Mike Kelly | b5fdf38 | 2019-06-11 16:35:25 +0100 | [diff] [blame] | 1 | // |
| 2 | // Copyright © 2017 Arm Ltd. All rights reserved. |
| 3 | // SPDX-License-Identifier: MIT |
| 4 | // |
| 5 | |
| 6 | #include "HalPolicy.hpp" |
Aron Virginas-Tar | 573a8fa | 2019-07-23 14:01:37 +0100 | [diff] [blame] | 7 | #include "Utils.hpp" |
Aron Virginas-Tar | f03fcf0 | 2019-07-09 17:44:24 +0100 | [diff] [blame] | 8 | |
Teresa Charlin | 8f6429d | 2019-10-01 13:10:15 +0100 | [diff] [blame] | 9 | #include <armnn/TypesUtils.hpp> |
| 10 | |
Matteo Martincigh | 00d6ed1 | 2019-11-28 17:13:24 +0000 | [diff] [blame] | 11 | #include <armnnUtils/DataLayoutIndexed.hpp> |
| 12 | #include <armnnUtils/TensorUtils.hpp> |
| 13 | |
| 14 | #include <Half.hpp> |
| 15 | |
Aron Virginas-Tar | 7a6d11b | 2019-07-03 15:27:08 +0100 | [diff] [blame] | 16 | #include <cmath> |
Aron Virginas-Tar | 3e0982b | 2019-10-29 14:25:09 +0000 | [diff] [blame] | 17 | #include <string> |
Aron Virginas-Tar | 7a6d11b | 2019-07-03 15:27:08 +0100 | [diff] [blame] | 18 | |
Mike Kelly | b5fdf38 | 2019-06-11 16:35:25 +0100 | [diff] [blame] | 19 | namespace armnn_driver |
| 20 | { |
| 21 | namespace hal_1_2 |
| 22 | { |
| 23 | |
Teresa Charlin | 8f6429d | 2019-10-01 13:10:15 +0100 | [diff] [blame] | 24 | using namespace armnn; |
| 25 | |
Aron Virginas-Tar | 65a1b1d | 2019-11-15 15:59:51 +0000 | [diff] [blame] | 26 | namespace |
| 27 | { |
| 28 | |
Kevin May | ec1e5b8 | 2020-02-26 17:00:39 +0000 | [diff] [blame] | 29 | bool IsQSymmDequantizeForWeights(const HalPolicy::Operation& operation, const HalPolicy::Model& model) |
Aron Virginas-Tar | 65a1b1d | 2019-11-15 15:59:51 +0000 | [diff] [blame] | 30 | { |
Kevin May | ec1e5b8 | 2020-02-26 17:00:39 +0000 | [diff] [blame] | 31 | const HalPolicy::Operand* operand = GetInputOperand<hal_1_2::HalPolicy>(operation, 0, model); |
Aron Virginas-Tar | 65a1b1d | 2019-11-15 15:59:51 +0000 | [diff] [blame] | 32 | if (!operand) |
| 33 | { |
| 34 | return false; |
| 35 | } |
| 36 | |
| 37 | if(!IsQSymm8(*operand)) |
| 38 | { |
| 39 | // Only QSymm8 weights are dequantized on the fly by the driver |
| 40 | return false; |
| 41 | } |
| 42 | |
| 43 | if (!IsOperandConstant<hal_1_2::HalPolicy>(*operand)) |
| 44 | { |
| 45 | // Non-const input is not accepted for weights |
| 46 | return false; |
| 47 | } |
| 48 | |
| 49 | // Iterate through all the operations and find the operation feeding from the Dequantize output |
| 50 | const size_t outputIndex = operation.outputs[0]; |
| 51 | for (uint32_t operationIdx = 0; operationIdx < model.operations.size(); ++operationIdx) |
| 52 | { |
| 53 | const auto& operationIt = model.operations[operationIdx]; |
| 54 | switch (operationIt.type) |
| 55 | { |
| 56 | case HalPolicy::OperationType::FULLY_CONNECTED: |
| 57 | if (outputIndex == operationIt.inputs[1]) // Weights are bound to slot 1 |
| 58 | { |
| 59 | // If the output is going into the FC weights return true |
| 60 | return true; |
| 61 | } |
| 62 | break; |
| 63 | case HalPolicy::OperationType::LSTM: |
| 64 | for (size_t k = 0; k < operationIt.inputs.size(); ++k) |
| 65 | { |
| 66 | if (outputIndex == operationIt.inputs[k]) |
| 67 | { |
| 68 | // If the output is going into the LSTM weights return true |
| 69 | return true; |
| 70 | } |
| 71 | } |
| 72 | break; |
| 73 | default: |
| 74 | break; |
| 75 | } |
| 76 | } |
| 77 | |
| 78 | return false; |
| 79 | } |
| 80 | |
| 81 | } // anonymous namespace |
| 82 | |
Mike Kelly | b5fdf38 | 2019-06-11 16:35:25 +0100 | [diff] [blame] | 83 | bool HalPolicy::ConvertOperation(const Operation& operation, const Model& model, ConversionData& data) |
| 84 | { |
Mike Kelly | b5fdf38 | 2019-06-11 16:35:25 +0100 | [diff] [blame] | 85 | switch (operation.type) |
| 86 | { |
Kevin May | 407718f | 2019-09-09 14:46:41 +0100 | [diff] [blame] | 87 | case V1_2::OperationType::ABS: |
josh minor | 00a963b | 2020-01-08 11:55:35 -0600 | [diff] [blame] | 88 | return ConvertElementwiseUnary(operation, model, data, UnaryOperation::Abs); |
Mike Kelly | 4627280 | 2019-08-14 17:00:48 +0100 | [diff] [blame] | 89 | case V1_2::OperationType::ADD: |
| 90 | return ConvertAdd(operation, model, data); |
Francis Murtagh | 19fa0cc | 2019-11-19 12:06:47 +0000 | [diff] [blame] | 91 | case V1_2::OperationType::ARGMAX: |
| 92 | return ConvertArgMinMax(operation, model, data, ArgMinMaxFunction::Max); |
| 93 | case V1_2::OperationType::ARGMIN: |
| 94 | return ConvertArgMinMax(operation, model, data, ArgMinMaxFunction::Min); |
Sadik Armagan | 15d63e2 | 2019-07-26 16:59:35 +0100 | [diff] [blame] | 95 | case V1_2::OperationType::AVERAGE_POOL_2D: |
| 96 | return ConvertAveragePool2d(operation, model, data); |
Finn Williams | 23b87b3 | 2019-07-30 11:44:05 +0100 | [diff] [blame] | 97 | case V1_2::OperationType::BATCH_TO_SPACE_ND: |
| 98 | return ConvertBatchToSpaceNd(operation, model, data); |
Mike Kelly | b880520 | 2019-07-31 17:25:43 +0100 | [diff] [blame] | 99 | case V1_2::OperationType::CONCATENATION: |
| 100 | return ConvertConcatenation(operation, model, data); |
Mike Kelly | b5fdf38 | 2019-06-11 16:35:25 +0100 | [diff] [blame] | 101 | case V1_2::OperationType::CONV_2D: |
Aron Virginas-Tar | 24e699d | 2019-06-17 14:47:46 +0100 | [diff] [blame] | 102 | return ConvertConv2d(operation, model, data); |
Aron Virginas-Tar | 8edb16d | 2019-10-01 13:34:59 +0100 | [diff] [blame] | 103 | case V1_2::OperationType::DEPTH_TO_SPACE: |
| 104 | return ConvertDepthToSpace(operation, model, data); |
Mike Kelly | b5fdf38 | 2019-06-11 16:35:25 +0100 | [diff] [blame] | 105 | case V1_2::OperationType::DEPTHWISE_CONV_2D: |
Aron Virginas-Tar | 24e699d | 2019-06-17 14:47:46 +0100 | [diff] [blame] | 106 | return ConvertDepthwiseConv2d(operation, model, data); |
Mike Kelly | 4627280 | 2019-08-14 17:00:48 +0100 | [diff] [blame] | 107 | case V1_2::OperationType::DEQUANTIZE: |
| 108 | return ConvertDequantize(operation, model, data); |
| 109 | case V1_2::OperationType::DIV: |
| 110 | return ConvertDiv(operation, model, data); |
Aron Virginas-Tar | 3e0982b | 2019-10-29 14:25:09 +0000 | [diff] [blame] | 111 | case V1_2::OperationType::EQUAL: |
| 112 | return ConvertComparison(operation, model, data, ComparisonOperation::Equal); |
Narumol Prangnawarat | 85f9654 | 2019-09-12 16:26:29 +0100 | [diff] [blame] | 113 | case V1_2::OperationType::EXPAND_DIMS: |
| 114 | return ConvertExpandDims(operation, model, data); |
Mike Kelly | 4627280 | 2019-08-14 17:00:48 +0100 | [diff] [blame] | 115 | case V1_2::OperationType::FLOOR: |
| 116 | return ConvertFloor(operation, model, data); |
| 117 | case V1_2::OperationType::FULLY_CONNECTED: |
| 118 | return ConvertFullyConnected(operation, model, data); |
Aron Virginas-Tar | 3e0982b | 2019-10-29 14:25:09 +0000 | [diff] [blame] | 119 | case V1_2::OperationType::GREATER: |
| 120 | return ConvertComparison(operation, model, data, ComparisonOperation::Greater); |
| 121 | case V1_2::OperationType::GREATER_EQUAL: |
| 122 | return ConvertComparison(operation, model, data, ComparisonOperation::GreaterOrEqual); |
Teresa Charlin | 8f6429d | 2019-10-01 13:10:15 +0100 | [diff] [blame] | 123 | case V1_2::OperationType::GROUPED_CONV_2D: |
| 124 | return ConvertGroupedConv2d(operation, model, data); |
