Jan Eilers | 38e05bd | 2019-06-26 13:10:09 +0100 | [diff] [blame] | 1 | // |
| 2 | // Copyright © 2017 Arm Ltd. All rights reserved. |
| 3 | // SPDX-License-Identifier: MIT |
| 4 | // |
| 5 | |
| 6 | //#pragma once |
| 7 | |
| 8 | #include "LstmUtils.hpp" |
| 9 | #include "BaseIterator.hpp" |
Colm Donelan | 0c47974 | 2021-12-10 12:43:54 +0000 | [diff] [blame] | 10 | #include <armnn/backends/TensorHandle.hpp> |
Jan Eilers | 38e05bd | 2019-06-26 13:10:09 +0100 | [diff] [blame] | 11 | |
| 12 | |
| 13 | // Helper functions ported from the Android code base |
| 14 | // Refer to: android/external/tensorflow/tensorflow/contrib/lite/kernels/internal/reference/portable_tensor_utils.cc |
| 15 | |
| 16 | void VectorBatchVectorAdd(armnn::Decoder<float>& vector, |
| 17 | uint32_t vSize, |
| 18 | armnn::Decoder<float>& batchVector, |
| 19 | uint32_t nBatch, |
| 20 | armnn::Encoder<float>& outResult ) |
| 21 | { |
| 22 | for (uint32_t b = 0; b < nBatch; b++) |
| 23 | { |
| 24 | for (uint32_t v = 0; v < vSize; v++) |
| 25 | { |
| 26 | outResult.Set(batchVector.Get() + vector.Get()); |
| 27 | ++outResult; |
| 28 | ++vector; |
| 29 | ++batchVector; |
| 30 | } |
| 31 | vector -= vSize; |
| 32 | } |
| 33 | batchVector -= vSize * nBatch; |
| 34 | outResult -= vSize * nBatch; |
| 35 | } |
| 36 | |
| 37 | |
| 38 | // Layer norm for each batch. |
| 39 | // normalization_epsilon is added to avoid divergence. |
| 40 | void MeanStddevNormalization(armnn::Decoder<float>& input_vector, |
| 41 | armnn::Encoder<float>& output_vector, |
| 42 | uint32_t v_size, |
| 43 | uint32_t n_batch, |
| 44 | float normalization_epsilon) |
| 45 | { |
| 46 | for (uint32_t batch = 0; batch < n_batch; ++batch) { |
| 47 | float sum = 0.0f; |
| 48 | float sum_sq = 0.0f; |
| 49 | for (uint32_t i = 0; i < v_size; ++i) { |
| 50 | sum += input_vector.Get(); |
| 51 | sum_sq += input_vector.Get() * input_vector.Get(); |
| 52 | ++input_vector; |
| 53 | } |
| 54 | input_vector -= v_size; |
| 55 | |
| 56 | const float mean = sum / static_cast<float>(v_size); |
| 57 | float stddev_inv = 0.0f; |
| 58 | const float variance = sum_sq / static_cast<float>(v_size) - mean * mean; |
| 59 | if (variance == 0) { |
| 60 | stddev_inv = 1.0f / std::sqrt(normalization_epsilon); |
| 61 | } else { |
| 62 | stddev_inv = 1.0f / std::sqrt(variance); |
| 63 | } |
| 64 | |
| 65 | for (uint32_t i = 0; i < v_size; ++i) { |
| 66 | output_vector.Set((input_vector.Get() - mean) * stddev_inv); |
| 67 | ++output_vector; |
| 68 | ++input_vector; |
| 69 | } |
| 70 | // Don't reset iterator to handle next batch |
| 71 | } |
| 72 | output_vector -= v_size * n_batch; |
| 73 | input_vector -= v_size * n_batch; |
| 74 | } |
| 75 | |
| 76 | void ZeroVector(armnn::Encoder<float>& vector, |
| 77 | uint32_t vSize) |
| 78 | { |
| 79 | for (uint32_t v = 0; v < vSize; v++) |
| 80 | { |
| 81 | vector.Set(0.0f); |
| 82 | ++vector; |
| 83 | } |
| 84 | vector -= vSize; |
| 85 | } |
| 86 | |
| 87 | void MatrixBatchVectorMultiplyAccumulate(armnn::Decoder<float>& matrix, |
| 88 | uint32_t mRows, |
| 89 | uint32_t mCols, |
| 90 | armnn::Decoder<float>& vector, |
| 91 | uint32_t nBatch, |
| 92 | armnn::Encoder<float>& outResult) |
| 93 | { |
