |
| 3571 | |
| 3572 | // Batch 0, Channel 1 |
| 3573 | 113.0f, 95.0f, 202.0f, |
| 3574 | 77.0f, 114.0f, 71.0f, |
| 3575 | 122.0f, 246.0f, 166.0f, |
| 3576 | 82.0f, 28.0f, 37.0f, |
| 3577 | |
| 3578 | // Batch 0, Channel 2 |
| 3579 | 56.0f, 170.0f, 162.0f, |
| 3580 | 194.0f, 89.0f, 254.0f, |
| 3581 | 12.0f, 209.0f, 200.0f, |
| 3582 | 1.0f, 64.0f, 54.0f, |
| 3583 | |
| 3584 | // Batch 1, Channel 0 |
| 3585 | 67.0f, 90.0f, 49.0f, |
| 3586 | 7.0f, 163.0f, 18.0f, |
| 3587 | 25.0f, 117.0f, 103.0f, |
| 3588 | 247.0f, 59.0f, 189.0f, |
| 3589 | |
| 3590 | // Batch 1, Channel 1 |
| 3591 | 239.0f, 104.0f, 199.0f, |
| 3592 | 17.0f, 124.0f, 153.0f, |
| 3593 | 222.0f, 217.0f, 75.0f, |
| 3594 | 32.0f, 126.0f, 21.0f, |
| 3595 | |
| 3596 | // Batch 1, Channel 2 |
| 3597 | 97.0f, 145.0f, 215.0f, |
| 3598 | 115.0f, 116.0f, 238.0f, |
| 3599 | 226.0f, 16.0f, 132.0f, |
| 3600 | 92.0f, 125.0f, 88.0f, |
| 3601 | }))); |
| 3602 | |
| 3603 | LayerTestResult<T, 4> result(outputTensorInfo); |
| 3604 | result.outputExpected = input; |
| 3605 | |
| 3606 | std::unique_ptr<armnn::ITensorHandle> outputHandle = workloadFactory.CreateTensorHandle(outputTensorInfo); |
| 3607 | |
| 3608 | armnn::ScopedCpuTensorHandle constantTensor(inputTensorInfo); |
| 3609 | AllocateAndCopyDataToITensorHandle(&constantTensor, &input[0][0][0][0]); |
| 3610 | |
| 3611 | armnn::ConstantQueueDescriptor descriptor; |
| 3612 | descriptor.m_LayerOutput = &constantTensor; |
| 3613 | |
| 3614 | armnn::WorkloadInfo info; |
| 3615 | AddOutputToWorkload(descriptor, info, outputTensorInfo, outputHandle.get()); |
| 3616 | |
| 3617 | std::unique_ptr<armnn::IWorkload> workload = workloadFactory.CreateConstant(descriptor, info); |
| 3618 | |
| 3619 | outputHandle->Allocate(); |
| 3620 | |
surmeh01 | 3537c2c | 2018-05-18 16:31:43 +0100 | [diff] [blame] | 3621 | workloadFactory.Finalize(); |
telsoa01 | 4fcda01 | 2018-03-09 14:13:49 +0000 | [diff] [blame] | 3622 | workload->Execute(); |
| 3623 | |
| 3624 | CopyDataFromITensorHandle(&result.output[0][0][0][0], outputHandle.get()); |
| 3625 | return result; |
| 3626 | } |
| 3627 | |
| 3628 | LayerTestResult<float, 4> ConstantTest(armnn::IWorkloadFactory& workloadFactory) |
| 3629 | { |
| 3630 | return ConstantTestImpl<float>(workloadFactory, 0.0f, 0); |
| 3631 | } |
| 3632 | |
| 3633 | LayerTestResult<uint8_t, 4> ConstantTestUint8(armnn::IWorkloadFactory& workloadFactory) |
| 3634 | { |
| 3635 | return ConstantTestImpl<uint8_t>(workloadFactory, 1.0f, 0); |
| 3636 | } |
| 3637 | |
| 3638 | LayerTestResult<uint8_t, 3> MergerUint8Test(armnn::IWorkloadFactory& workloadFactory) |
| 3639 | { |
surmeh01 | 3537c2c | 2018-05-18 16:31:43 +0100 | [diff] [blame] | 3640 | unsigned int outputWidth = 3; |
telsoa01 | 4fcda01 | 2018-03-09 14:13:49 +0000 | [diff] [blame] | 3641 | unsigned int outputHeight = 6; |
| 3642 | unsigned int outputChannels = 3; |
| 3643 | |
surmeh01 | 3537c2c | 2018-05-18 16:31:43 +0100 | [diff] [blame] | 3644 | unsigned int inputWidth1 = 3; |
| 3645 | unsigned int inputHeight1 = 6; |
| 3646 | unsigned int inputChannels1 = 2; |
telsoa01 | 4fcda01 | 2018-03-09 14:13:49 +0000 | [diff] [blame] | 3647 | |
surmeh01 | 3537c2c | 2018-05-18 16:31:43 +0100 | [diff] [blame] | 3648 | unsigned int inputWidth2 = 3; |
| 3649 | unsigned int inputHeight2 = 6; |
| 3650 | unsigned int inputChannels2 = 1; |
telsoa01 | 4fcda01 | 2018-03-09 14:13:49 +0000 | [diff] [blame] | 3651 | |
telsoa01 | c577f2c | 2018-08-31 09:22:23 +0100 | [diff] [blame] | 3652 | // Defines the tensor descriptors. |
telsoa01 | 4fcda01 | 2018-03-09 14:13:49 +0000 | [diff] [blame] | 3653 | armnn::TensorInfo outputTensorInfo({ outputChannels, outputHeight, outputWidth }, armnn::DataType::QuantisedAsymm8); |
| 3654 | armnn::TensorInfo inputTensorInfo1({ inputChannels1, inputHeight1, inputWidth1 }, armnn::DataType::QuantisedAsymm8); |
| 3655 | armnn::TensorInfo inputTensorInfo2({ inputChannels2, inputHeight2, inputWidth2 }, armnn::DataType::QuantisedAsymm8); |
telsoa01 | 4fcda01 | 2018-03-09 14:13:49 +0000 | [diff] [blame] | 3656 | |
telsoa01 | c577f2c | 2018-08-31 09:22:23 +0100 | [diff] [blame] | 3657 | // Arbitrary scale and offsets. They don't really matter as the merger operator doesn't dequantize/quantize them. |
telsoa01 | 4fcda01 | 2018-03-09 14:13:49 +0000 | [diff] [blame] | 3658 | const float scale = 0.13497836f; |
| 3659 | const int32_t offset = -7; |
| 3660 | |
| 3661 | outputTensorInfo.SetQuantizationScale(scale); |
| 3662 | outputTensorInfo.SetQuantizationOffset(offset); |
| 3663 | inputTensorInfo1.SetQuantizationScale(scale); |
| 3664 | inputTensorInfo1.SetQuantizationOffset(offset); |
| 3665 | inputTensorInfo2.SetQuantizationScale(scale); |
| 3666 | inputTensorInfo2.SetQuantizationOffset(offset); |
telsoa01 | 4fcda01 | 2018-03-09 14:13:49 +0000 | [diff] [blame] | 3667 | |
| 3668 | LayerTestResult<uint8_t, 3> ret(outputTensorInfo); |
| 3669 | |
| 3670 | ret.outputExpected = MakeTensor<uint8_t, 3>(outputTensorInfo, std::vector<uint8_t>( |
surmeh01 | 3537c2c | 2018-05-18 16:31:43 +0100 | [diff] [blame] | 3671 | { |
| 3672 | 1, 2, 3, |
| 3673 | 4, 5, 6, |
| 3674 | 7, 8, 9, |
| 3675 | 10, 11, 12, |
| 3676 | 13, 14, 15, |
| 3677 | 16, 17, 18, |
telsoa01 | 4fcda01 | 2018-03-09 14:13:49 +0000 | [diff] [blame] | 3678 | |
surmeh01 | 3537c2c | 2018-05-18 16:31:43 +0100 | [diff] [blame] | 3679 | 19, 20, 21, |
| 3680 | 22, 23, 24, |
| 3681 | 25, 26, 27, |
| 3682 | 28, 29, 30, |
| 3683 | 31, 32, 33, |
| 3684 | 34, 35, 36, |
telsoa01 | 4fcda01 | 2018-03-09 14:13:49 +0000 | [diff] [blame] | 3685 | |
surmeh01 | 3537c2c | 2018-05-18 16:31:43 +0100 | [diff] [blame] | 3686 | 37, 38, 39, |
| 3687 | 40, 41, 42, |
| 3688 | 43, 44, 45, |
| 3689 | 46, 47, 48, |
| 3690 | 49, 50, 51, |
| 3691 | 52, 53, 54, |
| 3692 | }) |
telsoa01 | 4fcda01 | 2018-03-09 14:13:49 +0000 | [diff] [blame] | 3693 | ); |
| 3694 | |
telsoa01 | 4fcda01 | 2018-03-09 14:13:49 +0000 | [diff] [blame] | 3695 | auto input1 = MakeTensor<uint8_t, 3>(inputTensorInfo1, std::vector<uint8_t>( |
| 3696 | { |
surmeh01 | 3537c2c | 2018-05-18 16:31:43 +0100 | [diff] [blame] | 3697 | 1, 2, 3, |
| 3698 | 4, 5, 6, |
| 3699 | 7, 8, 9, |
| 3700 | 10, 11, 12, |
| 3701 | 13, 14, 15, |
| 3702 | 16, 17, 18, |
telsoa01 | 4fcda01 | 2018-03-09 14:13:49 +0000 | [diff] [blame] | 3703 | |
surmeh01 | 3537c2c | 2018-05-18 16:31:43 +0100 | [diff] [blame] | 3704 | 19, 20, 21, |
| 3705 | 22, 23, 24, |
| 3706 | 25, 26, 27, |
| 3707 | 28, 29, 30, |
| 3708 | 31, 32, 33, |
| 3709 | 34, 35, 36, |
telsoa01 | 4fcda01 | 2018-03-09 14:13:49 +0000 | [diff] [blame] | 3710 | }) |
| 3711 | ); |
| 3712 | |
| 3713 | auto input2 = MakeTensor<uint8_t, 3>(inputTensorInfo2, std::vector<uint8_t>( |
| 3714 | { |
surmeh01 | 3537c2c | 2018-05-18 16:31:43 +0100 | [diff] [blame] | 3715 | 37, 38, 39, |
| 3716 | 40, 41, 42, |
telsoa01 | 4fcda01 | 2018-03-09 14:13:49 +0000 | [diff] [blame] | 3717 | 43, 44, 45, |
surmeh01 | 3537c2c | 2018-05-18 16:31:43 +0100 | [diff] [blame] | 3718 | 46, 47, 48, |
| 3719 | 49, 50, 51, |
| 3720 | 52, 53, 54, |
telsoa01 | 4fcda01 | 2018-03-09 14:13:49 +0000 | [diff] [blame] | 3721 | }) |
| 3722 | ); |
| 3723 | |
telsoa01 | c577f2c | 2018-08-31 09:22:23 +0100 | [diff] [blame] | 3724 | std::vector<unsigned int> wOrigin1 = { 0, 0, 0 }; //Extent of the window is defined by size of input[0]. |
telsoa01 | 4fcda01 | 2018-03-09 14:13:49 +0000 | [diff] [blame] | 3725 | armnn::MergerQueueDescriptor::ViewOrigin window1(wOrigin1); |
| 3726 | |
telsoa01 | c577f2c | 2018-08-31 09:22:23 +0100 | [diff] [blame] | 3727 | std::vector<unsigned int> wOrigin2 = { 2, 0, 0 }; //Extent of the window is defined by size of input[1]. |
telsoa01 | 4fcda01 | 2018-03-09 14:13:49 +0000 | [diff] [blame] | 3728 | armnn::MergerQueueDescriptor::ViewOrigin window2(wOrigin2); |
| 3729 | |
telsoa01 | 4fcda01 | 2018-03-09 14:13:49 +0000 | [diff] [blame] | 3730 | |
| 3731 | std::unique_ptr<armnn::ITensorHandle> outputHandle = workloadFactory.CreateTensorHandle(outputTensorInfo); |
| 3732 | |
| 3733 | bool subTensorsSupported = workloadFactory.SupportsSubTensors(); |
| 3734 | |
| 3735 | std::unique_ptr<armnn::ITensorHandle> inputHandle1 = |
| 3736 | subTensorsSupported ? |
| 3737 | workloadFactory.CreateSubTensorHandle(*outputHandle, inputTensorInfo1.GetShape(), wOrigin1.data()) : |
| 3738 | workloadFactory.CreateTensorHandle(inputTensorInfo1); |
| 3739 | |
| 3740 | std::unique_ptr<armnn::ITensorHandle> inputHandle2 = |
| 3741 | subTensorsSupported ? |