Aron Virginas-Tar | a2a7380 | 2019-10-09 15:30:40 +0100 | [diff] [blame] | 125 | case V1_2::OperationType::INSTANCE_NORMALIZATION: |
| 126 | return ConvertInstanceNormalization(operation, model, data); |
Mike Kelly | 4627280 | 2019-08-14 17:00:48 +0100 | [diff] [blame] | 127 | case V1_2::OperationType::L2_NORMALIZATION: |
| 128 | return ConvertL2Normalization(operation, model, data); |
Sadik Armagan | 15d63e2 | 2019-07-26 16:59:35 +0100 | [diff] [blame] | 129 | case V1_2::OperationType::L2_POOL_2D: |
| 130 | return ConvertL2Pool2d(operation, model, data); |
Aron Virginas-Tar | 3e0982b | 2019-10-29 14:25:09 +0000 | [diff] [blame] | 131 | case V1_2::OperationType::LESS: |
| 132 | return ConvertComparison(operation, model, data, ComparisonOperation::Less); |
| 133 | case V1_2::OperationType::LESS_EQUAL: |
| 134 | return ConvertComparison(operation, model, data, ComparisonOperation::LessOrEqual); |
Mike Kelly | 4627280 | 2019-08-14 17:00:48 +0100 | [diff] [blame] | 135 | case V1_2::OperationType::LOCAL_RESPONSE_NORMALIZATION: |
| 136 | return ConvertLocalResponseNormalization(operation, model, data); |
| 137 | case V1_2::OperationType::LOGISTIC: |
| 138 | return ConvertLogistic(operation, model, data); |
Aron Virginas-Tar | 75e6779 | 2019-10-15 13:33:03 +0100 | [diff] [blame] | 139 | case V1_2::OperationType::LOG_SOFTMAX: |
| 140 | return ConvertLogSoftmax(operation, model, data); |
Mike Kelly | 4627280 | 2019-08-14 17:00:48 +0100 | [diff] [blame] | 141 | case V1_2::OperationType::LSTM: |
| 142 | return ConvertLstm(operation, model, data); |
Sadik Armagan | 15d63e2 | 2019-07-26 16:59:35 +0100 | [diff] [blame] | 143 | case V1_2::OperationType::MAX_POOL_2D: |
| 144 | return ConvertMaxPool2d(operation, model, data); |
Narumol Prangnawarat | 95b1ef6 | 2019-07-15 12:02:20 +0100 | [diff] [blame] | 145 | case V1_2::OperationType::MAXIMUM: |
| 146 | return ConvertMaximum(operation, model, data); |
Mike Kelly | 4627280 | 2019-08-14 17:00:48 +0100 | [diff] [blame] | 147 | case V1_2::OperationType::MEAN: |
| 148 | return ConvertMean(operation, model, data); |
Ellen Norris-Thompson | 1cb29aa | 2019-07-11 17:27:37 +0100 | [diff] [blame] | 149 | case V1_2::OperationType::MINIMUM: |
| 150 | return ConvertMinimum(operation, model, data); |
Mike Kelly | 4627280 | 2019-08-14 17:00:48 +0100 | [diff] [blame] | 151 | case V1_2::OperationType::MUL: |
| 152 | return ConvertMul(operation, model, data); |
Aron Virginas-Tar | 3e0982b | 2019-10-29 14:25:09 +0000 | [diff] [blame] | 153 | case V1_2::OperationType::NOT_EQUAL: |
| 154 | return ConvertComparison(operation, model, data, ComparisonOperation::NotEqual); |
Mike Kelly | 3c67394 | 2019-07-25 09:26:06 +0100 | [diff] [blame] | 155 | case V1_2::OperationType::PAD: |
Aron Virginas-Tar | c921f6b | 2019-07-25 10:14:33 +0100 | [diff] [blame] | 156 | return ConvertPad(operation, model, data); |
Aron Virginas-Tar | cb8ac84 | 2019-07-05 15:47:07 +0100 | [diff] [blame] | 157 | case V1_2::OperationType::PAD_V2: |
| 158 | return ConvertPadV2(operation, model, data); |
Matteo Martincigh | 17ffff3 | 2019-06-27 14:12:55 +0100 | [diff] [blame] | 159 | case V1_2::OperationType::PRELU: |
| 160 | return ConvertPrelu(operation, model, data); |
Sadik Armagan | 5a476a8 | 2019-07-30 09:43:18 +0100 | [diff] [blame] | 161 | case V1_2::OperationType::QUANTIZE: |
| 162 | return ConvertQuantize(operation, model, data); |
Ellen Norris-Thompson | 7efb46d | 2019-07-24 17:39:19 +0100 | [diff] [blame] | 163 | case V1_2::OperationType::QUANTIZED_16BIT_LSTM: |
| 164 | return ConvertQuantizedLstm(operation, model, data); |
Sadik Armagan | 6111316 | 2019-07-25 09:09:40 +0100 | [diff] [blame] | 165 | case V1_2::OperationType::RELU: |
| 166 | return ConvertReLu(operation, model, data); |
| 167 | case V1_2::OperationType::RELU1: |
| 168 | return ConvertReLu1(operation, model, data); |
| 169 | case V1_2::OperationType::RELU6: |
| 170 | return ConvertReLu6(operation, model, data); |
Mike Kelly | 4627280 | 2019-08-14 17:00:48 +0100 | [diff] [blame] | 171 | case V1_2::OperationType::RESHAPE: |
| 172 | return ConvertReshape(operation, model, data); |
Aron Virginas-Tar | fb2fa29 | 2019-07-04 11:59:48 +0100 | [diff] [blame] | 173 | case V1_2::OperationType::RESIZE_BILINEAR: |
Teresa Charlin | 8f6429d | 2019-10-01 13:10:15 +0100 | [diff] [blame] | 174 | return ConvertResize(operation, model, data, ResizeMethod::Bilinear); |
Aron Virginas-Tar | 7a6d11b | 2019-07-03 15:27:08 +0100 | [diff] [blame] | 175 | case V1_2::OperationType::RESIZE_NEAREST_NEIGHBOR: |
Teresa Charlin | 8f6429d | 2019-10-01 13:10:15 +0100 | [diff] [blame] | 176 | return ConvertResize(operation, model, data, ResizeMethod::NearestNeighbor); |
Aron Virginas-Tar | fa6544e | 2019-09-10 14:42:22 +0100 | [diff] [blame] | 177 | case V1_2::OperationType::RSQRT: |
josh minor | 00a963b | 2020-01-08 11:55:35 -0600 | [diff] [blame] | 178 | return ConvertElementwiseUnary(operation, model, data, UnaryOperation::Rsqrt); |
Sadik Armagan | 701d9a0 | 2019-09-04 15:16:18 +0100 | [diff] [blame] | 179 | case V1_2::OperationType::SQRT: |
| 180 | return ConvertSqrt(operation, model, data); |
Mike Kelly | 4627280 | 2019-08-14 17:00:48 +0100 | [diff] [blame] | 181 | case V1_2::OperationType::SQUEEZE: |
| 182 | return ConvertSqueeze(operation, model, data); |
| 183 | case V1_2::OperationType::STRIDED_SLICE: |
| 184 | return ConvertStridedSlice(operation, model, data); |
| 185 | case V1_2::OperationType::TRANSPOSE: |
| 186 | return ConvertTranspose(operation, model, data); |
David Monahan | 613b49c | 2019-06-27 11:37:47 +0100 | [diff] [blame] | 187 | case V1_2::OperationType::TRANSPOSE_CONV_2D: |
Aron Virginas-Tar | 8b99168 | 2019-07-31 12:54:59 +0100 | [diff] [blame] | 188 | return ConvertTransposeConv2d(operation, model, data); |
Francis Murtagh | 074c25a | 2019-07-22 16:40:57 +0100 | [diff] [blame] | 189 | case V1_2::OperationType::SOFTMAX: |
| 190 | return ConvertSoftmax(operation, model, data); |
Finn Williams | d74c505 | 2019-07-30 17:06:00 +0100 | [diff] [blame] | 191 | case V1_2::OperationType::SPACE_TO_BATCH_ND : |
| 192 | return ConvertSpaceToBatchNd(operation, model, data); |
Aron Virginas-Tar | ad1ab53 | 2019-07-25 11:24:42 +0100 | [diff] [blame] | 193 | case V1_2::OperationType::SPACE_TO_DEPTH: |
| 194 | return ConvertSpaceToDepth(operation, model, data); |
Mike Kelly | 0a87936 | 2019-07-29 16:56:31 +0100 | [diff] [blame] | 195 | case V1_2::OperationType::SUB: |
| 196 | return ConvertSub(operation, model, data); |
Sadik Armagan | 6111316 | 2019-07-25 09:09:40 +0100 | [diff] [blame] | 197 | case V1_2::OperationType::TANH: |
| 198 | return ConvertTanH(operation, model, data); |
Mike Kelly | b5fdf38 | 2019-06-11 16:35:25 +0100 | [diff] [blame] | 199 | default: |
| 200 | return Fail("%s: Operation type %s not supported in ArmnnDriver", |
| 201 | __func__, toString(operation.type).c_str()); |
| 202 | } |
| 203 | } |
| 204 | |
Mike Kelly | 4627280 | 2019-08-14 17:00:48 +0100 | [diff] [blame] | 205 | bool HalPolicy::ConvertAdd(const Operation& operation, const Model& model, ConversionData& data) |
| 206 | { |
| 207 | ALOGV("hal_1_2::HalPolicy::ConvertAdd()"); |
| 208 | return ::ConvertAdd<hal_1_2::HalPolicy>(operation, model, data); |