| 94 | for (uint32_t b = 0; b < nBatch; b++) |
| 95 | { |
| 96 | for (uint32_t r = 0; r < mRows; r++) |
| 97 | { |
| 98 | vector += b * mCols; |
| 99 | for (uint32_t c = 0; c < mCols; c++) |
| 100 | { |
| 101 | outResult.Set(outResult.Get() + matrix.Get() * vector.Get()); |
| 102 | ++matrix; |
| 103 | ++vector; |
| 104 | } |
| 105 | outResult += 1; |
| 106 | vector -= (b+1) * mCols; |
| 107 | } |
| 108 | matrix -= (mRows * mCols); |
| 109 | } |
| 110 | outResult -= (mRows * nBatch); |
| 111 | } |
| 112 | |
| 113 | void VectorBatchVectorAssign(armnn::Decoder<float>& vector, |
| 114 | uint32_t vSize, |
| 115 | uint32_t nBatch, |
| 116 | armnn::Encoder<float>& outBatchVector) |
| 117 | { |
| 118 | for (uint32_t b = 0; b < nBatch; b++) |
| 119 | { |
| 120 | for (uint32_t v = 0; v < vSize; v++) |
| 121 | { |
| 122 | outBatchVector.Set(vector.Get()); |
| 123 | ++outBatchVector; |
| 124 | ++vector; |
| 125 | } |
| 126 | vector -= vSize; |
| 127 | } |
| 128 | outBatchVector -= (nBatch * vSize); |
| 129 | } |
| 130 | |
| 131 | void VectorBatchVectorCwiseProductAccumulate(armnn::Decoder<float>& vector, |
| 132 | uint32_t vSize, |
| 133 | armnn::Decoder<float>& batchVector, |
| 134 | uint32_t nBatch, |
| 135 | armnn::Encoder<float>& outResult) |
| 136 | { |
| 137 | for (uint32_t b = 0; b < nBatch; b++) |
| 138 | { |
| 139 | for (uint32_t v = 0; v < vSize; v++) |
| 140 | { |
| 141 | outResult.Set(outResult.Get() + vector.Get() * batchVector.Get()); |
| 142 | ++outResult; |
| 143 | ++vector; |
| 144 | ++batchVector; |
| 145 | } |
| 146 | vector -= vSize; |
| 147 | } |
| 148 | batchVector -= vSize * nBatch; |
| 149 | outResult -= vSize * nBatch; |
| 150 | } |
| 151 | |
| 152 | void VectorBatchVectorCwiseProduct(armnn::Decoder<float>& vector, |
| 153 | uint32_t vSize, |
| 154 | armnn::Decoder<float>& batchVector, |
| 155 | uint32_t nBatch, |
| 156 | armnn::Encoder<float>& outResult) |
| 157 | { |
| 158 | for (uint32_t b = 0; b < nBatch; b++) |
| 159 | { |
| 160 | for (uint32_t v = 0; v < vSize; v++) |
| 161 | { |
| 162 | outResult.Set(vector.Get() * batchVector.Get()); |
| 163 | ++outResult; |
| 164 | ++vector; |
| 165 | ++batchVector; |
| 166 | } |
| 167 | vector -= vSize; |
| 168 | } |
| 169 | batchVector -= vSize * nBatch; |
| 170 | outResult -= vSize * nBatch; |
| 171 | } |
| 172 | |
| 173 | void Sub1Vector(armnn::Decoder<float>& vector, |
| 174 | uint32_t vSize, |
| 175 | armnn::Encoder<float>& result) |
| 176 | { |
| 177 | for (uint32_t v = 0; v < vSize; v++) |
| 178 | { |
| 179 | result.Set(1.0f - vector.Get()); |
| 180 | ++vector; |
| 181 | ++result; |
| 182 | } |
| 183 | vector -= vSize; |
| 184 | result -= vSize; |
| 185 | } |
| 186 | |
| 187 | void VectorVectorCwiseProduct(armnn::Decoder<float>& vector1, |
| 188 | armnn::Decoder<float>& vector2, |
| 189 | uint32_t vSize, |
| 190 | armnn::Encoder<float>& outResult) |
| 191 | { |
| 192 | for (uint32_t v = 0; v < vSize; v++) |
| 193 | { |
| 194 | outResult.Set(vector1.Get() * vector2.Get()); |
| 195 | ++outResult; |
| 196 | ++vector1; |
| 197 | ++vector2; |
| 198 | } |
| 199 | outResult -= vSize; |
| 200 | vector1 -= vSize; |
| 201 | vector2 -= vSize; |
| 202 | } |
| 203 | |
