Tamás Nyíri | 7b885b3 | 2021-10-26 14:47:57 +0100 | [diff] [blame] | 1 | // |
| 2 | // Copyright © 2021 Arm Ltd and Contributors. All rights reserved. |
| 3 | // SPDX-License-Identifier: MIT |
| 4 | // |
| 5 | |
| 6 | #include "Pooling3d.hpp" |
| 7 | |
| 8 | #include <armnn/Exceptions.hpp> |
| 9 | #include <armnn/Types.hpp> |
| 10 | |
| 11 | #include <armnnUtils/DataLayoutIndexed.hpp> |
| 12 | #include <armnn/utility/NumericCast.hpp> |
| 13 | |
| 14 | #include <limits> |
| 15 | #include <algorithm> |
| 16 | #include <functional> |
| 17 | |
| 18 | namespace |
| 19 | { |
| 20 | using PoolingAlgorithm = armnn::PoolingAlgorithm; |
| 21 | |
| 22 | float DefaultInitializer(PoolingAlgorithm algorithm) |
| 23 | { |
| 24 | switch (algorithm) |
| 25 | { |
| 26 | case PoolingAlgorithm::Max: |
| 27 | { |
| 28 | return std::numeric_limits<float>::lowest(); |
| 29 | } |
| 30 | case PoolingAlgorithm::Average: |
| 31 | case PoolingAlgorithm::L2: |
| 32 | { |
| 33 | return 0.0f; |
| 34 | } |
| 35 | default: |
| 36 | { |
| 37 | throw armnn::InvalidArgumentException("Unsupported pooling algorithm"); |
| 38 | } |
| 39 | } |
| 40 | } |
| 41 | |
| 42 | using Accumulator = std::function<void(float & accu, float value)>; |
| 43 | |
| 44 | Accumulator GetAccumulator(PoolingAlgorithm algorithm) |
| 45 | { |
| 46 | switch (algorithm) |
| 47 | { |
| 48 | case PoolingAlgorithm::Max: |
| 49 | { |
| 50 | return [](float & accu, float value) { |
| 51 | if (value > accu) { |
| 52 | accu = value; |
| 53 | } |
| 54 | }; |
| 55 | } |
| 56 | |
| 57 | case PoolingAlgorithm::Average: |
| 58 | { |
| 59 | return [](float & accu, float value) { |
| 60 | accu += value; |
| 61 | }; |
| 62 | } |
| 63 | |
| 64 | case PoolingAlgorithm::L2: |
| 65 | { |
| 66 | return [](float & accu, float value) { |
| 67 | accu += (value*value); |
| 68 | }; |
| 69 | } |
| 70 | |
| 71 | default: |
| 72 | { |
| 73 | throw armnn::InvalidArgumentException("Unsupported pooling algorithm"); |
| 74 | } |
| 75 | } |
| 76 | } |
| 77 | |
| 78 | using Executor = std::function<void(float & accumulated, float kernelSize)>; |
| 79 | |
| 80 | Executor GetExecutor(PoolingAlgorithm algorithm) |
| 81 | { |
| 82 | switch (algorithm) |
| 83 | { |
| 84 | case PoolingAlgorithm::Max: |
| 85 | { |
| 86 | return [](float & /*accumulated*/, float /*kernelSize*/) {}; |
| 87 | } |
| 88 | |
| 89 | case PoolingAlgorithm::Average: |
| 90 | { |
| 91 | return [](float & accumulated, float kernelSize) { |
| 92 | accumulated /= kernelSize; |
| 93 | }; |
| 94 | } |
| 95 | |
| 96 | case PoolingAlgorithm::L2: |
| 97 | { |
| 98 | return [](float & accumulated, float kernelSize) { |
| 99 | accumulated = sqrtf(accumulated / kernelSize); |
| 100 | }; |
| 101 | } |
| 102 | |
| 103 | default: |
| 104 | { |
| 105 | throw armnn::InvalidArgumentException("Unsupported pooling algorithm"); |
| 106 | } |
| 107 | } |
| 108 | } |
| 109 | |
| 110 | bool OnPaddingOnly(int start, int end, int maxRange) |
| 111 | { |
| 112 | if (end <= 0 || start > maxRange) |
| 113 | { |
| 114 | return true; |
| 115 | } |
| 116 | else |
| 117 | { |
| 118 | return false; |
| 119 | } |
| 120 | } |
| 121 | |
| 122 | |
| 123 | bool ClampRange(int & start, int & end, int maxRange) |
| 124 | { |
| 125 | if (start < 0 || end > maxRange) |
| 126 | { |
| 127 | start = std::min(std::max(start, 0), maxRange); |
| 128 | end = std::min(std::max(end, 0), maxRange); |
| 129 | return true; |
| 130 | } |
| 131 | else |
| 132 | { |
| 133 | return false; |
| 134 | } |
| 135 | } |
| 136 | |
| 137 | int CalculateIndex(int channels, int depth, int height, int width, |
| 138 | int n, int c, int z, int y, int x, |
| 139 | armnnUtils::DataLayoutIndexed dataLayout) { |
| 140 | switch (dataLayout.GetDataLayout()) |
| 141 | { |
| 142 | case armnn::DataLayout::NDHWC: |
| 143 | { |
| 144 | int outputIndex = n * depth * height * width * channels + |
| 145 | z * height * width * channels + |
| 146 | y * width * channels + |
| 147 | x * channels + |