| 209 | } |
| 210 | |
Francis Murtagh | 19fa0cc | 2019-11-19 12:06:47 +0000 | [diff] [blame] | 211 | bool HalPolicy::ConvertArgMinMax(const V1_2::Operation& operation, |
| 212 | const V1_2::Model& model, |
| 213 | ConversionData& data, |
| 214 | armnn::ArgMinMaxFunction argMinMaxFunction) |
| 215 | { |
| 216 | ALOGV("hal_1_2::HalPolicy::ConvertArgMinMax()"); |
| 217 | return ::ConvertArgMinMax<hal_1_2::HalPolicy>(operation, model, data, argMinMaxFunction); |
| 218 | } |
| 219 | |
Sadik Armagan | 15d63e2 | 2019-07-26 16:59:35 +0100 | [diff] [blame] | 220 | bool HalPolicy::ConvertAveragePool2d(const Operation& operation, const Model& model, ConversionData& data) |
| 221 | { |
| 222 | ALOGV("hal_1_2::HalPolicy::ConvertAveragePool2d()"); |
Teresa Charlin | 8f6429d | 2019-10-01 13:10:15 +0100 | [diff] [blame] | 223 | return ConvertPooling2d<hal_1_2::HalPolicy>(operation, __func__, PoolingAlgorithm::Average, model, data); |
Sadik Armagan | 15d63e2 | 2019-07-26 16:59:35 +0100 | [diff] [blame] | 224 | } |
| 225 | |
Finn Williams | 23b87b3 | 2019-07-30 11:44:05 +0100 | [diff] [blame] | 226 | bool HalPolicy::ConvertBatchToSpaceNd(const Operation& operation, const Model& model, ConversionData& data) |
| 227 | { |
| 228 | ALOGV("hal_1_2::HalPolicy::ConvertBatchToSpaceNd()"); |
| 229 | return ::ConvertBatchToSpaceNd<hal_1_2::HalPolicy>(operation, model, data); |
| 230 | } |
| 231 | |
Aron Virginas-Tar | 3e0982b | 2019-10-29 14:25:09 +0000 | [diff] [blame] | 232 | bool HalPolicy::ConvertComparison(const Operation& operation, |
| 233 | const Model& model, |
| 234 | ConversionData& data, |
| 235 | ComparisonOperation comparisonOperation) |
| 236 | { |
| 237 | ALOGV("hal_1_2::HalPolicy::ConvertComparison()"); |
| 238 | ALOGV("comparisonOperation = %s", GetComparisonOperationAsCString(comparisonOperation)); |
| 239 | |
| 240 | LayerInputHandle input0 = ConvertToLayerInputHandle<hal_1_2::HalPolicy>(operation, 0, model, data); |
| 241 | LayerInputHandle input1 = ConvertToLayerInputHandle<hal_1_2::HalPolicy>(operation, 1, model, data); |
| 242 | |
| 243 | if (!(input0.IsValid() && input1.IsValid())) |
| 244 | { |
| 245 | return Fail("%s: Operation has invalid inputs", __func__); |
| 246 | } |
| 247 | |
| 248 | const Operand* output = GetOutputOperand<hal_1_2::HalPolicy>(operation, 0, model); |
| 249 | if (!output) |
| 250 | { |
| 251 | return Fail("%s: Could not read output 0", __func__); |
| 252 | } |
| 253 | |
| 254 | const TensorInfo& inputInfo0 = input0.GetTensorInfo(); |
| 255 | const TensorInfo& inputInfo1 = input1.GetTensorInfo(); |
| 256 | const TensorInfo& outputInfo = GetTensorInfoForOperand(*output); |
| 257 | |
| 258 | if (IsDynamicTensor(outputInfo)) |
| 259 | { |
| 260 | return Fail("%s: Dynamic output tensors are not supported", __func__); |
| 261 | } |
| 262 | |
| 263 | ComparisonDescriptor descriptor(comparisonOperation); |
| 264 | |
| 265 | bool isSupported = false; |
| 266 | FORWARD_LAYER_SUPPORT_FUNC(__func__, |
| 267 | IsComparisonSupported, |
| 268 | data.m_Backends, |
| 269 | isSupported, |
| 270 | inputInfo0, |
| 271 | inputInfo1, |
| 272 | outputInfo, |
| 273 | descriptor); |
| 274 | |
| 275 | if (!isSupported) |
| 276 | { |
| 277 | return false; |
| 278 | } |
| 279 | |
| 280 | IConnectableLayer* layer = data.m_Network->AddComparisonLayer(descriptor); |
| 281 | assert(layer != nullptr); |
Sadik Armagan | 793a70c | 2020-03-19 13:54:04 +0000 | [diff] [blame^] | 282 | bool isReshapeSupported = BroadcastTensor(input0, input1, layer, data); |
| 283 | if (!isReshapeSupported) |
| 284 | { |
| 285 | return false; |
| 286 | } |
Aron Virginas-Tar | 3e0982b | 2019-10-29 14:25:09 +0000 | [diff] [blame] | 287 | |
| 288 | return SetupAndTrackLayerOutputSlot<hal_1_2::HalPolicy>(operation, 0, *layer, model, data); |
| 289 | } |
| 290 | |
Mike Kelly | b880520 | 2019-07-31 17:25:43 +0100 | [diff] [blame] | 291 | bool HalPolicy::ConvertConcatenation(const Operation& operation, const Model& model, ConversionData& data) |
| 292 | { |
| 293 | ALOGV("hal_1_2::HalPolicy::ConvertConcatenation()"); |
| 294 | return ::ConvertConcatenation<hal_1_2::HalPolicy>(operation, model, data); |
| 295 | } |
| 296 | |
Aron Virginas-Tar | 24e699d | 2019-06-17 14:47:46 +0100 | [diff] [blame] | 297 | bool HalPolicy::ConvertConv2d(const Operation& operation, const Model& model, ConversionData& data) |
| 298 | { |
Aron Virginas-Tar | 29404fb | 2019-07-24 13:55:31 +0100 | [diff] [blame] | 299 | ALOGV("hal_1_2::HalPolicy::ConvertConv2d()"); |
| 300 | |
Aron Virginas-Tar | 24e699d | 2019-06-17 14:47:46 +0100 | [diff] [blame] | 301 | LayerInputHandle input = ConvertToLayerInputHandle<hal_1_2::HalPolicy>(operation, 0, model, data); |
| 302 | if (!input.IsValid()) |
| 303 | { |
| 304 | return Fail("%s: Operation has invalid inputs", __func__); |
| 305 | } |
| 306 | |
| 307 | const Operand* output = GetOutputOperand<hal_1_2::HalPolicy>(operation, 0, model); |
| 308 | if (!output) |
| 309 | { |
| 310 | return Fail("%s: Could not read output 0", __func__); |
| 311 | } |
| 312 | |
Teresa Charlin | 8f6429d | 2019-10-01 13:10:15 +0100 | [diff] [blame] | 313 | const TensorInfo& inputInfo = input.GetTensorInfo(); |
| 314 | const TensorInfo& outputInfo = GetTensorInfoForOperand(*output); |
Aron Virginas-Tar | b7421e5 | 2019-07-26 13:14:39 +0100 | [diff] [blame] | 315 | |
| 316 | if (IsDynamicTensor(outputInfo)) |
| 317 | { |
| 318 | return Fail("%s: Dynamic output tensors are not supported", __func__); |
| 319 | } |
Aron Virginas-Tar | 366e0a6 | 2019-07-10 13:01:41 +0100 | [diff] [blame] | 320 | |
Teresa Charlin | 8f6429d | 2019-10-01 13:10:15 +0100 | [diff] [blame] | 321 | Convolution2dDescriptor desc; |
| 322 | desc.m_DataLayout = DataLayout::NHWC; |
Mike Kelly | e1d60bb | 2019-07-11 11:44:52 +0100 | [diff] [blame] | 323 | |
| 324 | // Determine whether padding is implicit or explicit |
| 325 | bool implicitPadding = operation.inputs.size() == 7 || |
| 326 | (operation.inputs.size() >= 8 && |
| 327 | GetInputOperand<hal_1_2::HalPolicy>(operation, 7, model)->type == OperandType::BOOL); |
| 328 | |
| 329 | if (implicitPadding) |
| 330 | { |
| 331 | desc.m_DataLayout = OptionalDataLayout<hal_1_2::HalPolicy>(operation, 7, model, data); |
| 332 | } |
| 333 | else if (operation.inputs.size() >= 10) |
| 334 | { |
| 335 | desc.m_DataLayout = OptionalDataLayout<hal_1_2::HalPolicy>(operation, 10, model, data); |
| 336 | } |
| 337 | |
Teresa Charlin | 8f6429d | 2019-10-01 13:10:15 +0100 | [diff] [blame] | 338 | const PermutationVector OHWIToOIHW = {0, 2, 3, 1}; |
Mike Kelly | e1d60bb | 2019-07-11 11:44:52 +0100 | [diff] [blame] | 339 | |
Aron Virginas-Tar | 24e699d | 2019-06-17 14:47:46 +0100 | [diff] [blame] | 340 | // ArmNN does not currently support non-fixed weights or bias |
Mike Kelly | e1d60bb | 2019-07-11 11:44:52 +0100 | [diff] [blame] | 341 | // The NNAPI filter is always OHWI [depth_out, filter_height, filter_width, depth_in] but ArmNN expects the |
| 342 | // filter's height and width indices to match the input's height and width indices so we permute it to OIHW if |
| 343 | // the DataLayout is NCHW |