| 3742 | workloadFactory.CreateSubTensorHandle(*outputHandle, inputTensorInfo2.GetShape(), wOrigin2.data()) : |
| 3743 | workloadFactory.CreateTensorHandle(inputTensorInfo2); |
| 3744 | |
telsoa01 | 4fcda01 | 2018-03-09 14:13:49 +0000 | [diff] [blame] | 3745 | |
| 3746 | armnn::MergerQueueDescriptor data; |
| 3747 | armnn::WorkloadInfo info; |
| 3748 | AddInputToWorkload(data, info, inputTensorInfo1, inputHandle1.get()); |
| 3749 | AddInputToWorkload(data, info, inputTensorInfo2, inputHandle2.get()); |
telsoa01 | 4fcda01 | 2018-03-09 14:13:49 +0000 | [diff] [blame] | 3750 | AddOutputToWorkload(data, info, outputTensorInfo, outputHandle.get()); |
| 3751 | |
| 3752 | data.m_ViewOrigins.push_back(window1); |
| 3753 | data.m_ViewOrigins.push_back(window2); |
telsoa01 | 4fcda01 | 2018-03-09 14:13:49 +0000 | [diff] [blame] | 3754 | |
| 3755 | std::unique_ptr<armnn::IWorkload> workload = workloadFactory.CreateMerger(data, info); |
| 3756 | |
| 3757 | inputHandle1->Allocate(); |
| 3758 | inputHandle2->Allocate(); |
telsoa01 | 4fcda01 | 2018-03-09 14:13:49 +0000 | [diff] [blame] | 3759 | outputHandle->Allocate(); |
| 3760 | |
| 3761 | CopyDataToITensorHandle(inputHandle1.get(), &input1[0][0][0]); |
| 3762 | CopyDataToITensorHandle(inputHandle2.get(), &input2[0][0][0]); |
telsoa01 | 4fcda01 | 2018-03-09 14:13:49 +0000 | [diff] [blame] | 3763 | |
surmeh01 | 3537c2c | 2018-05-18 16:31:43 +0100 | [diff] [blame] | 3764 | workloadFactory.Finalize(); |
telsoa01 | 4fcda01 | 2018-03-09 14:13:49 +0000 | [diff] [blame] | 3765 | workload->Execute(); |
| 3766 | |
| 3767 | CopyDataFromITensorHandle(&ret.output[0][0][0], outputHandle.get()); |
| 3768 | |
| 3769 | return ret; |
| 3770 | } |
| 3771 | |
| 3772 | LayerTestResult<uint8_t, 4> AdditionUint8Test(armnn::IWorkloadFactory& workloadFactory) |
| 3773 | { |
| 3774 | unsigned int batchSize = 1; |
| 3775 | unsigned int channels = 2; |
| 3776 | unsigned int height = 2; |
| 3777 | unsigned int width = 3; |
| 3778 | |
| 3779 | const float scale = 7.0f; |
| 3780 | const int32_t offset = 3; |
| 3781 | |
| 3782 | armnn::TensorInfo inputTensorInfo1, inputTensorInfo2; |
| 3783 | armnn::TensorInfo outputTensorInfo; |
| 3784 | |
| 3785 | const unsigned int shape[] = { batchSize, channels, height, width }; |
| 3786 | inputTensorInfo1 = armnn::TensorInfo(4, shape, armnn::DataType::QuantisedAsymm8); |
| 3787 | inputTensorInfo1.SetQuantizationScale(scale); |
| 3788 | inputTensorInfo1.SetQuantizationOffset(offset); |
| 3789 | |
| 3790 | inputTensorInfo2 = armnn::TensorInfo(4, shape, armnn::DataType::QuantisedAsymm8); |
| 3791 | inputTensorInfo2.SetQuantizationScale(scale); |
| 3792 | inputTensorInfo2.SetQuantizationOffset(offset); |
| 3793 | |
| 3794 | outputTensorInfo = armnn::TensorInfo(4, shape, armnn::DataType::QuantisedAsymm8); |
| 3795 | outputTensorInfo.SetQuantizationScale(scale); |
| 3796 | outputTensorInfo.SetQuantizationOffset(offset); |
| 3797 | |
telsoa01 | c577f2c | 2018-08-31 09:22:23 +0100 | [diff] [blame] | 3798 | // See dequantized values to the right. |
telsoa01 | 4fcda01 | 2018-03-09 14:13:49 +0000 | [diff] [blame] | 3799 | auto input1 = MakeTensor<uint8_t, 4>(inputTensorInfo1, std::vector<uint8_t>( |
| 3800 | { |
| 3801 | 63, 35, 77, 70, 56, 112, // 420, 224, 518, 469, 371, 763 |
| 3802 | 203, 28, 252, 168, 245, 91 // 1400, 175, 1743, 1155, 1694, 616 |
| 3803 | })); |
| 3804 | |
telsoa01 | c577f2c | 2018-08-31 09:22:23 +0100 | [diff] [blame] | 3805 | // See dequantized values to the right. |
telsoa01 | 4fcda01 | 2018-03-09 14:13:49 +0000 | [diff] [blame] | 3806 | auto input2 = MakeTensor<uint8_t, 4>(inputTensorInfo1, std::vector<uint8_t>( |
| 3807 | { |
| 3808 | 21, 7, 175, 231, 175, 210, // 126, 28, 1204, 1596, 1204, 1449 |
| 3809 | 126, 161, 63, 21, 105, 126 // 861, 1106, 420, 126, 714, 861 |
| 3810 | })); |
| 3811 | |
telsoa01 | c577f2c | 2018-08-31 09:22:23 +0100 | [diff] [blame] | 3812 | // See dequantized values to the right. |
telsoa01 | 4fcda01 | 2018-03-09 14:13:49 +0000 | [diff] [blame] | 3813 | LayerTestResult<uint8_t, 4> result(outputTensorInfo); |
| 3814 | result.outputExpected = MakeTensor<uint8_t, 4>(outputTensorInfo, std::vector<uint8_t>( |
| 3815 | { |
| 3816 | 81, 39, 249, 255, 228, 255, // 546, 252, 1722, 2065(clamped), 1575, 2212(clamped) |
| 3817 | 255, 186, 255, 186, 255, 214, // 2261(clamped), 1281, 2163(clamped), 1281, 2408(clamped), 1477 |
| 3818 | })); |
| 3819 | |
| 3820 | std::unique_ptr<armnn::ITensorHandle> inputHandle1 = workloadFactory.CreateTensorHandle(inputTensorInfo1); |
| 3821 | std::unique_ptr<armnn::ITensorHandle> inputHandle2 = workloadFactory.CreateTensorHandle(inputTensorInfo2); |
| 3822 | std::unique_ptr<armnn::ITensorHandle> outputHandle = workloadFactory.CreateTensorHandle(outputTensorInfo); |
| 3823 | |
| 3824 | armnn::AdditionQueueDescriptor data; |
| 3825 | armnn::WorkloadInfo info; |
| 3826 | AddInputToWorkload(data, info, inputTensorInfo1, inputHandle1.get()); |
| 3827 | AddInputToWorkload(data, info, inputTensorInfo2, inputHandle2.get()); |
| 3828 | AddOutputToWorkload(data, info, outputTensorInfo, outputHandle.get()); |
| 3829 | |
| 3830 | std::unique_ptr<armnn::IWorkload> workload = workloadFactory.CreateAddition(data, info); |
| 3831 | |
| 3832 | inputHandle1->Allocate(); |
| 3833 | inputHandle2->Allocate(); |
| 3834 | outputHandle->Allocate(); |
| 3835 | |
| 3836 | CopyDataToITensorHandle(inputHandle1.get(), &input1[0][0][0][0]); |
| 3837 | CopyDataToITensorHandle(inputHandle2.get(), &input2[0][0][0][0]); |
| 3838 | |
surmeh01 | 3537c2c | 2018-05-18 16:31:43 +0100 | [diff] [blame] | 3839 | workloadFactory.Finalize(); |
telsoa01 | 4fcda01 | 2018-03-09 14:13:49 +0000 | [diff] [blame] | 3840 | workload->Execute(); |
| 3841 | |
| 3842 | CopyDataFromITensorHandle(&result.output[0][0][0][0], outputHandle.get()); |
| 3843 | |
| 3844 | return result; |
| 3845 | } |
| 3846 | |
surmeh01 | bceff2f | 2018-03-29 16:29:27 +0100 | [diff] [blame] | 3847 | namespace |
telsoa01 | 4fcda01 | 2018-03-09 14:13:49 +0000 | [diff] [blame] | 3848 | { |
surmeh01 | bceff2f | 2018-03-29 16:29:27 +0100 | [diff] [blame] | 3849 | LayerTestResult<uint8_t, 4> MultiplicationUint8TestHelper(armnn::IWorkloadFactory& workloadFactory, |
| 3850 | const unsigned int shape0[4], |
| 3851 | const std::vector<uint8_t> & values0, |
| 3852 | float scale0, |
| 3853 | int32_t offset0, |
| 3854 | const unsigned int shape1[4], |
| 3855 | const std::vector<uint8_t> & values1, |
| 3856 | float scale1, |
| 3857 | int32_t offset1, |
| 3858 | const unsigned int outShape[4], |
| 3859 | const std::vector<uint8_t> & outValues, |
| 3860 | float outScale, |
| 3861 | int32_t outOffset) |
| 3862 | { |
| 3863 | armnn::TensorInfo inputTensorInfo0(4, shape0, armnn::DataType::QuantisedAsymm8); |
| 3864 | armnn::TensorInfo inputTensorInfo1(4, shape1, armnn::DataType::QuantisedAsymm8); |
| 3865 | armnn::TensorInfo outputTensorInfo(4, outShape, armnn::DataType::QuantisedAsymm8); |
telsoa01 | 4fcda01 | 2018-03-09 14:13:49 +0000 | [diff] [blame] | 3866 | |
surmeh01 | bceff2f | 2018-03-29 16:29:27 +0100 | [diff] [blame] | 3867 | inputTensorInfo0.SetQuantizationScale(scale0); |
| 3868 | inputTensorInfo0.SetQuantizationOffset(offset0); |
telsoa01 | 4fcda01 | 2018-03-09 14:13:49 +0000 | [diff] [blame] | 3869 | |
surmeh01 | bceff2f | 2018-03-29 16:29:27 +0100 | [diff] [blame] | 3870 | inputTensorInfo1.SetQuantizationScale(scale1); |
| 3871 | inputTensorInfo1.SetQuantizationOffset(offset1); |
telsoa01 | 4fcda01 | 2018-03-09 14:13:49 +0000 | [diff] [blame] | 3872 | |
surmeh01 | bceff2f | 2018-03-29 16:29:27 +0100 | [diff] [blame] | 3873 | outputTensorInfo.SetQuantizationScale(outScale); |
| 3874 | outputTensorInfo.SetQuantizationOffset(outOffset); |
telsoa01 | 4fcda01 | 2018-03-09 14:13:49 +0000 | [diff] [blame] | 3875 | |
surmeh01 | bceff2f | 2018-03-29 16:29:27 +0100 | [diff] [blame] | 3876 | auto input0 = MakeTensor<uint8_t, 4>(inputTensorInfo0, values0); |
| 3877 | auto input1 = MakeTensor<uint8_t, 4>(inputTensorInfo1, values1); |
telsoa01 | 4fcda01 | 2018-03-09 14:13:49 +0000 | [diff] [blame] | 3878 | |
telsoa01 | 4fcda01 | 2018-03-09 14:13:49 +0000 | [diff] [blame] | 3879 | LayerTestResult<uint8_t, 4> result(outputTensorInfo); |
surmeh01 | bceff2f | 2018-03-29 16:29:27 +0100 | [diff] [blame] | 3880 | result.outputExpected = MakeTensor<uint8_t, 4>(outputTensorInfo, outValues); |
telsoa01 | 4fcda01 | 2018-03-09 14:13:49 +0000 | [diff] [blame] | 3881 | |