| 148 | c; |
| 149 | return outputIndex; |
| 150 | } |
| 151 | case armnn::DataLayout::NCDHW: |
| 152 | { |
| 153 | int outputIndex = n * channels * depth * height * width + |
| 154 | c * depth * height * width + |
| 155 | z * height * width + |
| 156 | y * width + |
| 157 | x; |
| 158 | return outputIndex; |
| 159 | } |
| 160 | default: |
| 161 | { |
| 162 | throw armnn::InvalidArgumentException("Unsupported data layout."); |
| 163 | } |
| 164 | } |
| 165 | } |
| 166 | } |
| 167 | |
| 168 | using namespace armnnUtils; |
| 169 | |
| 170 | namespace armnn |
| 171 | { |
| 172 | void Pooling3d(Decoder<float>& rInputDecoder, |
| 173 | Encoder<float>& rOutputEncoder, |
| 174 | const TensorInfo& inputInfo, |
| 175 | const TensorInfo& outputInfo, |
| 176 | const Pooling3dDescriptor& params) |
| 177 | { |
| 178 | const DataLayoutIndexed dataLayout(params.m_DataLayout); |
| 179 | |
| 180 | auto channelsIndex = dataLayout.GetChannelsIndex(); |
| 181 | |
| 182 | auto depthIndex = dataLayout.GetDepthIndex(); |
| 183 | auto heightIndex = dataLayout.GetHeightIndex(); |
| 184 | auto widthIndex = dataLayout.GetWidthIndex(); |
| 185 | |
| 186 | const int batchSize = armnn::numeric_cast<int>(outputInfo.GetShape()[0]); |
| 187 | const int channels = armnn::numeric_cast<int>(outputInfo.GetShape()[channelsIndex]); |
| 188 | |
| 189 | const int depthOutput = armnn::numeric_cast<int>(outputInfo.GetShape()[depthIndex]); |
| 190 | const int heightOutput = armnn::numeric_cast<int>(outputInfo.GetShape()[heightIndex]); |
| 191 | const int widthOutput = armnn::numeric_cast<int>(outputInfo.GetShape()[widthIndex]); |
| 192 | |
| 193 | const int depthInput = armnn::numeric_cast<int>(inputInfo.GetShape()[depthIndex]); |
| 194 | const int heightInput = armnn::numeric_cast<int>(inputInfo.GetShape()[heightIndex]); |
| 195 | const int widthInput = armnn::numeric_cast<int>(inputInfo.GetShape()[widthIndex]); |
| 196 | |
| 197 | const int padLeft = armnn::numeric_cast<int>(params.m_PadLeft); |
| 198 | const int padRight = armnn::numeric_cast<int>(params.m_PadRight); |
| 199 | const int padTop = armnn::numeric_cast<int>(params.m_PadTop); |
| 200 | const int padBottom = armnn::numeric_cast<int>(params.m_PadBottom); |
| 201 | const int padFront = armnn::numeric_cast<int>(params.m_PadFront); |
| 202 | const int padBack = armnn::numeric_cast<int>(params.m_PadBack); |
| 203 | |
| 204 | const int strideX = armnn::numeric_cast<int>(params.m_StrideX); |
| 205 | const int strideY = armnn::numeric_cast<int>(params.m_StrideY); |
| 206 | const int strideZ = armnn::numeric_cast<int>(params.m_StrideZ); |
| 207 | |
| 208 | const int poolHeight = armnn::numeric_cast<int>(params.m_PoolHeight); |
| 209 | const int poolWidth = armnn::numeric_cast<int>(params.m_PoolWidth); |
| 210 | const int poolDepth = armnn::numeric_cast<int>(params.m_PoolDepth); |
| 211 | |
| 212 | float defaultInitializer = DefaultInitializer(params.m_PoolType); |
| 213 | Accumulator accumulate = GetAccumulator(params.m_PoolType); |
| 214 | Executor execute = GetExecutor(params.m_PoolType); |
| 215 | |
| 216 | // Check supported padding methods outside the loop to simplify |
| 217 | // the inner loop. |
| 218 | if (params.m_PaddingMethod != PaddingMethod::Exclude && |
| 219 | params.m_PaddingMethod != PaddingMethod::IgnoreValue) |
| 220 | { |
| 221 | throw armnn::InvalidArgumentException("Unsupported padding type"); |
| 222 | } |
| 223 | |
| 224 | const std::vector<float> decodedInputVec = rInputDecoder.DecodeTensor(inputInfo.GetShape()); |
| 225 | |
| 226 | for (int n = 0; n < batchSize; n++) |
| 227 | { |
| 228 | for (int c = 0; c < channels; c++) |
| 229 | { |
| 230 | for (int zOutput = 0; zOutput < depthOutput; zOutput++) |
| 231 | { |
| 232 | // Calculate values independent of the x and y axis |
| 233 | int dstart = (zOutput * strideZ) - padFront; |
| 234 | int dend = dstart + poolDepth; |
| 235 | // Clamp the pooling region inside the valid input area (which includes the padding). |