Teresa Charlin | 8f6429d | 2019-10-01 13:10:15 +0100 | [diff] [blame] | 344 | const ConstTensorPin weightsPin = (desc.m_DataLayout == DataLayout::NCHW) ? |
Mike Kelly | e1d60bb | 2019-07-11 11:44:52 +0100 | [diff] [blame] | 345 | ConvertOperationInputToConstTensorPin<hal_1_2::HalPolicy>(operation, 1, model, data, OHWIToOIHW) : |
| 346 | ConvertOperationInputToConstTensorPin<hal_1_2::HalPolicy>(operation, 1, model, data); |
Aron Virginas-Tar | 24e699d | 2019-06-17 14:47:46 +0100 | [diff] [blame] | 347 | const ConstTensorPin biasPin = |
Mike Kelly | e1d60bb | 2019-07-11 11:44:52 +0100 | [diff] [blame] | 348 | ConvertOperationInputToConstTensorPin<hal_1_2::HalPolicy>(operation, 2, model, data); |
Aron Virginas-Tar | 24e699d | 2019-06-17 14:47:46 +0100 | [diff] [blame] | 349 | |
| 350 | if (!weightsPin.IsValid()) |
| 351 | { |
| 352 | return Fail("%s: Operation has invalid weights", __func__); |
| 353 | } |
| 354 | |
| 355 | if (!biasPin.IsValid()) |
| 356 | { |
| 357 | return Fail("%s: Operation has invalid biases", __func__); |
| 358 | } |
| 359 | |
Teresa Charlin | 8f6429d | 2019-10-01 13:10:15 +0100 | [diff] [blame] | 360 | ConstTensor weights = weightsPin.GetConstTensor(); |
| 361 | ConstTensor bias = biasPin.GetConstTensor(); |
Aron Virginas-Tar | 24e699d | 2019-06-17 14:47:46 +0100 | [diff] [blame] | 362 | SanitizeBiasQuantizationScale(bias.GetInfo(), weights.GetInfo(), inputInfo); |
| 363 | |
Aron Virginas-Tar | 24e699d | 2019-06-17 14:47:46 +0100 | [diff] [blame] | 364 | ActivationFn activation; |
| 365 | |
Aron Virginas-Tar | 24e699d | 2019-06-17 14:47:46 +0100 | [diff] [blame] | 366 | if (implicitPadding) |
| 367 | { |
| 368 | android::nn::PaddingScheme paddingScheme; |
| 369 | if (!GetInputPaddingScheme<hal_1_2::HalPolicy>(operation, 3, paddingScheme, model, data) || |
| 370 | !GetInputScalar<hal_1_2::HalPolicy>(operation, 4, OperandType::INT32, desc.m_StrideX, model, data) || |
| 371 | !GetInputScalar<hal_1_2::HalPolicy>(operation, 5, OperandType::INT32, desc.m_StrideY, model, data) || |
| 372 | !GetInputActivationFunction<hal_1_2::HalPolicy>(operation, 6, activation, model, data) || |
| 373 | !GetOptionalConvolutionDilationParams<hal_1_2::HalPolicy>(operation, 8, desc, model, data)) |
| 374 | { |
| 375 | return Fail("%s: Operation has invalid inputs (implicit padding)", __func__); |
| 376 | } |
| 377 | |
Mike Kelly | e1d60bb | 2019-07-11 11:44:52 +0100 | [diff] [blame] | 378 | armnnUtils::DataLayoutIndexed dataLayoutIndexed(desc.m_DataLayout); |
| 379 | unsigned int widthIndex = dataLayoutIndexed.GetWidthIndex(); |
| 380 | unsigned int heightIndex = dataLayoutIndexed.GetHeightIndex(); |
| 381 | const uint32_t kernelX = weights.GetShape()[widthIndex]; |
| 382 | const uint32_t kernelY = weights.GetShape()[heightIndex]; |
| 383 | const uint32_t inputX = inputInfo.GetShape()[widthIndex]; |
| 384 | const uint32_t inputY = inputInfo.GetShape()[heightIndex]; |
Aron Virginas-Tar | 24e699d | 2019-06-17 14:47:46 +0100 | [diff] [blame] | 385 | |
Mike Kelly | 86b36d4 | 2019-07-12 16:39:33 +0100 | [diff] [blame] | 386 | CalcPadding(inputX, kernelX, desc.m_StrideX, desc.m_DilationX, desc.m_PadLeft, desc.m_PadRight, paddingScheme); |
| 387 | CalcPadding(inputY, kernelY, desc.m_StrideY, desc.m_DilationY, desc.m_PadTop, desc.m_PadBottom, paddingScheme); |
Aron Virginas-Tar | 24e699d | 2019-06-17 14:47:46 +0100 | [diff] [blame] | 388 | |
Aron Virginas-Tar | 24e699d | 2019-06-17 14:47:46 +0100 | [diff] [blame] | 389 | } |
| 390 | else if (operation.inputs.size() >= 10) |
| 391 | { |
| 392 | // explicit padding |
| 393 | if (!GetInputScalar<hal_1_2::HalPolicy>(operation, 3, OperandType::INT32, desc.m_PadLeft, model, data) || |
| 394 | !GetInputScalar<hal_1_2::HalPolicy>(operation, 4, OperandType::INT32, desc.m_PadRight, model, data) || |
| 395 | !GetInputScalar<hal_1_2::HalPolicy>(operation, 5, OperandType::INT32, desc.m_PadTop, model, data) || |
| 396 | !GetInputScalar<hal_1_2::HalPolicy>(operation, 6, OperandType::INT32, desc.m_PadBottom, model, data) || |
| 397 | !GetInputScalar<hal_1_2::HalPolicy>(operation, 7, OperandType::INT32, desc.m_StrideX, model, data) || |
| 398 | !GetInputScalar<hal_1_2::HalPolicy>(operation, 8, OperandType::INT32, desc.m_StrideY, model, data) || |
| 399 | !GetInputActivationFunction<hal_1_2::HalPolicy>(operation, 9, activation, model, data) || |
| 400 | !GetOptionalConvolutionDilationParams<hal_1_2::HalPolicy>(operation, 11, desc, model, data)) |
| 401 | { |
| 402 | return Fail("%s: Operation has invalid inputs (explicit padding)", __func__); |
| 403 | } |
Aron Virginas-Tar | 24e699d | 2019-06-17 14:47:46 +0100 | [diff] [blame] | 404 | } |
| 405 | else |
| 406 | { |
| 407 | return Fail("%s: Unsupported number of operation inputs", __func__); |
| 408 | } |
| 409 | |
| 410 | desc.m_BiasEnabled = true; |
Teresa Charlin | 8f6429d | 2019-10-01 13:10:15 +0100 | [diff] [blame] | 411 | Optional<TensorInfo> biases(bias.GetInfo()); |
Aron Virginas-Tar | 24e699d | 2019-06-17 14:47:46 +0100 | [diff] [blame] | 412 | |
Ferran Balaguer | d30093c | 2019-07-09 17:04:47 +0100 | [diff] [blame] | 413 | bool isSupported = false; |
| 414 | FORWARD_LAYER_SUPPORT_FUNC(__func__, |
| 415 | IsConvolution2dSupported, |
| 416 | data.m_Backends, |
| 417 | isSupported, |
| 418 | inputInfo, |
| 419 | outputInfo, |
| 420 | desc, |
| 421 | weights.GetInfo(), |
| 422 | biases); |
Aron Virginas-Tar | 2b17312 | 2019-07-15 14:29:09 +0100 | [diff] [blame] | 423 | |
Ferran Balaguer | d30093c | 2019-07-09 17:04:47 +0100 | [diff] [blame] | 424 | if (!isSupported) |
Aron Virginas-Tar | 24e699d | 2019-06-17 14:47:46 +0100 | [diff] [blame] | 425 | { |
| 426 | return false; |
| 427 | } |
| 428 | |
Teresa Charlin | 8f6429d | 2019-10-01 13:10:15 +0100 | [diff] [blame] | 429 | IConnectableLayer* startLayer = |
| 430 | data.m_Network->AddConvolution2dLayer(desc, weights, Optional<ConstTensor>(bias)); |
Aron Virginas-Tar | 24e699d | 2019-06-17 14:47:46 +0100 | [diff] [blame] | 431 | |
| 432 | if (!startLayer) |
| 433 | { |
| 434 | return Fail("%s: AddConvolution2dLayer failed", __func__); |
| 435 | } |
| 436 | |
Teresa Charlin | 8f6429d | 2019-10-01 13:10:15 +0100 | [diff] [blame] | 437 | IConnectableLayer* endLayer = ProcessActivation(outputInfo, activation, startLayer, data); |
Aron Virginas-Tar | 24e699d | 2019-06-17 14:47:46 +0100 | [diff] [blame] | 438 | |
| 439 | if (!endLayer) |
| 440 | { |
| 441 | return Fail("%s: ProcessActivation failed", __func__); |
| 442 | } |
| 443 | |
| 444 | input.Connect(startLayer->GetInputSlot(0)); |
| 445 | |
Aron Virginas-Tar | b7421e5 | 2019-07-26 13:14:39 +0100 | [diff] [blame] | 446 | return SetupAndTrackLayerOutputSlot<hal_1_2::HalPolicy>(operation, 0, *endLayer, model, data); |
Aron Virginas-Tar | 24e699d | 2019-06-17 14:47:46 +0100 | [diff] [blame] | 447 | } |
| 448 | |
Aron Virginas-Tar | 8edb16d | 2019-10-01 13:34:59 +0100 | [diff] [blame] | 449 | bool HalPolicy::ConvertDepthToSpace(const Operation& operation, const Model& model, ConversionData& data) |
| 450 | { |
| 451 | ALOGV("hal_1_2::HalPolicy::ConvertDepthToSpace()"); |
| 452 | return ::ConvertDepthToSpace<hal_1_2::HalPolicy>(operation, model, data); |
| 453 | } |
| 454 | |