surmeh01 | bceff2f | 2018-03-29 16:29:27 +0100 | [diff] [blame] | 3882 | std::unique_ptr<armnn::ITensorHandle> inputHandle0 = workloadFactory.CreateTensorHandle(inputTensorInfo0); |
telsoa01 | 4fcda01 | 2018-03-09 14:13:49 +0000 | [diff] [blame] | 3883 | std::unique_ptr<armnn::ITensorHandle> inputHandle1 = workloadFactory.CreateTensorHandle(inputTensorInfo1); |
telsoa01 | 4fcda01 | 2018-03-09 14:13:49 +0000 | [diff] [blame] | 3884 | std::unique_ptr<armnn::ITensorHandle> outputHandle = workloadFactory.CreateTensorHandle(outputTensorInfo); |
| 3885 | |
| 3886 | armnn::MultiplicationQueueDescriptor data; |
| 3887 | armnn::WorkloadInfo info; |
surmeh01 | bceff2f | 2018-03-29 16:29:27 +0100 | [diff] [blame] | 3888 | AddInputToWorkload(data, info, inputTensorInfo0, inputHandle0.get()); |
| 3889 | AddInputToWorkload(data, info, inputTensorInfo1, inputHandle1.get()); |
telsoa01 | 4fcda01 | 2018-03-09 14:13:49 +0000 | [diff] [blame] | 3890 | AddOutputToWorkload(data, info, outputTensorInfo, outputHandle.get()); |
| 3891 | |
| 3892 | std::unique_ptr<armnn::IWorkload> workload = workloadFactory.CreateMultiplication(data, info); |
| 3893 | |
surmeh01 | bceff2f | 2018-03-29 16:29:27 +0100 | [diff] [blame] | 3894 | inputHandle0->Allocate(); |
telsoa01 | 4fcda01 | 2018-03-09 14:13:49 +0000 | [diff] [blame] | 3895 | inputHandle1->Allocate(); |
telsoa01 | 4fcda01 | 2018-03-09 14:13:49 +0000 | [diff] [blame] | 3896 | outputHandle->Allocate(); |
| 3897 | |
surmeh01 | bceff2f | 2018-03-29 16:29:27 +0100 | [diff] [blame] | 3898 | CopyDataToITensorHandle(inputHandle0.get(), &input0[0][0][0][0]); |
telsoa01 | 4fcda01 | 2018-03-09 14:13:49 +0000 | [diff] [blame] | 3899 | CopyDataToITensorHandle(inputHandle1.get(), &input1[0][0][0][0]); |
telsoa01 | 4fcda01 | 2018-03-09 14:13:49 +0000 | [diff] [blame] | 3900 | |
surmeh01 | 3537c2c | 2018-05-18 16:31:43 +0100 | [diff] [blame] | 3901 | workloadFactory.Finalize(); |
telsoa01 | 4fcda01 | 2018-03-09 14:13:49 +0000 | [diff] [blame] | 3902 | workload->Execute(); |
| 3903 | |
| 3904 | CopyDataFromITensorHandle(&result.output[0][0][0][0], outputHandle.get()); |
| 3905 | |
| 3906 | return result; |
| 3907 | } |
surmeh01 | bceff2f | 2018-03-29 16:29:27 +0100 | [diff] [blame] | 3908 | } // anonymous namespace |
| 3909 | |
| 3910 | LayerTestResult<uint8_t, 4> MultiplicationUint8Test(armnn::IWorkloadFactory& workloadFactory) |
| 3911 | { |
| 3912 | unsigned int batchSize = 1; |
| 3913 | unsigned int channels = 2; |
| 3914 | unsigned int height = 2; |
| 3915 | unsigned int width = 3; |
| 3916 | const unsigned int shape[] = { batchSize, channels, height, width }; |
| 3917 | |
telsoa01 | c577f2c | 2018-08-31 09:22:23 +0100 | [diff] [blame] | 3918 | // See dequantized values to the right. |
surmeh01 | bceff2f | 2018-03-29 16:29:27 +0100 | [diff] [blame] | 3919 | std::vector<uint8_t> input0({ |
| 3920 | 62, 37, 3, 172, 13, 111, // 244, 144, 8, 684, 48, 440, |
| 3921 | 188, 20, 73, 31, 23, 31 // 748, 76, 288, 120, 88, 120 |
| 3922 | }); |
| 3923 | |
telsoa01 | c577f2c | 2018-08-31 09:22:23 +0100 | [diff] [blame] | 3924 | // See dequantized values to the right. |
surmeh01 | bceff2f | 2018-03-29 16:29:27 +0100 | [diff] [blame] | 3925 | std::vector<uint8_t> input1({ |
| 3926 | 126, 240, 252, 183, 121, 247, // 384, 726, 762, 555, 369, 747, |
| 3927 | 48, 115, 151, 79, 78, 97 // 150, 351, 459, 243, 240, 297 |
| 3928 | }); |
| 3929 | |
telsoa01 | c577f2c | 2018-08-31 09:22:23 +0100 | [diff] [blame] | 3930 | // See dequantized values to the right. |
surmeh01 | bceff2f | 2018-03-29 16:29:27 +0100 | [diff] [blame] | 3931 | std::vector<uint8_t> output( |
| 3932 | { |
| 3933 | 64, 72, 0, 255, 8, 236, // 93696, 104544, 6096(clamped), 379620(clamped), 17712, 328680, |
| 3934 | 77, 15, 92, 16, 10, 21, // 112200, 26676, 132192, 29160, 21120, 35640 |
| 3935 | }); |
| 3936 | |
| 3937 | return MultiplicationUint8TestHelper(workloadFactory, |
| 3938 | shape, |
| 3939 | input0, |
| 3940 | 4.0f, |
| 3941 | 1, |
| 3942 | shape, |
| 3943 | input1, |
| 3944 | 3.0f, |
| 3945 | -2, |
| 3946 | shape, |
| 3947 | output, |
telsoa01 | c577f2c | 2018-08-31 09:22:23 +0100 | [diff] [blame] | 3948 | 1366.255f, // Scale/offset chosen to have output values out of range. |
surmeh01 | bceff2f | 2018-03-29 16:29:27 +0100 | [diff] [blame] | 3949 | -5); |
| 3950 | } |
| 3951 | |
| 3952 | LayerTestResult<uint8_t, 4> MultiplicationBroadcast1ElementUint8Test(armnn::IWorkloadFactory& workloadFactory) |
| 3953 | { |
| 3954 | const unsigned int shape0[] = { 1, 2, 2, 3 }; |
| 3955 | const unsigned int shape1[] = { 1, 1, 1, 1 }; |
| 3956 | |
| 3957 | std::vector<uint8_t> input0({ |
| 3958 | 1, 2, 3, 4, 5, 6, |
| 3959 | 7, 8, 9, 10, 11, 12 |
| 3960 | }); |
| 3961 | |
| 3962 | std::vector<uint8_t> input1({2}); |
| 3963 | |
| 3964 | std::vector<uint8_t> output({ |
| 3965 | 2, 4, 6, 8, 10, 12, |
| 3966 | 14, 16, 18, 20, 22, 24 |
| 3967 | }); |
| 3968 | |
| 3969 | return MultiplicationUint8TestHelper(workloadFactory, |
| 3970 | shape0, |
| 3971 | input0, |
| 3972 | 1.0f, |
| 3973 | 0, |
| 3974 | shape1, |
| 3975 | input1, |
| 3976 | 1.0f, |
| 3977 | 0, |
| 3978 | shape0, |
| 3979 | output, |
| 3980 | 1.0f, |
| 3981 | 0); |
| 3982 | } |
| 3983 | |
| 3984 | LayerTestResult<uint8_t, 4> MultiplicationBroadcast1DVectorUint8Test(armnn::IWorkloadFactory& workloadFactory) |
| 3985 | { |
| 3986 | const unsigned int shape0[] = { 1, 2, 2, 3 }; |
| 3987 | const unsigned int shape1[] = { 1, 1, 1, 3 }; |
| 3988 | |
| 3989 | std::vector<uint8_t> input0({ |
| 3990 | 1, 2, 3, 4, 5, 6, |
| 3991 | 7, 8, 9, 10, 11, 12 |
| 3992 | }); |
| 3993 | |
| 3994 | std::vector<uint8_t> input1({1, 2, 3}); |
| 3995 | |
| 3996 | std::vector<uint8_t> output({ |
| 3997 | 1, 4, 9, 4, 10, 18, |
| 3998 | 7, 16, 27, 10, 22, 36 |
| 3999 | }); |
| 4000 | |
| 4001 | return MultiplicationUint8TestHelper(workloadFactory, |
| 4002 | shape0, |
| 4003 | input0, |
| 4004 | 1.0f, |
| 4005 | 0, |
| 4006 | shape1, |
| 4007 | input1, |
| 4008 | 1.0f, |
| 4009 | 0, |
| 4010 | shape0, |
| 4011 | output, |
| 4012 | 1.0f, |
| 4013 | 0); |
| 4014 | } |
telsoa01 | 4fcda01 | 2018-03-09 14:13:49 +0000 | [diff] [blame] | 4015 | |
David Beck | f195f03 | 2018-09-06 16:46:34 +0100 | [diff] [blame] | 4016 | namespace |
| 4017 | { |
| 4018 | template <typename T> |
| 4019 | LayerTestResult<T, 4> SubtractionTestHelper(armnn::IWorkloadFactory& workloadFactory, |
| 4020 | const unsigned int shape0[4], |
| 4021 | const std::vector<T>& values0, |
| 4022 | float scale0, |
| 4023 | int32_t offset0, |
| 4024 | const unsigned int shape1[4], |
| 4025 | const std::vector<T> & values1, |
| 4026 | float scale1, |
| 4027 | int32_t offset1, |
| 4028 | const unsigned int outShape[4], |
| 4029 | const std::vector<T> & outValues, |
| 4030 | float outScale, |
| 4031 | int32_t outOffset) |
| 4032 | { |
| 4033 | auto dataType = (std::is_same<T, uint8_t>::value ? |
| 4034 | armnn::DataType::QuantisedAsymm8 : |
| 4035 | armnn::DataType::Float32); |
| 4036 | |
| 4037 | armnn::TensorInfo inputTensorInfo0(4, shape0, dataType); |
| 4038 | armnn::TensorInfo inputTensorInfo1(4, shape1, dataType); |
| 4039 | armnn::TensorInfo outputTensorInfo(4, outShape, dataType); |
| 4040 | |
| 4041 | inputTensorInfo0.SetQuantizationScale(scale0); |
| 4042 | inputTensorInfo0.SetQuantizationOffset(offset0); |
| 4043 | |
| 4044 | inputTensorInfo1.SetQuantizationScale(scale1); |
| 4045 | inputTensorInfo1.SetQuantizationOffset(offset1); |
| 4046 | |
| 4047 | outputTensorInfo.SetQuantizationScale(outScale); |
| 4048 | outputTensorInfo.SetQuantizationOffset(outOffset); |
| 4049 | |
| 4050 | auto input0 = MakeTensor<T, 4>(inputTensorInfo0, values0); |
| 4051 | auto input1 = MakeTensor<T, 4>(inputTensorInfo1, values1); |
| 4052 | |
| 4053 | LayerTestResult<T, 4> result(outputTensorInfo); |
| 4054 | result.outputExpected = MakeTensor<T, 4>(outputTensorInfo, outValues); |
| 4055 | |
| 4056 | std::unique_ptr<armnn::ITensorHandle> inputHandle0 = workloadFactory.CreateTensorHandle(inputTensorInfo0); |
| 4057 | std::unique_ptr<armnn::ITensorHandle> inputHandle1 = workloadFactory.CreateTensorHandle(inputTensorInfo1); |
| 4058 | std::unique_ptr<armnn::ITensorHandle> outputHandle = workloadFactory.CreateTensorHandle(outputTensorInfo); |
| 4059 | |
| 4060 | armnn::SubtractionQueueDescriptor data; |
| 4061 | armnn::WorkloadInfo info; |
| 4062 | AddInputToWorkload(data, info, inputTensorInfo0, inputHandle0.get()); |