| 204 | void VectorVectorCwiseProductAccumulate(armnn::Decoder<float>& vector1, |
| 205 | armnn::Decoder<float>& vector2, |
| 206 | uint32_t vSize, |
| 207 | armnn::Encoder<float>& outResult) |
| 208 | { |
| 209 | for (uint32_t v = 0; v < vSize; v++) |
| 210 | { |
| 211 | outResult.Set(outResult.Get() + vector1.Get() * vector2.Get()); |
| 212 | ++outResult; |
| 213 | ++vector1; |
| 214 | ++vector2; |
| 215 | } |
| 216 | outResult -= vSize; |
| 217 | vector1 -= vSize; |
| 218 | vector2 -= vSize; |
| 219 | } |
| 220 | |
| 221 | float Clip(float f, |
| 222 | float absLimit) |
| 223 | { |
| 224 | float result = (absLimit < f) ? absLimit : f; |
| 225 | result = (-absLimit > result) ? -absLimit : result; |
| 226 | return result; |
| 227 | } |
| 228 | |
| 229 | void ClipVector(armnn::Decoder<float>& vector, |
| 230 | uint32_t vSize, |
| 231 | float absLimit, |
| 232 | armnn::Encoder<float>& outResult) |
| 233 | { |
| 234 | for (uint32_t v = 0; v < vSize; v++) |
| 235 | { |
| 236 | outResult.Set(Clip(vector.Get(), absLimit)); |
| 237 | ++vector; |
| 238 | ++outResult; |
| 239 | } |
| 240 | vector -= vSize; |
| 241 | outResult -= vSize; |
| 242 | } |
| 243 | |
| 244 | void CopyVector(armnn::Decoder<float>& vector, |
| 245 | uint32_t vSize, |
| 246 | armnn::Encoder<float>& outResult) |
| 247 | { |
| 248 | for (uint32_t v = 0; v < vSize; v++) |
| 249 | { |
| 250 | outResult.Set(vector.Get()); |
| 251 | ++outResult; |
| 252 | ++vector; |
| 253 | } |
| 254 | outResult -= vSize; |
| 255 | vector -= vSize; |
| 256 | } |
| 257 | |
| 258 | void SetActivationParameters(uint32_t activation, |
| 259 | armnn::ActivationFunction& outArmnnActivation, |
| 260 | float& outA, |
| 261 | float& outB) |
| 262 | { |
| 263 | switch (activation) |
| 264 | { |
| 265 | case 0: // None |
| 266 | outA = 0; |
| 267 | outB = 0; |
| 268 | return; |
| 269 | |
| 270 | case 1: // Relu |
| 271 | outArmnnActivation = armnn::ActivationFunction::ReLu; |
| 272 | outA = 0; |
| 273 | outB = 0; |
| 274 | return; |
| 275 | |
| 276 | case 3: // Relu6 |
| 277 | outArmnnActivation = armnn::ActivationFunction::BoundedReLu; |
| 278 | outA = 6; |
| 279 | outB = 0; |
| 280 | return; |
| 281 | |
| 282 | case 4: // Tanh |
| 283 | outArmnnActivation = armnn::ActivationFunction::TanH; |
| 284 | outA = 1; |
| 285 | outB = 1; |
| 286 | return; |
| 287 | |
| 288 | case 6: // Sigmoid |
| 289 | outArmnnActivation = armnn::ActivationFunction::Sigmoid; |
| 290 | outA = 0; |
| 291 | outB = 0; |
| 292 | return; |
| 293 | |
| 294 | default: |
| 295 | throw armnn::Exception("Unsupported activation function: " + std::to_string(activation)); |
| 296 | } |
| 297 | } |
| 298 | |
James Conroy | 1f58f03 | 2021-04-27 17:13:27 +0100 | [diff] [blame] | 299 | std::unique_ptr<armnn::ScopedTensorHandle> AssignScopedTensorHandle(const armnn::ConstTensorHandle *ptr) |
Jan Eilers | 38e05bd | 2019-06-26 13:10:09 +0100 | [diff] [blame] | 300 | { |
| 301 | if (!ptr) |
| 302 | { |
| 303 | return nullptr; |
| 304 | } |
| 305 | |
James Conroy | 1f58f03 | 2021-04-27 17:13:27 +0100 | [diff] [blame] | 306 | return std::make_unique<armnn::ScopedTensorHandle>(*ptr); |
Jan Eilers | 38e05bd | 2019-06-26 13:10:09 +0100 | [diff] [blame] | 307 | } |