Aron Virginas-Tar | 24e699d | 2019-06-17 14:47:46 +0100 | [diff] [blame] | 455 | bool HalPolicy::ConvertDepthwiseConv2d(const Operation& operation, const Model& model, ConversionData& data) |
| 456 | { |
Aron Virginas-Tar | 29404fb | 2019-07-24 13:55:31 +0100 | [diff] [blame] | 457 | ALOGV("hal_1_2::HalPolicy::ConvertDepthwiseConv2d()"); |
| 458 | |
Aron Virginas-Tar | 24e699d | 2019-06-17 14:47:46 +0100 | [diff] [blame] | 459 | LayerInputHandle input = ConvertToLayerInputHandle<hal_1_2::HalPolicy>(operation, 0, model, data); |
| 460 | |
| 461 | if (!input.IsValid()) |
| 462 | { |
| 463 | return Fail("%s: Operation has invalid inputs", __func__); |
| 464 | } |
| 465 | |
| 466 | const Operand* output = GetOutputOperand<hal_1_2::HalPolicy>(operation, 0, model); |
| 467 | |
| 468 | if (!output) |
| 469 | { |
| 470 | return Fail("%s: Could not read output 0", __func__); |
| 471 | } |
| 472 | |
Teresa Charlin | 8f6429d | 2019-10-01 13:10:15 +0100 | [diff] [blame] | 473 | const TensorInfo& inputInfo = input.GetTensorInfo(); |
| 474 | const TensorInfo& outputInfo = GetTensorInfoForOperand(*output); |
Aron Virginas-Tar | b7421e5 | 2019-07-26 13:14:39 +0100 | [diff] [blame] | 475 | |
| 476 | if (IsDynamicTensor(outputInfo)) |
| 477 | { |
| 478 | return Fail("%s: Dynamic output tensors are not supported", __func__); |
| 479 | } |
Aron Virginas-Tar | 24e699d | 2019-06-17 14:47:46 +0100 | [diff] [blame] | 480 | |
| 481 | // ArmNN does not currently support non-fixed weights or bias |
| 482 | // Find the shape of the weights tensor. In AndroidNN this will be [ 1, H, W, I * M ] |
| 483 | const Operand* weightsOperand = GetInputOperand<hal_1_2::HalPolicy>(operation, 1, model); |
| 484 | |
| 485 | if (weightsOperand == nullptr) |
| 486 | { |
| 487 | return Fail("%s: Operand is invalid", __func__); |
| 488 | } |
Teresa Charlin | 3b95960 | 2019-10-31 17:05:47 +0000 | [diff] [blame] | 489 | if ( weightsOperand->dimensions[0] != 1) |
| 490 | { |
| 491 | return Fail("%s: Invalid weights; for depthwise convolution, dimension 0 must be 1 but it is %i", |
| 492 | __func__, weightsOperand->dimensions[0] ); |
| 493 | } |
| 494 | |
Teresa Charlin | 8f6429d | 2019-10-01 13:10:15 +0100 | [diff] [blame] | 495 | DepthwiseConvolution2dDescriptor desc; |
| 496 | desc.m_DataLayout = DataLayout::NHWC; |
Aron Virginas-Tar | 24e699d | 2019-06-17 14:47:46 +0100 | [diff] [blame] | 497 | |
| 498 | // Determine whether padding is implicit or explicit |
| 499 | bool implicitPadding = operation.inputs.size() == 8 || |
| 500 | (operation.inputs.size() >= 9 && |
| 501 | GetInputOperand<hal_1_2::HalPolicy>(operation, 8, model)->type == OperandType::BOOL); |
| 502 | |
| 503 | // Look ahead to find the optional DataLayout, if present |
| 504 | const uint32_t dataLayoutFlagIndex = implicitPadding ? 8 : 11; |
| 505 | desc.m_DataLayout = OptionalDataLayout<hal_1_2::HalPolicy>(operation, dataLayoutFlagIndex, model, data); |
| 506 | |
| 507 | armnnUtils::DataLayoutIndexed dataLayoutIndexed(desc.m_DataLayout); |
| 508 | unsigned int channelsIndex = dataLayoutIndexed.GetChannelsIndex(); |
| 509 | unsigned int widthIndex = dataLayoutIndexed.GetWidthIndex(); |
| 510 | unsigned int heightIndex = dataLayoutIndexed.GetHeightIndex(); |
| 511 | |
| 512 | // Reinterpret weight data as [ H, W, I, M ] |
Teresa Charlin | 8f6429d | 2019-10-01 13:10:15 +0100 | [diff] [blame] | 513 | TensorShape weightsShape({ weightsOperand->dimensions[1], |
Aron Virginas-Tar | 24e699d | 2019-06-17 14:47:46 +0100 | [diff] [blame] | 514 | weightsOperand->dimensions[2], |
| 515 | inputInfo.GetShape()[channelsIndex], |
| 516 | weightsOperand->dimensions[3] / inputInfo.GetShape()[channelsIndex] }); |
| 517 | |
| 518 | // Swizzle weight data [ H, W, I, M ] -> [ M, I, H, W ] |
Teresa Charlin | 8f6429d | 2019-10-01 13:10:15 +0100 | [diff] [blame] | 519 | const PermutationVector HWIMToMIHW = { 2U, 3U, 1U, 0U }; |
Aron Virginas-Tar | 24e699d | 2019-06-17 14:47:46 +0100 | [diff] [blame] | 520 | |
| 521 | const ConstTensorPin weightsPin = |
| 522 | ConvertOperationInputToConstTensorPin<hal_1_2::HalPolicy>(operation, |
| 523 | 1, |
| 524 | model, |
| 525 | data, |
| 526 | HWIMToMIHW, |
| 527 | &weightsShape); |
| 528 | |
| 529 | // Bias is a 1D tensor |
| 530 | const ConstTensorPin biasPin = |
| 531 | ConvertOperationInputToConstTensorPin<hal_1_2::HalPolicy>(operation, 2, model, data); |
| 532 | |
| 533 | if (!weightsPin.IsValid()) |
| 534 | { |
| 535 | return Fail("%s: Operation has invalid weights", __func__); |
| 536 | } |
| 537 | |
| 538 | if (!biasPin.IsValid()) |
| 539 | { |
| 540 | return Fail("%s: Operation has invalid biases", __func__); |
| 541 | } |
| 542 | |
Teresa Charlin | 8f6429d | 2019-10-01 13:10:15 +0100 | [diff] [blame] | 543 | ConstTensor weights = weightsPin.GetConstTensor(); |
| 544 | ConstTensor bias = biasPin.GetConstTensor(); |
Aron Virginas-Tar | 24e699d | 2019-06-17 14:47:46 +0100 | [diff] [blame] | 545 | SanitizeBiasQuantizationScale(bias.GetInfo(), weights.GetInfo(), inputInfo); |
| 546 | |
| 547 | ActivationFn activation; |
| 548 | |
| 549 | if (implicitPadding) |
| 550 | { |
| 551 | android::nn::PaddingScheme paddingScheme; |
| 552 | if (!GetInputPaddingScheme<hal_1_2::HalPolicy>(operation, 3, paddingScheme, model, data) || |
| 553 | !GetInputScalar<hal_1_2::HalPolicy>(operation, 4, OperandType::INT32, desc.m_StrideX, model, data) || |
| 554 | !GetInputScalar<hal_1_2::HalPolicy>(operation, 5, OperandType::INT32, desc.m_StrideY, model, data) || |
| 555 | !GetInputActivationFunction<hal_1_2::HalPolicy>(operation, 7, activation, model, data) || |
| 556 | !GetOptionalConvolutionDilationParams<hal_1_2::HalPolicy>(operation, 9, desc, model, data)) |
| 557 | { |
| 558 | return Fail("%s: Operation has invalid inputs (implicit padding)", __func__); |
| 559 | } |
| 560 | |
| 561 | const uint32_t kernelX = weights.GetShape()[3]; |
| 562 | const uint32_t kernelY = weights.GetShape()[2]; |
| 563 | const uint32_t inputX = inputInfo.GetShape()[widthIndex]; |
| 564 | const uint32_t inputY = inputInfo.GetShape()[heightIndex]; |
| 565 | |
Mike Kelly | 86b36d4 | 2019-07-12 16:39:33 +0100 | [diff] [blame] | 566 | CalcPadding(inputX, kernelX, desc.m_StrideX, desc.m_DilationX, desc.m_PadLeft, desc.m_PadRight, paddingScheme); |
| 567 | CalcPadding(inputY, kernelY, desc.m_StrideY, desc.m_DilationY, desc.m_PadTop, desc.m_PadBottom, paddingScheme); |
Aron Virginas-Tar | 24e699d | 2019-06-17 14:47:46 +0100 | [diff] [blame] | 568 | } |
| 569 | else if (operation.inputs.size() >= 11) |
| 570 | { |
| 571 | // explicit padding |
| 572 | if (!GetInputScalar<hal_1_2::HalPolicy>(operation, 3, OperandType::INT32, desc.m_PadLeft, model, data) || |
| 573 | !GetInputScalar<hal_1_2::HalPolicy>(operation, 4, OperandType::INT32, desc.m_PadRight, model, data) || |
| 574 | !GetInputScalar<hal_1_2::HalPolicy>(operation, 5, OperandType::INT32, desc.m_PadTop, model, data) || |
| 575 | !GetInputScalar<hal_1_2::HalPolicy>(operation, 6, OperandType::INT32, desc.m_PadBottom, model, data) || |
| 576 | !GetInputScalar<hal_1_2::HalPolicy>(operation, 7, OperandType::INT32, desc.m_StrideX, model, data) || |
| 577 | !GetInputScalar<hal_1_2::HalPolicy>(operation, 8, OperandType::INT32, desc.m_StrideY, model, data) || |