| 4063 | AddInputToWorkload(data, info, inputTensorInfo1, inputHandle1.get()); |
| 4064 | AddOutputToWorkload(data, info, outputTensorInfo, outputHandle.get()); |
| 4065 | |
| 4066 | std::unique_ptr<armnn::IWorkload> workload = workloadFactory.CreateSubtraction(data, info); |
| 4067 | |
| 4068 | inputHandle0->Allocate(); |
| 4069 | inputHandle1->Allocate(); |
| 4070 | outputHandle->Allocate(); |
| 4071 | |
| 4072 | CopyDataToITensorHandle(inputHandle0.get(), &input0[0][0][0][0]); |
| 4073 | CopyDataToITensorHandle(inputHandle1.get(), &input1[0][0][0][0]); |
| 4074 | |
| 4075 | workloadFactory.Finalize(); |
| 4076 | workload->Execute(); |
| 4077 | |
| 4078 | CopyDataFromITensorHandle(&result.output[0][0][0][0], outputHandle.get()); |
| 4079 | |
| 4080 | return result; |
| 4081 | } |
| 4082 | } // anonymous namespace |
| 4083 | |
| 4084 | LayerTestResult<uint8_t, 4> SubtractionUint8Test(armnn::IWorkloadFactory& workloadFactory) |
| 4085 | { |
| 4086 | const unsigned int shape0[] = { 1, 1, 2, 2 }; |
| 4087 | const unsigned int shape1[] = { 1, 1, 2, 2 }; |
| 4088 | |
| 4089 | std::vector<uint8_t> input0({ 10, 12, 14, 16 }); |
| 4090 | std::vector<uint8_t> input1({ 1, 2, 1, 2 }); |
| 4091 | std::vector<uint8_t> output({ 3, 3, 5, 5 }); |
| 4092 | |
| 4093 | return SubtractionTestHelper(workloadFactory, |
| 4094 | shape0, input0, 0.5f, 2, |
| 4095 | shape1, input1, 1.0f, 0, |
| 4096 | shape0, output, 1.0f, 0); |
| 4097 | } |
| 4098 | |
| 4099 | LayerTestResult<uint8_t, 4> SubtractionBroadcast1ElementUint8Test(armnn::IWorkloadFactory& workloadFactory) |
| 4100 | { |
| 4101 | const unsigned int shape0[] = { 1, 1, 2, 2 }; |
| 4102 | const unsigned int shape1[] = { 1, 1, 1, 1 }; |
| 4103 | |
| 4104 | std::vector<uint8_t> input0({ 10, 12, 14, 16 }); |
| 4105 | std::vector<uint8_t> input1({ 2 }); |
| 4106 | std::vector<uint8_t> output({ 5, 6, 7, 8 }); |
| 4107 | |
| 4108 | return SubtractionTestHelper(workloadFactory, |
| 4109 | shape0, input0, 0.5f, 2, |
| 4110 | shape1, input1, 1.0f, 0, |
| 4111 | shape0, output, 1.0f, 3); |
| 4112 | } |
| 4113 | |
| 4114 | LayerTestResult<uint8_t, 4> SubtractionBroadcastUint8Test(armnn::IWorkloadFactory& workloadFactory) |
| 4115 | { |
| 4116 | const unsigned int shape0[] = { 1, 1, 2, 2 }; |
| 4117 | const unsigned int shape1[] = { 1, 1, 2, 1 }; |
| 4118 | |
| 4119 | std::vector<uint8_t> input0({ 10, 12, 14, 16 }); |
| 4120 | std::vector<uint8_t> input1({ 2, 1 }); |
| 4121 | std::vector<uint8_t> output({ 8, 11, 12, 15 }); |
| 4122 | |
| 4123 | return SubtractionTestHelper(workloadFactory, |
| 4124 | shape0, input0, 1.0f, 0, |
| 4125 | shape1, input1, 1.0f, 0, |
| 4126 | shape0, output, 1.0f, 0); |
| 4127 | } |
| 4128 | |
| 4129 | LayerTestResult<float, 4> SubtractionTest(armnn::IWorkloadFactory& workloadFactory) |
| 4130 | { |
| 4131 | const unsigned int shape0[] = { 1, 1, 2, 2 }; |
| 4132 | const unsigned int shape1[] = { 1, 1, 2, 2 }; |
| 4133 | |
| 4134 | std::vector<float> input0({ 1, 2, 3, 4 }); |
| 4135 | std::vector<float> input1({ 1, -1, 0, 2 }); |
| 4136 | std::vector<float> output({ 0, 3, 3, 2 }); |
| 4137 | |
| 4138 | return SubtractionTestHelper(workloadFactory, |
| 4139 | shape0, input0, 1.0f, 0, |
| 4140 | shape1, input1, 1.0f, 0, |
| 4141 | shape0, output, 1.0f, 0); |
| 4142 | } |
| 4143 | |
| 4144 | LayerTestResult<float, 4> SubtractionBroadcast1ElementTest(armnn::IWorkloadFactory& workloadFactory) |
| 4145 | { |
| 4146 | const unsigned int shape0[] = { 1, 1, 2, 2 }; |
| 4147 | const unsigned int shape1[] = { 1, 1, 1, 1 }; |
| 4148 | |
| 4149 | std::vector<float> input0({ 1, 2, 3, 4 }); |
| 4150 | std::vector<float> input1({ 10 }); |
| 4151 | std::vector<float> output({ -9, -8, -7, -6 }); |
| 4152 | |
| 4153 | return SubtractionTestHelper(workloadFactory, |
| 4154 | shape0, input0, 1.0f, 0, |
| 4155 | shape1, input1, 1.0f, 0, |
| 4156 | shape0, output, 1.0f, 0); |
| 4157 | } |
| 4158 | |
| 4159 | LayerTestResult<float, 4> SubtractionBroadcastTest(armnn::IWorkloadFactory& workloadFactory) |
| 4160 | { |
| 4161 | const unsigned int shape0[] = { 1, 1, 2, 2 }; |
| 4162 | const unsigned int shape1[] = { 1, 1, 1, 2 }; |
| 4163 | |
| 4164 | std::vector<float> input0({ 1, 2, 3, 4 }); |
| 4165 | std::vector<float> input1({ 10, -5 }); |
| 4166 | std::vector<float> output({ -9, 7, -7, 9 }); |
| 4167 | |
| 4168 | return SubtractionTestHelper(workloadFactory, |
| 4169 | shape0, input0, 1.0f, 0, |
| 4170 | shape1, input1, 1.0f, 0, |
| 4171 | shape0, output, 1.0f, 0); |
| 4172 | } |
| 4173 | |
telsoa01 | 4fcda01 | 2018-03-09 14:13:49 +0000 | [diff] [blame] | 4174 | LayerTestResult<uint8_t, 4> ResizeBilinearNopUint8Test(armnn::IWorkloadFactory& workloadFactory) |
| 4175 | { |
| 4176 | constexpr unsigned int inputWidth = 4; |
| 4177 | constexpr unsigned int inputHeight = 4; |
| 4178 | constexpr unsigned int inputChannels = 1; |
| 4179 | constexpr unsigned int inputBatchSize = 1; |
| 4180 | |
| 4181 | constexpr unsigned int outputWidth = inputWidth; |
| 4182 | constexpr unsigned int outputHeight = inputHeight; |
| 4183 | constexpr unsigned int outputChannels = inputChannels; |
| 4184 | constexpr unsigned int outputBatchSize = inputBatchSize; |
| 4185 | |
| 4186 | armnn::TensorInfo inputTensorInfo({ inputBatchSize, inputChannels, inputHeight, inputWidth }, |
| 4187 | armnn::DataType::QuantisedAsymm8); |
| 4188 | inputTensorInfo.SetQuantizationScale(1.5f); |
| 4189 | inputTensorInfo.SetQuantizationOffset(-3); |
| 4190 | |
| 4191 | armnn::TensorInfo outputTensorInfo({ outputBatchSize, outputChannels, outputHeight, outputWidth }, |
| 4192 | armnn::DataType::QuantisedAsymm8); |
| 4193 | outputTensorInfo.SetQuantizationScale(1.5f); |
| 4194 | outputTensorInfo.SetQuantizationOffset(-3); |
| 4195 | |
| 4196 | auto input = MakeTensor<uint8_t, 4>(inputTensorInfo, std::vector<uint8_t>({ |
| 4197 | 1, 2, 3, 4, |
| 4198 | 2, 3, 4, 5, |
| 4199 | 3, 4, 5, 6, |
| 4200 | 4, 5, 6, 7 |
| 4201 | })); |
| 4202 | |
| 4203 | LayerTestResult<uint8_t, 4> result(outputTensorInfo); |
| 4204 | result.outputExpected = input; |
| 4205 | |
| 4206 | std::unique_ptr<armnn::ITensorHandle> inputHandle = workloadFactory.CreateTensorHandle(inputTensorInfo); |
| 4207 | std::unique_ptr<armnn::ITensorHandle> outputHandle = workloadFactory.CreateTensorHandle(outputTensorInfo); |
| 4208 | |
| 4209 | armnn::ResizeBilinearQueueDescriptor descriptor; |
| 4210 | armnn::WorkloadInfo info; |
| 4211 | AddInputToWorkload(descriptor, info, inputTensorInfo, inputHandle.get()); |
| 4212 | AddOutputToWorkload(descriptor, info, outputTensorInfo, outputHandle.get()); |
| 4213 | |
| 4214 | std::unique_ptr<armnn::IWorkload> workload = workloadFactory.CreateResizeBilinear(descriptor, info); |
| 4215 | |
| 4216 | inputHandle->Allocate(); |
| 4217 | outputHandle->Allocate(); |
| 4218 | CopyDataToITensorHandle(inputHandle.get(), &input[0][0][0][0]); |
| 4219 | |
surmeh01 | 3537c2c | 2018-05-18 16:31:43 +0100 | [diff] [blame] | 4220 | workloadFactory.Finalize(); |
telsoa01 | 4fcda01 | 2018-03-09 14:13:49 +0000 | [diff] [blame] | 4221 | workload->Execute(); |
| 4222 | |
| 4223 | CopyDataFromITensorHandle(&result.output[0][0][0][0], outputHandle.get()); |
| 4224 | return result; |
| 4225 | } |
| 4226 | |
| 4227 | LayerTestResult<uint8_t, 4> SimpleResizeBilinearUint8Test(armnn::IWorkloadFactory& workloadFactory) |
| 4228 | { |
| 4229 | constexpr unsigned int inputWidth = 2; |
| 4230 | constexpr unsigned int inputHeight = 2; |
| 4231 | constexpr unsigned int inputChannels = 1; |
| 4232 | constexpr unsigned int inputBatchSize = 1; |
| 4233 | |
| 4234 | constexpr unsigned int outputWidth = inputWidth / 2; |
| 4235 | constexpr unsigned int outputHeight = inputHeight / 2; |
| 4236 | constexpr unsigned int outputChannels = inputChannels; |
| 4237 | constexpr unsigned int outputBatchSize = inputBatchSize; |
| 4238 | |
| 4239 | armnn::TensorInfo inputTensorInfo({ inputBatchSize, inputChannels, inputHeight, inputWidth }, |
| 4240 | armnn::DataType::QuantisedAsymm8); |
| 4241 | inputTensorInfo.SetQuantizationScale(0.1567f); |
| 4242 | inputTensorInfo.SetQuantizationOffset(1); |
| 4243 | |
| 4244 | armnn::TensorInfo outputTensorInfo({ outputBatchSize, outputChannels, outputHeight, outputWidth }, |
| 4245 | armnn::DataType::QuantisedAsymm8); |
| 4246 | outputTensorInfo.SetQuantizationScale(0.1567f); |
| 4247 | outputTensorInfo.SetQuantizationOffset(1); |
| 4248 | |
| 4249 | auto input = MakeTensor<uint8_t, 4>(inputTensorInfo, std::vector<uint8_t>({ |