| 236 | // This is necessary because the final pooling in a row may overlap beyond the padding. |
| 237 | dend = std::min(dend, depthInput + padBack); |
| 238 | |
| 239 | int depth = dend - dstart; |
| 240 | bool dclamped = ClampRange(dstart, dend, depthInput); |
| 241 | int depthClamped = dend - dstart; |
| 242 | |
| 243 | for (int yOutput = 0; yOutput < heightOutput; yOutput++) |
| 244 | { |
| 245 | int hstart = (yOutput * strideY) - padTop; |
| 246 | int hend = hstart + poolHeight; |
| 247 | // Clamp the pooling region inside the valid input area (which includes the padding). |
| 248 | // This is necessary because the final pooling in a row may overlap beyond the padding. |
| 249 | hend = std::min(hend, heightInput + padBottom); |
| 250 | |
| 251 | int height = hend - hstart; |
| 252 | bool hclamped = ClampRange(hstart, hend, heightInput); |
| 253 | int heightClamped = hend - hstart; |
| 254 | |
| 255 | for (int xOutput = 0; xOutput < widthOutput; xOutput++) |
| 256 | { |
| 257 | int wstart = (xOutput * strideX) - padLeft; |
| 258 | int wend = wstart + poolWidth; |
| 259 | // Clamp the pooling region inside the valid input area (which includes the padding). |
| 260 | // This is necessary because the final pooling in a row may overlap beyond the padding. |
| 261 | wend = std::min(wend, widthInput + padRight); |
| 262 | |
| 263 | int width = wend - wstart; |
| 264 | bool wclamped = ClampRange(wstart, wend, widthInput); |
| 265 | int widthClamped = wend - wstart; |
| 266 | |
| 267 | float result = defaultInitializer; |
| 268 | float poolAreaSize = armnn::numeric_cast<float>(depth * height * width); |
| 269 | |
| 270 | // Special case: when the pooling kernel is over a padding region and the padding |
| 271 | // size is larger or equal to the kernel and the kernel only covers |
| 272 | // padding and no real values, then we initialize the result as zero |
| 273 | // by convention. This is because we need to choose a value here and |
| 274 | // all values we have are padding, which we ignore. |
| 275 | if (OnPaddingOnly(dstart, dend, depthInput) || |
| 276 | OnPaddingOnly(hstart, hend, heightInput) || |
| 277 | OnPaddingOnly(wstart, wend, widthInput)) |
| 278 | { |
| 279 | result = 0.0f; |
| 280 | |
| 281 | int outputIndex = CalculateIndex(channels, depthOutput, heightOutput, widthOutput, |
| 282 | n, c, zOutput, yOutput, xOutput, dataLayout); |
| 283 | |
| 284 | rOutputEncoder[static_cast<unsigned int>(outputIndex)]; |
| 285 | rOutputEncoder.Set(result); |
| 286 | |
| 287 | continue; |
| 288 | } |
| 289 | |
| 290 | bool clamped = (dclamped | hclamped | wclamped); |
| 291 | |
| 292 | if (clamped && params.m_PaddingMethod == PaddingMethod::Exclude) |
| 293 | { |
| 294 | // When we exclude the padding, it means we calculate with a smaller |
| 295 | // kernel size, so I changed the divisor here. |
| 296 | poolAreaSize = armnn::numeric_cast<float>(depthClamped * heightClamped * widthClamped); |
| 297 | } |
| 298 | |
| 299 | for (auto zInput = dstart; zInput < dend; zInput++) |
| 300 | { |
| 301 | for (auto yInput = hstart; yInput < hend; yInput++) |
| 302 | { |
| 303 | for (auto xInput = wstart; xInput < wend; xInput++) |
| 304 | { |
| 305 | |
| 306 | int inputIndex = CalculateIndex(channels, depthInput, heightInput, widthInput, |
| 307 | n, c, zInput, yInput, xInput, dataLayout); |
| 308 | |
| 309 | accumulate(result, decodedInputVec[static_cast<unsigned int>(inputIndex)]); |
| 310 | } |
| 311 | } |
| 312 | } |
| 313 | |
| 314 | execute(result, poolAreaSize); |
| 315 | |
| 316 | int outputIndex = CalculateIndex(channels, depthOutput, heightOutput, widthOutput, |
| 317 | n, c, zOutput, yOutput, xOutput, dataLayout); |
| 318 | |
| 319 | rOutputEncoder[static_cast<unsigned int>(outputIndex)]; |
| 320 | rOutputEncoder.Set(result); |
| 321 | } |
| 322 | } |
| 323 | } |
| 324 | } |
| 325 | } |
| 326 | } |
| 327 | |
| 328 | } //namespace armnn |