| 578 | !GetInputActivationFunction<hal_1_2::HalPolicy>(operation, 10, activation, model, data) || |
| 579 | !GetOptionalConvolutionDilationParams<hal_1_2::HalPolicy>(operation, 12, desc, model, data)) |
| 580 | { |
| 581 | return Fail("%s: Operation has invalid inputs (explicit padding)", __func__); |
| 582 | } |
| 583 | } |
| 584 | else |
| 585 | { |
| 586 | return Fail("%s: Unsupported number of operation inputs", __func__); |
| 587 | } |
| 588 | |
| 589 | desc.m_BiasEnabled = true; |
Teresa Charlin | 8f6429d | 2019-10-01 13:10:15 +0100 | [diff] [blame] | 590 | Optional<TensorInfo> biases(bias.GetInfo()); |
Aron Virginas-Tar | 24e699d | 2019-06-17 14:47:46 +0100 | [diff] [blame] | 591 | |
Ferran Balaguer | d30093c | 2019-07-09 17:04:47 +0100 | [diff] [blame] | 592 | bool isSupported = false; |
| 593 | FORWARD_LAYER_SUPPORT_FUNC(__func__, |
| 594 | IsDepthwiseConvolutionSupported, |
| 595 | data.m_Backends, |
| 596 | isSupported, |
| 597 | inputInfo, |
| 598 | outputInfo, |
| 599 | desc, |
| 600 | weights.GetInfo(), |
| 601 | biases); |
Aron Virginas-Tar | 9fd3739 | 2019-07-15 18:04:32 +0100 | [diff] [blame] | 602 | |
Ferran Balaguer | d30093c | 2019-07-09 17:04:47 +0100 | [diff] [blame] | 603 | if (!isSupported) |
Aron Virginas-Tar | 24e699d | 2019-06-17 14:47:46 +0100 | [diff] [blame] | 604 | { |
| 605 | return false; |
| 606 | } |
| 607 | |
Teresa Charlin | 8f6429d | 2019-10-01 13:10:15 +0100 | [diff] [blame] | 608 | IConnectableLayer* startLayer = |
| 609 | data.m_Network->AddDepthwiseConvolution2dLayer(desc, weights, Optional<ConstTensor>(bias)); |
Aron Virginas-Tar | 9fd3739 | 2019-07-15 18:04:32 +0100 | [diff] [blame] | 610 | |
Aron Virginas-Tar | 24e699d | 2019-06-17 14:47:46 +0100 | [diff] [blame] | 611 | if (!startLayer) |
| 612 | { |
| 613 | return Fail("%s: AddDepthwiseConvolution2dLayer failed", __func__); |
| 614 | } |
| 615 | |
Teresa Charlin | 8f6429d | 2019-10-01 13:10:15 +0100 | [diff] [blame] | 616 | IConnectableLayer* endLayer = ProcessActivation(outputInfo, activation, startLayer, data); |
Aron Virginas-Tar | 24e699d | 2019-06-17 14:47:46 +0100 | [diff] [blame] | 617 | if (!endLayer) |
| 618 | { |
| 619 | return Fail("%s: ProcessActivation failed", __func__); |
| 620 | } |
| 621 | |
| 622 | input.Connect(startLayer->GetInputSlot(0)); |
| 623 | |
Aron Virginas-Tar | b7421e5 | 2019-07-26 13:14:39 +0100 | [diff] [blame] | 624 | return SetupAndTrackLayerOutputSlot<hal_1_2::HalPolicy>(operation, 0, *endLayer, model, data); |
Aron Virginas-Tar | 24e699d | 2019-06-17 14:47:46 +0100 | [diff] [blame] | 625 | } |
| 626 | |
Mike Kelly | 4627280 | 2019-08-14 17:00:48 +0100 | [diff] [blame] | 627 | bool HalPolicy::ConvertDequantize(const Operation& operation, const Model& model, ConversionData& data) |
| 628 | { |
| 629 | ALOGV("hal_1_2::HalPolicy::ConvertDequantize()"); |
Aron Virginas-Tar | 65a1b1d | 2019-11-15 15:59:51 +0000 | [diff] [blame] | 630 | |
| 631 | if (IsQSymmDequantizeForWeights(operation, model)) |
| 632 | { |
| 633 | // NOTE: QSymm8 weights are dequantized internally by the driver, |
| 634 | // therefore this type of Dequantize is implicitly supported |
| 635 | return true; |
| 636 | } |
| 637 | |
Mike Kelly | 4627280 | 2019-08-14 17:00:48 +0100 | [diff] [blame] | 638 | return ::ConvertDequantize<hal_1_2::HalPolicy>(operation, model, data); |
| 639 | } |
| 640 | |
| 641 | bool HalPolicy::ConvertDiv(const Operation& operation, const Model& model, ConversionData& data) |
| 642 | { |
| 643 | ALOGV("hal_1_2::HalPolicy::ConvertDiv()"); |
| 644 | return ::ConvertDiv<hal_1_2::HalPolicy>(operation, model, data); |
| 645 | } |
| 646 | |
josh minor | 00a963b | 2020-01-08 11:55:35 -0600 | [diff] [blame] | 647 | bool HalPolicy::ConvertElementwiseUnary(const Operation& operation, |
| 648 | const Model& model, |
| 649 | ConversionData& data, |
| 650 | UnaryOperation unaryOperation) |
| 651 | { |
| 652 | ALOGV("hal_1_2::HalPolicy::ConvertElementwiseUnary()"); |
| 653 | ALOGV("unaryOperation = %s", GetUnaryOperationAsCString(unaryOperation)); |
| 654 | |
| 655 | LayerInputHandle input = ConvertToLayerInputHandle<hal_1_2::HalPolicy>(operation, 0, model, data); |
| 656 | |
| 657 | if (!input.IsValid()) |
| 658 | { |
| 659 | return Fail("%s: Operation has invalid input", __func__); |
| 660 | } |
| 661 | |
| 662 | const Operand* output = GetOutputOperand<hal_1_2::HalPolicy>(operation, 0, model); |
| 663 | if (!output) |
| 664 | { |
| 665 | return Fail("%s: Could not read output 0", __func__); |
| 666 | } |
| 667 | |
| 668 | const TensorInfo& inputInfo = input.GetTensorInfo(); |
| 669 | const TensorInfo& outputInfo = GetTensorInfoForOperand(*output); |
| 670 | |
| 671 | if (IsDynamicTensor(outputInfo)) |
| 672 | { |
| 673 | return Fail("%s: Dynamic output tensors are not supported", __func__); |
| 674 | } |
| 675 | |
| 676 | ElementwiseUnaryDescriptor descriptor(unaryOperation); |
| 677 | |
| 678 | bool isSupported = false; |
| 679 | FORWARD_LAYER_SUPPORT_FUNC(__func__, |
| 680 | IsElementwiseUnarySupported, |
| 681 | data.m_Backends, |
| 682 | isSupported, |
| 683 | inputInfo, |
| 684 | outputInfo, |
| 685 | descriptor); |
| 686 | |
| 687 | if (!isSupported) |
| 688 | { |
| 689 | return false; |
| 690 | } |
| 691 | |
| 692 | IConnectableLayer* layer = data.m_Network->AddElementwiseUnaryLayer(descriptor); |
| 693 | assert(layer != nullptr); |
| 694 | |
| 695 | input.Connect(layer->GetInputSlot(0)); |
| 696 | |
| 697 | return SetupAndTrackLayerOutputSlot<hal_1_2::HalPolicy>(operation, 0, *layer, model, data); |
| 698 | } |
| 699 | |
Narumol Prangnawarat | 85f9654 | 2019-09-12 16:26:29 +0100 | [diff] [blame] | 700 | bool HalPolicy::ConvertExpandDims(const Operation& operation, const Model& model, ConversionData& data) |
| 701 | { |
| 702 | ALOGV("hal_1_2::HalPolicy::ConvertExpandDims()"); |
| 703 | |
| 704 | LayerInputHandle input = ConvertToLayerInputHandle<HalPolicy>(operation, 0, model, data); |
| 705 | |
| 706 | if (!input.IsValid()) |
| 707 | { |
| 708 | return Fail("%s: Operation has invalid input", __func__); |
| 709 | } |
| 710 | |
| 711 | const Operand* output = GetOutputOperand<HalPolicy>(operation, 0, model); |
| 712 | if (!output) |
| 713 | { |
| 714 | return Fail("%s: Operation has invalid output", __func__); |
| 715 | } |
| 716 | |
Teresa Charlin | 8f6429d | 2019-10-01 13:10:15 +0100 | [diff] [blame] | 717 | const TensorInfo& outputInfo = GetTensorInfoForOperand(*output); |
Narumol Prangnawarat | 85f9654 | 2019-09-12 16:26:29 +0100 | [diff] [blame] | 718 | if (IsDynamicTensor(outputInfo)) |
| 719 | { |
| 720 | return Fail("%s: Dynamic output tensors are not supported", __func__); |
| 721 | } |
| 722 | |
| 723 | int32_t axis; |
| 724 | if (!GetInputScalar<HalPolicy>(operation, 1, OperandType::INT32, axis, model, data)) |
| 725 | { |
| 726 | return Fail("%s: failed to get axis input value", __func__); |
| 727 | } |
| 728 | |
Teresa Charlin | 8f6429d | 2019-10-01 13:10:15 +0100 | [diff] [blame] | 729 | TensorShape targetShape; |
Narumol Prangnawarat | 85f9654 | 2019-09-12 16:26:29 +0100 | [diff] [blame] | 730 | |
| 731 | try |
| 732 | { |
| 733 | targetShape = armnnUtils::ExpandDims(input.GetTensorInfo().GetShape(), axis); |
| 734 | } |