| 4250 | 1, 255, |
| 4251 | 200, 250 |
| 4252 | })); |
| 4253 | |
| 4254 | // The 'resize bilinear' operation projects the top-left corner of output texels into the input image, |
| 4255 | // then figures out the interpolants and weights. Note this is different to projecting the centre of the |
telsoa01 | c577f2c | 2018-08-31 09:22:23 +0100 | [diff] [blame] | 4256 | // output texel - and thus we'll expect the output 1x1 matrix to contain, as its single element, the value |
telsoa01 | 4fcda01 | 2018-03-09 14:13:49 +0000 | [diff] [blame] | 4257 | // that was at position (0,0) of the input matrix (rather than an average, which we would expect if projecting |
| 4258 | // the centre). |
| 4259 | LayerTestResult<uint8_t, 4> result(outputTensorInfo); |
| 4260 | result.outputExpected = MakeTensor<uint8_t, 4>(outputTensorInfo, std::vector<uint8_t>({ |
| 4261 | 1 |
| 4262 | })); |
| 4263 | |
| 4264 | std::unique_ptr<armnn::ITensorHandle> inputHandle = workloadFactory.CreateTensorHandle(inputTensorInfo); |
| 4265 | std::unique_ptr<armnn::ITensorHandle> outputHandle = workloadFactory.CreateTensorHandle(outputTensorInfo); |
| 4266 | |
| 4267 | armnn::ResizeBilinearQueueDescriptor descriptor; |
| 4268 | armnn::WorkloadInfo info; |
| 4269 | AddInputToWorkload(descriptor, info, inputTensorInfo, inputHandle.get()); |
| 4270 | AddOutputToWorkload(descriptor, info, outputTensorInfo, outputHandle.get()); |
| 4271 | |
| 4272 | std::unique_ptr<armnn::IWorkload> workload = workloadFactory.CreateResizeBilinear(descriptor, info); |
| 4273 | |
| 4274 | inputHandle->Allocate(); |
| 4275 | outputHandle->Allocate(); |
| 4276 | CopyDataToITensorHandle(inputHandle.get(), &input[0][0][0][0]); |
| 4277 | |
surmeh01 | 3537c2c | 2018-05-18 16:31:43 +0100 | [diff] [blame] | 4278 | workloadFactory.Finalize(); |
telsoa01 | 4fcda01 | 2018-03-09 14:13:49 +0000 | [diff] [blame] | 4279 | workload->Execute(); |
| 4280 | |
| 4281 | CopyDataFromITensorHandle(&result.output[0][0][0][0], outputHandle.get()); |
| 4282 | return result; |
| 4283 | } |
| 4284 | |
| 4285 | LayerTestResult<uint8_t, 4> ResizeBilinearSqMinUint8Test(armnn::IWorkloadFactory& workloadFactory) |
| 4286 | { |
| 4287 | constexpr unsigned int inputWidth = 4; |
| 4288 | constexpr unsigned int inputHeight = 4; |
| 4289 | constexpr unsigned int inputChannels = 1; |
| 4290 | constexpr unsigned int inputBatchSize = 1; |
| 4291 | |
| 4292 | constexpr unsigned int outputWidth = inputWidth / 2; |
| 4293 | constexpr unsigned int outputHeight = inputHeight / 2; |
| 4294 | constexpr unsigned int outputChannels = inputChannels; |
| 4295 | constexpr unsigned int outputBatchSize = inputBatchSize; |
| 4296 | |
| 4297 | armnn::TensorInfo inputTensorInfo({ inputBatchSize, inputChannels, inputHeight, inputWidth }, |
| 4298 | armnn::DataType::QuantisedAsymm8); |
| 4299 | inputTensorInfo.SetQuantizationScale(3.141592f); |
| 4300 | inputTensorInfo.SetQuantizationOffset(3); |
| 4301 | |
| 4302 | armnn::TensorInfo outputTensorInfo({ outputBatchSize, outputChannels, outputHeight, outputWidth }, |
| 4303 | armnn::DataType::QuantisedAsymm8); |
| 4304 | outputTensorInfo.SetQuantizationScale(3.141592f); |
| 4305 | outputTensorInfo.SetQuantizationOffset(3); |
| 4306 | |
| 4307 | auto input = MakeTensor<uint8_t, 4>(inputTensorInfo, std::vector<uint8_t>({ |
| 4308 | 1, 2, 3, 4, |
| 4309 | 2, 3, 4, 5, |
| 4310 | 3, 4, 5, 6, |
| 4311 | 4, 5, 6, 7 |
| 4312 | })); |
| 4313 | |
| 4314 | LayerTestResult<uint8_t, 4> result(outputTensorInfo); |
| 4315 | result.outputExpected = MakeTensor<uint8_t, 4>(outputTensorInfo, std::vector<uint8_t>({ |
| 4316 | 1, 3, |
| 4317 | 3, 5 |
| 4318 | })); |
| 4319 | |
| 4320 | std::unique_ptr<armnn::ITensorHandle> inputHandle = workloadFactory.CreateTensorHandle(inputTensorInfo); |
| 4321 | std::unique_ptr<armnn::ITensorHandle> outputHandle = workloadFactory.CreateTensorHandle(outputTensorInfo); |
| 4322 | |
| 4323 | armnn::ResizeBilinearQueueDescriptor descriptor; |
| 4324 | armnn::WorkloadInfo info; |
| 4325 | AddInputToWorkload(descriptor, info, inputTensorInfo, inputHandle.get()); |
| 4326 | AddOutputToWorkload(descriptor, info, outputTensorInfo, outputHandle.get()); |
| 4327 | |
| 4328 | std::unique_ptr<armnn::IWorkload> workload = workloadFactory.CreateResizeBilinear(descriptor, info); |
| 4329 | |
| 4330 | inputHandle->Allocate(); |
| 4331 | outputHandle->Allocate(); |
| 4332 | CopyDataToITensorHandle(inputHandle.get(), &input[0][0][0][0]); |
| 4333 | |
surmeh01 | 3537c2c | 2018-05-18 16:31:43 +0100 | [diff] [blame] | 4334 | workloadFactory.Finalize(); |
telsoa01 | 4fcda01 | 2018-03-09 14:13:49 +0000 | [diff] [blame] | 4335 | workload->Execute(); |
| 4336 | |
| 4337 | CopyDataFromITensorHandle(&result.output[0][0][0][0], outputHandle.get()); |
| 4338 | return result; |
| 4339 | } |
| 4340 | |
| 4341 | LayerTestResult<uint8_t, 4> ResizeBilinearMinUint8Test(armnn::IWorkloadFactory& workloadFactory) |
| 4342 | { |
| 4343 | constexpr unsigned int inputWidth = 3; |
| 4344 | constexpr unsigned int inputHeight = 2; |
| 4345 | constexpr unsigned int inputChannels = 1; |
| 4346 | constexpr unsigned int inputBatchSize = 1; |
| 4347 | |
| 4348 | constexpr unsigned int outputWidth = 2; |
| 4349 | constexpr unsigned int outputHeight = 1; |
| 4350 | constexpr unsigned int outputChannels = inputChannels; |
| 4351 | constexpr unsigned int outputBatchSize = inputBatchSize; |
| 4352 | |
| 4353 | armnn::TensorInfo inputTensorInfo({ inputBatchSize, inputChannels, inputHeight, inputWidth }, |
| 4354 | armnn::DataType::QuantisedAsymm8); |
| 4355 | inputTensorInfo.SetQuantizationScale(1.5f); |
| 4356 | inputTensorInfo.SetQuantizationOffset(-1); |
| 4357 | |
| 4358 | armnn::TensorInfo outputTensorInfo({ outputBatchSize, outputChannels, outputHeight, outputWidth }, |
| 4359 | armnn::DataType::QuantisedAsymm8); |
| 4360 | outputTensorInfo.SetQuantizationScale(1.5f); |
| 4361 | outputTensorInfo.SetQuantizationOffset(-1); |
| 4362 | |
| 4363 | auto input = MakeTensor<uint8_t, 4>(inputTensorInfo, std::vector<uint8_t>({ |
| 4364 | 1, 2, 3, // 3.0, 4.5, 6.0 |
| 4365 | 5, 8, 13 // 9.0, 13.5, 21.0 |
| 4366 | })); |
| 4367 | |
| 4368 | LayerTestResult<uint8_t, 4> result(outputTensorInfo); |
| 4369 | result.outputExpected = MakeTensor<uint8_t, 4>(outputTensorInfo, std::vector<uint8_t>({ |
| 4370 | 1, 3 // 3.0, 5.25 |
| 4371 | })); |
| 4372 | |
| 4373 | std::unique_ptr<armnn::ITensorHandle> inputHandle = workloadFactory.CreateTensorHandle(inputTensorInfo); |
| 4374 | std::unique_ptr<armnn::ITensorHandle> outputHandle = workloadFactory.CreateTensorHandle(outputTensorInfo); |
| 4375 | |
| 4376 | armnn::ResizeBilinearQueueDescriptor descriptor; |
| 4377 | armnn::WorkloadInfo info; |
| 4378 | AddInputToWorkload(descriptor, info, inputTensorInfo, inputHandle.get()); |
| 4379 | AddOutputToWorkload(descriptor, info, outputTensorInfo, outputHandle.get()); |
| 4380 | |
| 4381 | std::unique_ptr<armnn::IWorkload> workload = workloadFactory.CreateResizeBilinear(descriptor, info); |
| 4382 | |
| 4383 | inputHandle->Allocate(); |
| 4384 | outputHandle->Allocate(); |
| 4385 | |
| 4386 | CopyDataToITensorHandle(inputHandle.get(), &input[0][0][0][0]); |
| 4387 | |
surmeh01 | 3537c2c | 2018-05-18 16:31:43 +0100 | [diff] [blame] | 4388 | workloadFactory.Finalize(); |
telsoa01 | 4fcda01 | 2018-03-09 14:13:49 +0000 | [diff] [blame] | 4389 | workload->Execute(); |
| 4390 | |
| 4391 | CopyDataFromITensorHandle(&result.output[0][0][0][0], outputHandle.get()); |
| 4392 | return result; |
| 4393 | } |
| 4394 | |
| 4395 | LayerTestResult<uint8_t, 4> ResizeBilinearMagUint8Test(armnn::IWorkloadFactory& workloadFactory) |
| 4396 | { |
| 4397 | constexpr unsigned int inputWidth = 2; |
| 4398 | constexpr unsigned int inputHeight = 3; |
| 4399 | constexpr unsigned int inputChannels = 1; |
| 4400 | constexpr unsigned int inputBatchSize = 1; |
| 4401 | |
| 4402 | constexpr unsigned int outputWidth = 5; |
| 4403 | constexpr unsigned int outputHeight = 3; |
| 4404 | constexpr unsigned int outputChannels = inputChannels; |
| 4405 | constexpr unsigned int outputBatchSize = inputBatchSize; |
| 4406 | |
| 4407 | armnn::TensorInfo inputTensorInfo({ inputBatchSize, inputChannels, inputHeight, inputWidth }, |
| 4408 | armnn::DataType::QuantisedAsymm8); |
| 4409 | inputTensorInfo.SetQuantizationScale(0.010765f); |
| 4410 | inputTensorInfo.SetQuantizationOffset(7); |
| 4411 | |