| 735 | catch (const std::exception &e) |
| 736 | { |
| 737 | return Fail("%s: %s", __func__, e.what()); |
| 738 | } |
| 739 | |
| 740 | if (targetShape != outputInfo.GetShape()) |
| 741 | { |
| 742 | return Fail("%s: Shape of the output operand does not match the resolved expanded shape", __func__); |
| 743 | } |
| 744 | |
Teresa Charlin | 8f6429d | 2019-10-01 13:10:15 +0100 | [diff] [blame] | 745 | ReshapeDescriptor reshapeDescriptor; |
Narumol Prangnawarat | 85f9654 | 2019-09-12 16:26:29 +0100 | [diff] [blame] | 746 | reshapeDescriptor.m_TargetShape = targetShape; |
| 747 | |
| 748 | bool isSupported = false; |
| 749 | FORWARD_LAYER_SUPPORT_FUNC(__func__, |
| 750 | IsReshapeSupported, |
| 751 | data.m_Backends, |
| 752 | isSupported, |
| 753 | input.GetTensorInfo(), |
Kevin May | aed08ac | 2019-12-12 16:33:31 +0000 | [diff] [blame] | 754 | outputInfo, |
Narumol Prangnawarat | 85f9654 | 2019-09-12 16:26:29 +0100 | [diff] [blame] | 755 | reshapeDescriptor); |
| 756 | |
| 757 | if (!isSupported) |
| 758 | { |
| 759 | return false; |
| 760 | } |
| 761 | |
Teresa Charlin | 8f6429d | 2019-10-01 13:10:15 +0100 | [diff] [blame] | 762 | IConnectableLayer* layer = data.m_Network->AddReshapeLayer(reshapeDescriptor); |
Narumol Prangnawarat | 85f9654 | 2019-09-12 16:26:29 +0100 | [diff] [blame] | 763 | assert(layer != nullptr); |
| 764 | input.Connect(layer->GetInputSlot(0)); |
| 765 | |
| 766 | return SetupAndTrackLayerOutputSlot<HalPolicy>(operation, 0, *layer, model, data); |
| 767 | } |
| 768 | |
Mike Kelly | 4627280 | 2019-08-14 17:00:48 +0100 | [diff] [blame] | 769 | bool HalPolicy::ConvertFloor(const Operation& operation, const Model& model, ConversionData& data) |
| 770 | { |
| 771 | ALOGV("hal_1_2::HalPolicy::ConvertFloor()"); |
| 772 | return ::ConvertFloor<hal_1_2::HalPolicy>(operation, model, data); |
| 773 | } |
| 774 | |
| 775 | bool HalPolicy::ConvertFullyConnected(const Operation& operation, const Model& model, ConversionData& data) |
| 776 | { |
| 777 | ALOGV("hal_1_2::HalPolicy::ConvertFullyConnected()"); |
| 778 | return ::ConvertFullyConnected<hal_1_2::HalPolicy>(operation, model, data); |
| 779 | } |
| 780 | |
Teresa Charlin | 8f6429d | 2019-10-01 13:10:15 +0100 | [diff] [blame] | 781 | bool HalPolicy::ConvertGroupedConv2d(const Operation& operation, const Model& model, ConversionData& data) |
| 782 | { |
| 783 | ALOGV("hal_1_2::HalPolicy::ConvertGroupedConv2d()"); |
| 784 | |
| 785 | // |
| 786 | // Parse data |
| 787 | // |
| 788 | LayerInputHandle input = ConvertToLayerInputHandle<hal_1_2::HalPolicy>(operation, 0, model, data); |
| 789 | if (!input.IsValid()) |
| 790 | { |
| 791 | return Fail("%s: Operation has invalid inputs", __func__); |
| 792 | } |
| 793 | const TensorInfo& inputInfo = input.GetTensorInfo(); |
| 794 | |
| 795 | const Operand* output = GetOutputOperand<hal_1_2::HalPolicy>(operation, 0, model); |
| 796 | if (!output) |
| 797 | { |
| 798 | return Fail("%s: Could not read output 0", __func__); |
| 799 | } |
| 800 | const TensorInfo& outputInfo = GetTensorInfoForOperand(*output); |
| 801 | if (IsDynamicTensor(outputInfo)) |
| 802 | { |
| 803 | return Fail("%s: Dynamic output tensors are not supported", __func__); |
| 804 | } |
| 805 | |
| 806 | // Look ahead to determine data layout |
| 807 | DataLayout dataLayout = DataLayout::NHWC; |
| 808 | if (operation.inputs.size() == 12) |
| 809 | { |
| 810 | dataLayout = OptionalDataLayout<hal_1_2::HalPolicy>(operation, 11, model, data); |
| 811 | } |
| 812 | else |
| 813 | { |
| 814 | dataLayout = OptionalDataLayout<hal_1_2::HalPolicy>(operation, 8, model, data); |
| 815 | } |
| 816 | |
| 817 | // NOTE: |
| 818 | // NNAPI weights are always OHWI, i.e. [depth_out, filter_height, filter_width, depth_group], |
| 819 | // but Arm NN expects the filter's height and width indices to match the input's height and |
| 820 | // width indices so when the DataLayout is NCHW, we need to permute the weights to OIHW |
| 821 | const PermutationVector ohwiToOihw = { 0u, 2u, 3u, 1u }; |
| 822 | const ConstTensorPin weightsPin = (dataLayout == DataLayout::NCHW) ? |
| 823 | ConvertOperationInputToConstTensorPin<hal_1_2::HalPolicy>(operation, 1, model, data, ohwiToOihw) : |
| 824 | ConvertOperationInputToConstTensorPin<hal_1_2::HalPolicy>(operation, 1, model, data); |
| 825 | const ConstTensorPin biasesPin = |
| 826 | ConvertOperationInputToConstTensorPin<hal_1_2::HalPolicy>(operation, 2, model, data); |
| 827 | if (!weightsPin.IsValid() || !biasesPin.IsValid()) |
| 828 | { |
| 829 | return Fail("%s: Operation has invalid inputs", __func__); |
| 830 | } |
| 831 | |
| 832 | ConstTensor weights = weightsPin.GetConstTensor(); |
Aron Virginas-Tar | 60a346b | 2019-11-07 14:49:26 +0000 | [diff] [blame] | 833 | ConstTensor biases = biasesPin.GetConstTensor(); |
Teresa Charlin | 8f6429d | 2019-10-01 13:10:15 +0100 | [diff] [blame] | 834 | SanitizeBiasQuantizationScale(biases.GetInfo(), weights.GetInfo(), inputInfo); |
| 835 | |
| 836 | const TensorShape& inputShape = inputInfo.GetShape(); |
| 837 | const TensorShape& outputShape = outputInfo.GetShape(); |
| 838 | const TensorShape& weightsShape = weights.GetShape(); |
| 839 | const TensorShape& biasesShape = biases.GetShape(); |
| 840 | |
| 841 | armnnUtils::DataLayoutIndexed dataLayoutIndexed(dataLayout); |
| 842 | const unsigned int channelsIndex = dataLayoutIndexed.GetChannelsIndex(); |
| 843 | const unsigned int heightIndex = dataLayoutIndexed.GetHeightIndex(); |
| 844 | const unsigned int widthIndex = dataLayoutIndexed.GetWidthIndex(); |
| 845 | |
| 846 | Convolution2dDescriptor desc; |
| 847 | desc.m_DataLayout = dataLayout; |
| 848 | desc.m_BiasEnabled = true; |
| 849 | |
| 850 | int numGroups; |
| 851 | ActivationFn activation; |
| 852 | |
| 853 | if (operation.inputs.size() == 12) |
| 854 | { |
| 855 | if (!GetInputScalar<hal_1_2::HalPolicy>(operation, 3, OperandType::INT32, desc.m_PadLeft, model, data) || |
| 856 | !GetInputScalar<hal_1_2::HalPolicy>(operation, 4, OperandType::INT32, desc.m_PadRight, model, data) || |
| 857 | !GetInputScalar<hal_1_2::HalPolicy>(operation, 5, OperandType::INT32, desc.m_PadTop, model, data) || |
| 858 | !GetInputScalar<hal_1_2::HalPolicy>(operation, 6, OperandType::INT32, desc.m_PadBottom, model, data) || |
| 859 | !GetInputScalar<hal_1_2::HalPolicy>(operation, 7, OperandType::INT32, desc.m_StrideX, model, data) || |
| 860 | !GetInputScalar<hal_1_2::HalPolicy>(operation, 8, OperandType::INT32, desc.m_StrideY, model, data) || |
| 861 | !GetInputScalar<hal_1_2::HalPolicy>(operation, 9, OperandType::INT32, numGroups, model, data) || |
| 862 | !GetInputActivationFunction<hal_1_2::HalPolicy>(operation, 10, activation, model, data)) |
| 863 | { |
| 864 | return Fail("%s: Operation has invalid inputs (explicit padding)", __func__); |
| 865 | } |
| 866 | |
| 867 | } |
| 868 | else if (operation.inputs.size() == 9) |
| 869 | { |
| 870 | android::nn::PaddingScheme paddingScheme; |
| 871 | if (!GetInputPaddingScheme<hal_1_2::HalPolicy>(operation, 3, paddingScheme, model, data) || |
| 872 | !GetInputScalar<hal_1_2::HalPolicy>(operation, 4, OperandType::INT32, desc.m_StrideX, model, data) || |