| 4412 | armnn::TensorInfo outputTensorInfo({ outputBatchSize, outputChannels, outputHeight, outputWidth }, |
| 4413 | armnn::DataType::QuantisedAsymm8); |
| 4414 | outputTensorInfo.SetQuantizationScale(0.010132f); |
| 4415 | outputTensorInfo.SetQuantizationOffset(-18); |
| 4416 | |
| 4417 | auto input = MakeTensor<uint8_t, 4>(inputTensorInfo, std::vector<uint8_t>({ |
| 4418 | 24, 228, // 0.183005, 2.379065, |
| 4419 | 105, 128, // 1.05497, 1.302565 |
| 4420 | 230, 71 // 2.400595, 0.68896 |
| 4421 | })); |
| 4422 | |
| 4423 | LayerTestResult<uint8_t, 4> result(outputTensorInfo); |
| 4424 | result.outputExpected = MakeTensor<uint8_t, 4>(outputTensorInfo, std::vector<uint8_t>({ |
| 4425 | 0, 87, 173, 217, 217, // 0.18300501, 1.06142902, 1.93985295, 2.37906504, 2.37906504 |
| 4426 | 86, 96, 106, 111, 111, // 1.05497003, 1.15400803, 1.25304604, 1.30256498, 1.30256498 |
| 4427 | 219, 151, 84, 50, 50 // 2.40059495, 1.71594095, 1.03128707, 0.68896002, 0.68896002 |
| 4428 | })); |
| 4429 | |
| 4430 | std::unique_ptr<armnn::ITensorHandle> inputHandle = workloadFactory.CreateTensorHandle(inputTensorInfo); |
| 4431 | std::unique_ptr<armnn::ITensorHandle> outputHandle = workloadFactory.CreateTensorHandle(outputTensorInfo); |
| 4432 | |
| 4433 | armnn::ResizeBilinearQueueDescriptor descriptor; |
| 4434 | armnn::WorkloadInfo info; |
| 4435 | AddInputToWorkload(descriptor, info, inputTensorInfo, inputHandle.get()); |
| 4436 | AddOutputToWorkload(descriptor, info, outputTensorInfo, outputHandle.get()); |
| 4437 | |
| 4438 | std::unique_ptr<armnn::IWorkload> workload = workloadFactory.CreateResizeBilinear(descriptor, info); |
| 4439 | |
| 4440 | inputHandle->Allocate(); |
| 4441 | outputHandle->Allocate(); |
| 4442 | CopyDataToITensorHandle(inputHandle.get(), &input[0][0][0][0]); |
| 4443 | |
surmeh01 | 3537c2c | 2018-05-18 16:31:43 +0100 | [diff] [blame] | 4444 | workloadFactory.Finalize(); |
telsoa01 | 4fcda01 | 2018-03-09 14:13:49 +0000 | [diff] [blame] | 4445 | workload->Execute(); |
| 4446 | |
| 4447 | CopyDataFromITensorHandle(&result.output[0][0][0][0], outputHandle.get()); |
| 4448 | return result; |
| 4449 | } |
| 4450 | |
| 4451 | LayerTestResult<float, 4> BatchNormTest(armnn::IWorkloadFactory& workloadFactory) |
| 4452 | { |
| 4453 | auto ret = BatchNormTestImpl<float>(workloadFactory, 0.f, 0); |
| 4454 | return ret; |
| 4455 | } |
| 4456 | |
| 4457 | LayerTestResult<uint8_t, 4> BatchNormUint8Test(armnn::IWorkloadFactory& workloadFactory) |
| 4458 | { |
| 4459 | auto ret = BatchNormTestImpl<uint8_t>(workloadFactory, 1.f/20.f, 50); |
| 4460 | return ret; |
| 4461 | } |
| 4462 | |
| 4463 | LayerTestResult<uint8_t, 4> ConstantUint8Test(armnn::IWorkloadFactory& workloadFactory) |
| 4464 | { |
| 4465 | return ConstantTestImpl<uint8_t>(workloadFactory, 2e-6f, 1); |
| 4466 | } |
| 4467 | |
| 4468 | LayerTestResult<uint8_t, 1> Concatenation1dUint8Test(armnn::IWorkloadFactory& workloadFactory) |
| 4469 | { |
| 4470 | return Concatenation1dTestImpl<uint8_t>(workloadFactory, 0.5f, -1); |
| 4471 | } |
| 4472 | |
| 4473 | LayerTestResult<uint8_t, 2> Concatenation2dDim0Uint8Test(armnn::IWorkloadFactory& workloadFactory) |
| 4474 | { |
| 4475 | return Concatenation2dDim0TestImpl<uint8_t>(workloadFactory, 0.5f, -1); |
| 4476 | } |
| 4477 | |
| 4478 | LayerTestResult<uint8_t, 2> Concatenation2dDim1Uint8Test(armnn::IWorkloadFactory& workloadFactory) |
| 4479 | { |
| 4480 | return Concatenation2dDim1TestImpl<uint8_t>(workloadFactory, 0.5f, -1); |
| 4481 | } |
| 4482 | |
| 4483 | LayerTestResult<uint8_t, 2> Concatenation2dDim0DiffInputDimsUint8Test(armnn::IWorkloadFactory& workloadFactory) |
| 4484 | { |
| 4485 | return Concatenation2dDim0DiffInputDimsTestImpl<uint8_t>(workloadFactory, 0.5f, -1); |
| 4486 | } |
| 4487 | |
| 4488 | LayerTestResult<uint8_t, 2> Concatenation2dDim1DiffInputDimsUint8Test(armnn::IWorkloadFactory& workloadFactory) |
| 4489 | { |
| 4490 | return Concatenation2dDim1DiffInputDimsTestImpl<uint8_t>(workloadFactory, 0.5f, -1); |
| 4491 | } |
| 4492 | |
| 4493 | LayerTestResult<uint8_t, 3> Concatenation3dDim0Uint8Test(armnn::IWorkloadFactory& workloadFactory) |
| 4494 | { |
| 4495 | return Concatenation3dDim0TestImpl<uint8_t>(workloadFactory, 0.5f, -1); |
| 4496 | } |
| 4497 | |
| 4498 | LayerTestResult<uint8_t, 3> Concatenation3dDim1Uint8Test(armnn::IWorkloadFactory& workloadFactory) |
| 4499 | { |
| 4500 | return Concatenation3dDim1TestImpl<uint8_t>(workloadFactory, 0.5f, -1); |
| 4501 | } |
| 4502 | |
| 4503 | LayerTestResult<uint8_t, 3> Concatenation3dDim2Uint8Test(armnn::IWorkloadFactory& workloadFactory) |
| 4504 | { |
| 4505 | return Concatenation3dDim2TestImpl<uint8_t>(workloadFactory, 0.5f, -1); |
| 4506 | } |
| 4507 | |
| 4508 | LayerTestResult<uint8_t, 3> Concatenation3dDim0DiffInputDimsUint8Test(armnn::IWorkloadFactory& workloadFactory) |
| 4509 | { |
| 4510 | return Concatenation3dDim0TestImpl<uint8_t>(workloadFactory, 0.5f, -1); |
| 4511 | } |
| 4512 | |
| 4513 | LayerTestResult<uint8_t, 3> Concatenation3dDim1DiffInputDimsUint8Test(armnn::IWorkloadFactory& workloadFactory) |
| 4514 | { |
| 4515 | return Concatenation3dDim1DiffInputDimsTestImpl<uint8_t>(workloadFactory, 0.5f, -1); |
| 4516 | } |
| 4517 | |
| 4518 | LayerTestResult<uint8_t, 3> Concatenation3dDim2DiffInputDimsUint8Test(armnn::IWorkloadFactory& workloadFactory) |
| 4519 | { |
| 4520 | return Concatenation3dDim2DiffInputDimsTestImpl<uint8_t>(workloadFactory, 0.5f, -1); |
| 4521 | } |
| 4522 | |
| 4523 | LayerTestResult<float, 4> SimpleMaxPooling2dSize2x2Stride2x2Test(armnn::IWorkloadFactory& workloadFactory, |
| 4524 | bool forceNoPadding) |
| 4525 | { |
| 4526 | return SimpleMaxPooling2dSize2x2Stride2x2TestCommon<float>(workloadFactory, forceNoPadding); |
| 4527 | } |
| 4528 | |
| 4529 | LayerTestResult<uint8_t, 4> SimpleMaxPooling2dSize2x2Stride2x2Uint8Test(armnn::IWorkloadFactory& workloadFactory, |
| 4530 | bool forceNoPadding) |
| 4531 | { |
| 4532 | return SimpleMaxPooling2dSize2x2Stride2x2TestCommon<uint8_t>(workloadFactory, forceNoPadding, 3.0f, -5); |
| 4533 | } |
| 4534 | |
| 4535 | LayerTestResult<float, 4> SimpleMaxPooling2dSize3x3Stride2x4Test(armnn::IWorkloadFactory& workloadFactory, |
| 4536 | bool forceNoPadding) |
| 4537 | { |
| 4538 | return SimpleMaxPooling2dSize3x3Stride2x4TestCommon<float>(workloadFactory, forceNoPadding); |
| 4539 | } |
| 4540 | |
| 4541 | LayerTestResult<uint8_t, 4> SimpleMaxPooling2dSize3x3Stride2x4Uint8Test(armnn::IWorkloadFactory& workloadFactory, |
| 4542 | bool forceNoPadding) |
| 4543 | { |
| 4544 | return SimpleMaxPooling2dSize3x3Stride2x4TestCommon<uint8_t>(workloadFactory, forceNoPadding, 0.1f, 128); |
| 4545 | } |
| 4546 | |
| 4547 | LayerTestResult<float, 4> SimpleAveragePooling2dTest(armnn::IWorkloadFactory& workloadFactory) |
| 4548 | { |
| 4549 | return SimpleAveragePooling2dTestCommon<float>(workloadFactory); |
| 4550 | } |
| 4551 | |
| 4552 | LayerTestResult<uint8_t, 4> SimpleAveragePooling2dUint8Test(armnn::IWorkloadFactory& workloadFactory) |
| 4553 | { |
| 4554 | return SimpleAveragePooling2dTestCommon<uint8_t>(workloadFactory, 0.5, -1); |
| 4555 | } |
| 4556 | |
surmeh01 | bceff2f | 2018-03-29 16:29:27 +0100 | [diff] [blame] | 4557 | LayerTestResult<float, 4> IgnorePaddingAveragePooling2dSize3x2Stride2x2Test(armnn::IWorkloadFactory& workloadFactory, |
| 4558 | bool forceNoPadding) |
| 4559 | { |
| 4560 | return IgnorePaddingAveragePooling2dSize3x2Stride2x2TestCommon<float>(workloadFactory, forceNoPadding); |
| 4561 | } |
| 4562 | |
telsoa01 | 4fcda01 | 2018-03-09 14:13:49 +0000 | [diff] [blame] | 4563 | LayerTestResult<float, 4> LargeTensorsAveragePooling2dTest(armnn::IWorkloadFactory& workloadFactory) |
| 4564 | { |
| 4565 | return LargeTensorsAveragePooling2dTestCommon<float>(workloadFactory); |
| 4566 | } |
| 4567 | |
| 4568 | LayerTestResult<uint8_t, 4> LargeTensorsAveragePooling2dUint8Test(armnn::IWorkloadFactory& workloadFactory) |
| 4569 | { |
| 4570 | return LargeTensorsAveragePooling2dTestCommon<uint8_t>(workloadFactory, 0.5, -1); |
| 4571 | } |
| 4572 | |
| 4573 | LayerTestResult<float, 4> SimpleL2Pooling2dTest(armnn::IWorkloadFactory& workloadFactory) |
| 4574 | { |
| 4575 | return SimpleL2Pooling2dTestCommon<float>(workloadFactory); |
| 4576 | } |
| 4577 | |
| 4578 | LayerTestResult<uint8_t, 4> SimpleL2Pooling2dUint8Test(armnn::IWorkloadFactory& workloadFactory) |
| 4579 | { |
| 4580 | return SimpleL2Pooling2dTestCommon<uint8_t>(workloadFactory); |
| 4581 | } |
| 4582 | |
| 4583 | LayerTestResult<float, 4> L2Pooling2dSize3Stride1Test(armnn::IWorkloadFactory& workloadFactory) |
| 4584 | { |
| 4585 | return L2Pooling2dSize3Stride1TestCommon<float>(workloadFactory); |
| 4586 | } |
| 4587 | |
| 4588 | LayerTestResult<uint8_t, 4> L2Pooling2dSize3Stride1Uint8Test(armnn::IWorkloadFactory& workloadFactory) |
| 4589 | { |
| 4590 | return L2Pooling2dSize3Stride1TestCommon<uint8_t>(workloadFactory); |
| 4591 | } |
| 4592 | |
| 4593 | LayerTestResult<float, 4> L2Pooling2dSize3Stride3Test(armnn::IWorkloadFactory& workloadFactory) |
| 4594 | { |
| 4595 | return L2Pooling2dSize3Stride3TestCommon<float>(workloadFactory); |
| 4596 | } |
| 4597 | |
| 4598 | LayerTestResult<uint8_t, 4> L2Pooling2dSize3Stride3Uint8Test(armnn::IWorkloadFactory& workloadFactory) |
| 4599 | { |
| 4600 | return L2Pooling2dSize3Stride3TestCommon<uint8_t>(workloadFactory); |
| 4601 | } |
| 4602 | |
| 4603 | LayerTestResult<float, 4> L2Pooling2dSize3Stride4Test(armnn::IWorkloadFactory& workloadFactory) |
| 4604 | { |
| 4605 | return L2Pooling2dSize3Stride4TestCommon<float>(workloadFactory); |
| 4606 | } |
| 4607 | |
| 4608 | LayerTestResult<uint8_t, 4> L2Pooling2dSize3Stride4Uint8Test(armnn::IWorkloadFactory& workloadFactory) |
| 4609 | { |
| 4610 | return L2Pooling2dSize3Stride4TestCommon<uint8_t>(workloadFactory); |
| 4611 | } |
| 4612 | |
| 4613 | LayerTestResult<float, 4> L2Pooling2dSize7Test(armnn::IWorkloadFactory& workloadFactory) |
| 4614 | { |
| 4615 | return L2Pooling2dSize7TestCommon<float>(workloadFactory); |
| 4616 | } |
| 4617 | |
| 4618 | LayerTestResult<uint8_t, 4> L2Pooling2dSize7Uint8Test(armnn::IWorkloadFactory& workloadFactory) |
| 4619 | { |
| 4620 | return L2Pooling2dSize7TestCommon<uint8_t>(workloadFactory); |
| 4621 | } |
| 4622 | |
| 4623 | LayerTestResult<float, 4> L2Pooling2dSize9Test(armnn::IWorkloadFactory& workloadFactory) |
| 4624 | { |
| 4625 | return L2Pooling2dSize9TestCommon<float>(workloadFactory); |
| 4626 | } |
| 4627 | |
| 4628 | LayerTestResult<uint8_t, 4> L2Pooling2dSize9Uint8Test(armnn::IWorkloadFactory& workloadFactory) |
| 4629 | { |
| 4630 | return L2Pooling2dSize9TestCommon<uint8_t>(workloadFactory); |
| 4631 | } |
| 4632 | |
| 4633 | LayerTestResult<float, 4> AsymmetricNonSquarePooling2dTest(armnn::IWorkloadFactory& workloadFactory) |
| 4634 | { |
| 4635 | return AsymmetricNonSquarePooling2dTestCommon<float>(workloadFactory); |
| 4636 | } |
| 4637 | |
| 4638 | LayerTestResult<uint8_t, 4> AsymmetricNonSquarePooling2dUint8Test(armnn::IWorkloadFactory& workloadFactory) |
| 4639 | { |
| 4640 | return AsymmetricNonSquarePooling2dTestCommon<uint8_t>(workloadFactory); |
| 4641 | } |
| 4642 | |
| 4643 | LayerTestResult<float, 4> ComparePooling2dTest(armnn::IWorkloadFactory& workloadFactory, |
| 4644 | armnn::IWorkloadFactory& refWorkloadFactory, |
| 4645 | armnn::PoolingAlgorithm poolingType) |
| 4646 | { |
| 4647 | return ComparePooling2dTestCommon<float>(workloadFactory, refWorkloadFactory, poolingType); |
| 4648 | } |
| 4649 | |
| 4650 | LayerTestResult<uint8_t, 4> ComparePooling2dUint8Test(armnn::IWorkloadFactory& workloadFactory, |
| 4651 | armnn::IWorkloadFactory& refWorkloadFactory, |
| 4652 | armnn::PoolingAlgorithm poolingType) |
| 4653 | { |
| 4654 | return ComparePooling2dTestCommon<uint8_t>(workloadFactory, refWorkloadFactory, poolingType, 0.1f, 128); |
| 4655 | } |
| 4656 | |
| 4657 | LayerTestResult<float, 2> FullyConnectedLargeTest(armnn::IWorkloadFactory& workloadFactory, |
| 4658 | bool transposeWeights) |
| 4659 | { |
| 4660 | return FullyConnectedLargeTestCommon<float>(workloadFactory, transposeWeights); |
| 4661 | } |
| 4662 | |
| 4663 | LayerTestResult<float, 4> IgnorePaddingSimpleMaxPooling2dTest(armnn::IWorkloadFactory& workloadFactory) |
| 4664 | { |
| 4665 | return IgnorePaddingSimpleMaxPooling2dTestCommon<float>(workloadFactory); |
| 4666 | } |
| 4667 | |
| 4668 | LayerTestResult<uint8_t, 4> IgnorePaddingSimpleMaxPooling2dUint8Test(armnn::IWorkloadFactory& workloadFactory) |
| 4669 | { |
| 4670 | return IgnorePaddingSimpleMaxPooling2dTestCommon<uint8_t>(workloadFactory, 1.0f, -5); |
| 4671 | } |
| 4672 | |
| 4673 | LayerTestResult<float, 4> IgnorePaddingMaxPooling2dSize3Test(armnn::IWorkloadFactory& workloadFactory) |
| 4674 | { |
| 4675 | return IgnorePaddingMaxPooling2dSize3TestCommon<float>(workloadFactory); |
| 4676 | } |
| 4677 | |
| 4678 | LayerTestResult<uint8_t, 4> IgnorePaddingMaxPooling2dSize3Uint8Test(armnn::IWorkloadFactory& workloadFactory) |
| 4679 | { |
| 4680 | return IgnorePaddingMaxPooling2dSize3TestCommon<uint8_t>(workloadFactory, 1.0f, -5); |
| 4681 | } |
| 4682 | |
| 4683 | LayerTestResult<float, 4> IgnorePaddingSimpleAveragePooling2dTest(armnn::IWorkloadFactory& workloadFactory) |
| 4684 | { |
| 4685 | return IgnorePaddingSimpleAveragePooling2dTestCommon<float>(workloadFactory); |
| 4686 | } |
| 4687 | |
| 4688 | LayerTestResult<uint8_t, 4> IgnorePaddingSimpleAveragePooling2dUint8Test(armnn::IWorkloadFactory& workloadFactory) |
| 4689 | { |
| 4690 | return IgnorePaddingSimpleAveragePooling2dTestCommon<uint8_t>(workloadFactory); |
| 4691 | } |
| 4692 | |
| 4693 | LayerTestResult<float, 4> IgnorePaddingSimpleAveragePooling2dNoPaddingTest(armnn::IWorkloadFactory& workloadFactory) |
| 4694 | { |
| 4695 | return IgnorePaddingSimpleAveragePooling2dNoPaddingTestCommon<float>(workloadFactory); |
| 4696 | } |
| 4697 | |
| 4698 | LayerTestResult<uint8_t, 4> IgnorePaddingSimpleAveragePooling2dNoPaddingUint8Test( |
| 4699 | armnn::IWorkloadFactory& workloadFactory) |
| 4700 | { |
| 4701 | return IgnorePaddingSimpleAveragePooling2dNoPaddingTestCommon<uint8_t>(workloadFactory); |
| 4702 | } |
| 4703 | |
| 4704 | LayerTestResult<float, 4> IgnorePaddingAveragePooling2dSize3Test(armnn::IWorkloadFactory& workloadFactory) |
| 4705 | { |
| 4706 | return IgnorePaddingAveragePooling2dSize3TestCommon<float>(workloadFactory); |
| 4707 | } |
| 4708 | |
| 4709 | LayerTestResult<uint8_t, 4> IgnorePaddingAveragePooling2dSize3Uint8Test(armnn::IWorkloadFactory& workloadFactory) |
| 4710 | { |
| 4711 | return IgnorePaddingAveragePooling2dSize3TestCommon<uint8_t>(workloadFactory); |
| 4712 | } |
| 4713 | |
| 4714 | LayerTestResult<float, 4> IgnorePaddingSimpleL2Pooling2dTest(armnn::IWorkloadFactory& workloadFactory) |
| 4715 | { |
| 4716 | return IgnorePaddingSimpleL2Pooling2dTestCommon<float>(workloadFactory); |
| 4717 | } |
| 4718 | |
| 4719 | LayerTestResult<uint8_t, 4> IgnorePaddingSimpleL2Pooling2dUint8Test(armnn::IWorkloadFactory& workloadFactory) |
| 4720 | { |
| 4721 | return IgnorePaddingSimpleL2Pooling2dTestCommon<uint8_t>(workloadFactory); |
| 4722 | } |
| 4723 | |
| 4724 | LayerTestResult<float, 4> IgnorePaddingL2Pooling2dSize3Test(armnn::IWorkloadFactory& workloadFactory) |
| 4725 | { |
| 4726 | return IgnorePaddingL2Pooling2dSize3TestCommon<float>(workloadFactory); |
| 4727 | } |
| 4728 | |
| 4729 | LayerTestResult<uint8_t, 4> IgnorePaddingL2Pooling2dSize3Uint8Test(armnn::IWorkloadFactory& workloadFactory) |
| 4730 | { |
| 4731 | return IgnorePaddingL2Pooling2dSize3TestCommon<uint8_t>(workloadFactory); |
| 4732 | } |
| 4733 | |
| 4734 | LayerTestResult<float, 4> SimplePermuteFloat32Test(armnn::IWorkloadFactory& workloadFactory) |
| 4735 | { |
| 4736 | return SimplePermuteFloat32TestCommon(workloadFactory); |
| 4737 | }; |
| 4738 | |
| 4739 | LayerTestResult<uint8_t, 4> SimplePermuteUint8Test(armnn::IWorkloadFactory& workloadFactory) |
| 4740 | { |
| 4741 | return SimplePermuteUint8TestCommon(workloadFactory); |
| 4742 | }; |
surmeh01 | bceff2f | 2018-03-29 16:29:27 +0100 | [diff] [blame] | 4743 | |
| 4744 | LayerTestResult<float, 4> PermuteFloat32ValueSet1Test(armnn::IWorkloadFactory& workloadFactory) |
| 4745 | { |
| 4746 | return PermuteFloat32ValueSet1TestCommon(workloadFactory); |
| 4747 | }; |
| 4748 | |
| 4749 | LayerTestResult<float, 4> PermuteFloat32ValueSet2Test(armnn::IWorkloadFactory& workloadFactory) |
| 4750 | { |
| 4751 | return PermuteFloat32ValueSet2TestCommon(workloadFactory); |
| 4752 | }; |
| 4753 | |
| 4754 | LayerTestResult<float, 4> PermuteFloat32ValueSet3Test(armnn::IWorkloadFactory& workloadFactory) |
| 4755 | { |
| 4756 | return PermuteFloat32ValueSet3TestCommon(workloadFactory); |
narpra01 | 1e4c31d | 2018-09-28 11:07:51 +0100 | [diff] [blame] | 4757 | }; |
| 4758 | |
| 4759 | namespace |
| 4760 | { |
| 4761 | template <typename T, std::size_t InputDim, std::size_t OutputDim> |
| 4762 | LayerTestResult<T, OutputDim> MeanTestHelper(armnn::IWorkloadFactory& workloadFactory, |
| 4763 | const unsigned int* inputShape, |
| 4764 | const std::vector<T>& inputData, |
| 4765 | const std::vector<unsigned int>& axis, |
| 4766 | bool keepDims, |
| 4767 | const unsigned int* outputShape, |
| 4768 | const std::vector<T>& outputData, |
| 4769 | float scale = 1.0f, |
| 4770 | int32_t offset = 0) |
| 4771 | { |
| 4772 | auto dataType = (std::is_same<T, uint8_t>::value ? |
| 4773 | armnn::DataType::QuantisedAsymm8 : |
| 4774 | armnn::DataType::Float32); |
| 4775 | |
| 4776 | armnn::TensorInfo inputTensorInfo(InputDim, inputShape, dataType); |
| 4777 | armnn::TensorInfo outputTensorInfo(OutputDim, outputShape, dataType); |
| 4778 | |
| 4779 | inputTensorInfo.SetQuantizationScale(scale); |
| 4780 | inputTensorInfo.SetQuantizationOffset(offset); |
| 4781 | |
| 4782 | outputTensorInfo.SetQuantizationScale(scale); |
| 4783 | outputTensorInfo.SetQuantizationOffset(offset); |
| 4784 | |
| 4785 | auto input = MakeTensor<T, InputDim>(inputTensorInfo, inputData); |
| 4786 | |
| 4787 | LayerTestResult<T, OutputDim> result(outputTensorInfo); |
| 4788 | result.outputExpected = MakeTensor<T, OutputDim>(outputTensorInfo, outputData); |
| 4789 | |
| 4790 | std::unique_ptr<armnn::ITensorHandle> inputHandle = workloadFactory.CreateTensorHandle(inputTensorInfo); |
| 4791 | std::unique_ptr<armnn::ITensorHandle> outputHandle = workloadFactory.CreateTensorHandle(outputTensorInfo); |
| 4792 | |
| 4793 | armnn::MeanQueueDescriptor data; |
| 4794 | data.m_Parameters.m_Axis = axis; |
| 4795 | data.m_Parameters.m_KeepDims = keepDims; |
| 4796 | armnn::WorkloadInfo info; |
| 4797 | AddInputToWorkload(data, info, inputTensorInfo, inputHandle.get()); |
| 4798 | AddOutputToWorkload(data, info, outputTensorInfo, outputHandle.get()); |
| 4799 | |
| 4800 | std::unique_ptr<armnn::IWorkload> workload = workloadFactory.CreateMean(data, info); |
| 4801 | |
| 4802 | inputHandle->Allocate(); |
| 4803 | outputHandle->Allocate(); |
| 4804 | |
| 4805 | CopyDataToITensorHandle(inputHandle.get(), input.origin()); |
| 4806 | |
| 4807 | workloadFactory.Finalize(); |
| 4808 | workload->Execute(); |
| 4809 | |
| 4810 | CopyDataFromITensorHandle(result.output.origin(), outputHandle.get()); |
| 4811 | |
| 4812 | return result; |
| 4813 | } |
| 4814 | } // anonymous namespace |
| 4815 | |
| 4816 | LayerTestResult<uint8_t, 1> MeanUint8SimpleTest(armnn::IWorkloadFactory& workloadFactory) |
| 4817 | { |
| 4818 | const unsigned int inputShape[] = { 3, 2 }; |
| 4819 | const unsigned int outputShape[] = { 1 }; |
| 4820 | |
| 4821 | std::vector<uint8_t> input({ 1, 1, 2, 2, 3, 3 }); |
| 4822 | std::vector<uint8_t> output({ 2 }); |
| 4823 | |
| 4824 | return MeanTestHelper<uint8_t, 2, 1>(workloadFactory, inputShape, input, {}, false, outputShape, output); |
| 4825 | } |
| 4826 | |
| 4827 | LayerTestResult<uint8_t, 3> MeanUint8SimpleAxisTest(armnn::IWorkloadFactory& workloadFactory) |
| 4828 | { |
| 4829 | const unsigned int inputShape[] = { 1, 1, 3, 2 }; |
| 4830 | const unsigned int outputShape[] = { 1, 1, 2 }; |
| 4831 | |
| 4832 | std::vector<uint8_t> input({ 1, 1, 2, 2, 3, 3 }); |
| 4833 | std::vector<uint8_t> output({ 2, 2 }); |
| 4834 | |
| 4835 | return MeanTestHelper<uint8_t, 4, 3>(workloadFactory, inputShape, input, {2}, false, outputShape, output); |
| 4836 | } |
| 4837 | |
| 4838 | LayerTestResult<uint8_t, 4> MeanUint8KeepDimsTest(armnn::IWorkloadFactory& workloadFactory) |
| 4839 | { |
| 4840 | const unsigned int inputShape[] = { 1, 1, 3, 2 }; |
| 4841 | const unsigned int outputShape[] = { 1, 1, 1, 2 }; |
| 4842 | |
| 4843 | std::vector<uint8_t> input({ 1, 1, 2, 2, 3, 3 }); |
| 4844 | std::vector<uint8_t> output({ 2, 2 }); |
| 4845 | |
| 4846 | return MeanTestHelper<uint8_t, 4, 4>(workloadFactory, inputShape, input, {2}, true, outputShape, output); |
| 4847 | } |
| 4848 | |
| 4849 | LayerTestResult<uint8_t, 4> MeanUint8MultipleDimsTest(armnn::IWorkloadFactory& workloadFactory) |
| 4850 | { |
| 4851 | const unsigned int inputShape[] = { 2, 3, 1, 2 }; |
| 4852 | const unsigned int outputShape[] = { 1, 3, 1, 1 }; |
| 4853 | |
| 4854 | std::vector<uint8_t> input({ 1, 2, 3, 4, 5, 6, 1, 2, 3, 4, 5, 6}); |
| 4855 | std::vector<uint8_t> output({ 1, 3, 5 }); |
| 4856 | |
| 4857 | return MeanTestHelper<uint8_t, 4, 4>(workloadFactory, inputShape, input, {0, 3}, true, outputShape, output); |
| 4858 | } |
| 4859 | |
| 4860 | LayerTestResult<uint8_t, 1> MeanVtsUint8Test(armnn::IWorkloadFactory& workloadFactory) |
| 4861 | { |
| 4862 | const unsigned int inputShape[] = {4, 3, 2}; |
| 4863 | const unsigned int outputShape[] = { 2 }; |
| 4864 | |
| 4865 | std::vector<uint8_t> input({1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24}); |
| 4866 | std::vector<uint8_t> output({12, 13}); |
| 4867 | |
| 4868 | return MeanTestHelper<uint8_t, 3, 1>(workloadFactory, inputShape, input, {0, 1}, false, outputShape, |
| 4869 | output, 0.8f, 5); |
| 4870 | } |
| 4871 | |
| 4872 | LayerTestResult<float, 1> MeanFloatSimpleTest(armnn::IWorkloadFactory& workloadFactory) |
| 4873 | { |
| 4874 | const unsigned int inputShape[] = { 3, 2 }; |
| 4875 | const unsigned int outputShape[] = { 1 }; |
| 4876 | |
| 4877 | std::vector<float> input({ 1., 1., 2., 2., 3., 3. }); |
| 4878 | std::vector<float> output({ 2. }); |
| 4879 | |
| 4880 | return MeanTestHelper<float, 2, 1>(workloadFactory, inputShape, input, {}, false, outputShape, output); |
| 4881 | } |
| 4882 | |
| 4883 | LayerTestResult<float, 3> MeanFloatSimpleAxisTest(armnn::IWorkloadFactory& workloadFactory) |
| 4884 | { |
| 4885 | const unsigned int inputShape[] = { 2, 3, 1, 2 }; |
| 4886 | const unsigned int outputShape[] = { 3, 1, 2 }; |
| 4887 | |
| 4888 | std::vector<float> input({ 1., 2., 3., 4., 5., 6., 1., 2., 3., 4., 5., 6.}); |
| 4889 | std::vector<float> output({ 1., 2., 3., 4., 5., 6. }); |
| 4890 | |
| 4891 | return MeanTestHelper<float, 4, 3>(workloadFactory, inputShape, input, {0}, false, outputShape, output); |
| 4892 | } |
| 4893 | |
| 4894 | LayerTestResult<float, 4> MeanFloatKeepDimsTest(armnn::IWorkloadFactory& workloadFactory) |
| 4895 | { |
| 4896 | const unsigned int inputShape[] = { 1, 1, 3, 2 }; |
| 4897 | const unsigned int outputShape[] = { 1, 1, 1, 2 }; |
| 4898 | |
| 4899 | std::vector<float> input({ 1., 1., 2., 2., 3., 3. }); |
| 4900 | std::vector<float> output({ 2., 2. }); |
| 4901 | |
| 4902 | return MeanTestHelper<float, 4, 4>(workloadFactory, inputShape, input, {2}, true, outputShape, output); |
| 4903 | } |
| 4904 | |
| 4905 | LayerTestResult<float, 4> MeanFloatMultipleDimsTest(armnn::IWorkloadFactory& workloadFactory) |
| 4906 | { |
| 4907 | const unsigned int inputShape[] = { 2, 3, 1, 2 }; |
| 4908 | const unsigned int outputShape[] = { 1, 3, 1, 1 }; |
| 4909 | |
| 4910 | std::vector<float> input({ 1., 2., 3., 4., 5., 6., 1., 2., 3., 4., 5., 6.}); |
| 4911 | std::vector<float> output({ 1.5, 3.5, 5.5 }); |
| 4912 | |
| 4913 | return MeanTestHelper<float, 4, 4>(workloadFactory, inputShape, input, {0, 3}, true, outputShape, output); |
| 4914 | } |
| 4915 | |
| 4916 | LayerTestResult<float, 1> MeanVtsFloat1Test(armnn::IWorkloadFactory& workloadFactory) |
| 4917 | { |
| 4918 | const unsigned int inputShape[] = {4, 3, 2}; |
| 4919 | const unsigned int outputShape[] = { 2 }; |
| 4920 | |
| 4921 | std::vector<float> input({1.0f, 2.0f, 3.0f, 4.0f, 5.0f, 6.0f, 7.0f, 8.0f, 9.0f, 10.0f, 11.0f, 12.0f, 13.0f, 14.0f, |
| 4922 | 15.0f, 16.0f, 17.0f, 18.0f, 19.0f, 20.0f, 21.0f, 22.0f, 23.0f, 24.0f}); |
| 4923 | std::vector<float> output({12.0f, 13.0f}); |
| 4924 | |
| 4925 | return MeanTestHelper<float, 3, 1>(workloadFactory, inputShape, input, {0, 1}, false, outputShape, output); |
| 4926 | } |
| 4927 | |
| 4928 | LayerTestResult<float, 3> MeanVtsFloat2Test(armnn::IWorkloadFactory& workloadFactory) |
| 4929 | { |
| 4930 | const unsigned int inputShape[] = {4, 3, 2}; |
| 4931 | const unsigned int outputShape[] = {1, 3, 1 }; |
| 4932 | |
| 4933 | std::vector<float> input({1.0f, 2.0f, 3.0f, 4.0f, 5.0f, 6.0f, 7.0f, 8.0f, 9.0f, 10.0f, 11.0f, 12.0f, 13.0f, 14.0f, |
| 4934 | 15.0f, 16.0f, 17.0f, 18.0f, 19.0f, 20.0f, 21.0f, 22.0f, 23.0f, 24.0f}); |
| 4935 | std::vector<float> output({10.5f, 12.5f, 14.5f}); |
| 4936 | |
| 4937 | return MeanTestHelper<float, 3, 3>(workloadFactory, inputShape, input, {0, 2}, true, outputShape, output); |
| 4938 | } |