| 873 | !GetInputScalar<hal_1_2::HalPolicy>(operation, 5, OperandType::INT32, desc.m_StrideY, model, data) || |
| 874 | !GetInputScalar<hal_1_2::HalPolicy>(operation, 6, OperandType::INT32, numGroups, model, data) || |
| 875 | !GetInputActivationFunction<hal_1_2::HalPolicy>(operation, 7, activation, model, data)) |
| 876 | { |
| 877 | return Fail("%s: Operation has invalid inputs (implicit padding)", __func__); |
| 878 | } |
| 879 | |
| 880 | const uint32_t inputX = inputInfo.GetShape()[widthIndex]; |
| 881 | const uint32_t inputY = inputInfo.GetShape()[heightIndex]; |
| 882 | |
| 883 | const uint32_t kernelX = weightsShape[widthIndex]; |
| 884 | const uint32_t kernelY = weightsShape[heightIndex]; |
| 885 | |
| 886 | CalcPadding(inputX, kernelX, desc.m_StrideX, desc.m_PadLeft, desc.m_PadRight, paddingScheme); |
| 887 | CalcPadding(inputY, kernelY, desc.m_StrideY, desc.m_PadTop, desc.m_PadBottom, paddingScheme); |
| 888 | } |
| 889 | else |
| 890 | { |
| 891 | return Fail("%s: Unsupported number of operation inputs", __func__); |
| 892 | } |
| 893 | |
| 894 | const unsigned int outputChannels = outputShape[channelsIndex]; |
| 895 | |
| 896 | const unsigned int channelsPerGroup = weightsShape[channelsIndex]; |
| 897 | const unsigned int channelMultiplier = outputChannels / numGroups; |
| 898 | |
| 899 | // |
| 900 | // Validate all relevant inputs |
| 901 | // |
| 902 | if (numGroups <= 0) |
| 903 | { |
| 904 | return Fail("%s: Number of groups must be greater than 0. Got: %d", __func__, numGroups); |
| 905 | } |
| 906 | |
| 907 | if (outputChannels % numGroups != 0u) |
| 908 | { |
| 909 | return Fail("%s: Output channels must be divisible by the number of groups", __func__); |
| 910 | } |
| 911 | |
| 912 | // |
| 913 | // Set up Splitter layer |
| 914 | // |
| 915 | unsigned int splitterDimSizes[4] = { inputShape[0], inputShape[1], inputShape[2], inputShape[3] }; |
| 916 | splitterDimSizes[channelsIndex] /= numGroups; // split in depth |
| 917 | |
| 918 | TensorInfo splitterOutputInfo(4, |
| 919 | splitterDimSizes, |
| 920 | inputInfo.GetDataType(), |
| 921 | inputInfo.GetQuantizationScale(), |
| 922 | inputInfo.GetQuantizationOffset()); |
| 923 | |
| 924 | std::vector<std::reference_wrapper<TensorInfo>> splitterOutputInfos(numGroups, std::ref(splitterOutputInfo)); |
| 925 | |
| 926 | ViewsDescriptor splitterDesc(numGroups); |
| 927 | for (unsigned int group = 0u; group < numGroups; ++group) |
| 928 | { |
| 929 | splitterDesc.SetViewOriginCoord(group, channelsIndex, splitterDimSizes[channelsIndex] * group); |
| 930 | for (unsigned int dimIdx = 0u; dimIdx < 4u; dimIdx++) |
| 931 | { |
| 932 | splitterDesc.SetViewSize(group, dimIdx, splitterDimSizes[dimIdx]); |
| 933 | } |
| 934 | } |
| 935 | |
| 936 | bool isSupported = false; |
| 937 | FORWARD_LAYER_SUPPORT_FUNC(__func__, |
| 938 | IsSplitterSupported, |
| 939 | data.m_Backends, |
| 940 | isSupported, |
| 941 | inputInfo, |
| 942 | splitterOutputInfos, |
| 943 | splitterDesc); |
| 944 | if (!isSupported) |
| 945 | { |
| 946 | return false; |
| 947 | } |
| 948 | |
| 949 | IConnectableLayer* splitterLayer = data.m_Network->AddSplitterLayer(splitterDesc); |
| 950 | if (!splitterLayer) |
| 951 | { |
| 952 | return Fail("%s: Failed to add SplitterLayer", __func__); |
| 953 | } |
| 954 | |
| 955 | input.Connect(splitterLayer->GetInputSlot(0)); |
| 956 | for (unsigned int group = 0u; group < splitterLayer->GetNumOutputSlots(); ++group) |
| 957 | { |
| 958 | splitterLayer->GetOutputSlot(group).SetTensorInfo(splitterOutputInfo); |
| 959 | } |
| 960 | |
| 961 | // |
| 962 | // Set up Convolution2d layers for each group |
| 963 | // |
Aron Virginas-Tar | 60a346b | 2019-11-07 14:49:26 +0000 | [diff] [blame] | 964 | |
| 965 | // Set up group tensor shapes |
Teresa Charlin | 8f6429d | 2019-10-01 13:10:15 +0100 | [diff] [blame] | 966 | TensorShape groupInputShape(inputShape); |
| 967 | groupInputShape[channelsIndex] = channelsPerGroup; |
| 968 | |
| 969 | TensorShape groupOutputShape(outputShape); |
| 970 | groupOutputShape[channelsIndex] = 1; |
| 971 | |
| 972 | TensorShape groupWeightsShape(weightsShape); |
| 973 | groupWeightsShape[0] /= channelMultiplier * numGroups; |
| 974 | |
| 975 | TensorShape groupBiasesShape({ 1 }); |
| 976 | |
Aron Virginas-Tar | 60a346b | 2019-11-07 14:49:26 +0000 | [diff] [blame] | 977 | // Set up group tensor infos |
| 978 | TensorInfo groupInputInfo(inputInfo); |
| 979 | groupInputInfo.SetShape(groupInputShape); |
| 980 | |
| 981 | const TensorInfo& weightsInfo = weights.GetInfo(); |
| 982 | TensorInfo groupWeightsInfo(weightsInfo); |
| 983 | groupWeightsInfo.SetShape(groupWeightsShape); |
| 984 | |
| 985 | const TensorInfo& biasesInfo = biases.GetInfo(); |
| 986 | TensorInfo groupBiasesInfo(biasesInfo); |
| 987 | groupBiasesInfo.SetShape(groupBiasesShape); |
| 988 | |
| 989 | TensorInfo groupOutputInfo(outputInfo); |
| 990 | groupOutputInfo.SetShape(groupOutputShape); |
Teresa Charlin | 8f6429d | 2019-10-01 13:10:15 +0100 | [diff] [blame] | 991 | |
| 992 | const unsigned int weightsDataTypeSize = GetDataTypeSize(groupWeightsInfo.GetDataType()); |
| 993 | const unsigned int biasesDataTypeSize = GetDataTypeSize(groupBiasesInfo.GetDataType()); |
| 994 | |
Aron Virginas-Tar | 60a346b | 2019-11-07 14:49:26 +0000 | [diff] [blame] | 995 | std::vector<IConnectableLayer*> convLayers(numGroups * channelMultiplier, nullptr); |
Teresa Charlin | 8f6429d | 2019-10-01 13:10:15 +0100 | [diff] [blame] | 996 | for (unsigned int group = 0u; group < numGroups; ++group) |
| 997 | { |
| 998 | for (unsigned int m = 0u; m < channelMultiplier; ++m) |
| 999 | { |
| 1000 | auto index = group * channelMultiplier + m; |
| 1001 | |
| 1002 | const unsigned int weightsDataOffset = groupWeightsShape.GetNumElements() * index * weightsDataTypeSize; |
| 1003 | const unsigned int biasesDataOffset = groupBiasesShape.GetNumElements() * index * biasesDataTypeSize; |
| 1004 | |
Aron Virginas-Tar | 60a346b | 2019-11-07 14:49:26 +0000 | [diff] [blame] | 1005 | if (weightsInfo.HasPerAxisQuantization()) |
| 1006 | { |
| 1007 | // Extract per-axis quantization scales for group weights |
| 1008 | const std::vector<float>& weightsQuantScales = weightsInfo.GetQuantizationScales(); |
| 1009 | groupWeightsInfo.SetQuantizationScales( |
| 1010 | std::vector<float>(weightsQuantScales.begin() + index, |
| 1011 | weightsQuantScales.begin() + index + groupWeightsShape[0])); |
| 1012 | |
| 1013 | // Extract per-axis quantization scales for group biases |
| 1014 | const std::vector<float>& biasesQuantScales = biasesInfo.GetQuantizationScales(); |
| 1015 | groupBiasesInfo.SetQuantizationScales( |
| 1016 | std::vector<float>(biasesQuantScales.begin() + index, |
| 1017 | biasesQuantScales.begin() + index + groupWeightsShape[0])); |
| 1018 | } |
| 1019 | |
Teresa Charlin | 8f6429d | 2019-10-01 13:10:15 +0100 | [diff] [blame] | 1020 | // Extract weights and biases data for current group convolution |
| 1021 | ConstTensor groupWeights(groupWeightsInfo, |
| 1022 | static_cast<const void *>(reinterpret_cast<const char *>(weights.GetMemoryArea()) + |
| 1023 | weightsDataOffset)); |
| 1024 | ConstTensor groupBiases(groupBiasesInfo, |
| 1025 | static_cast<const void *>(reinterpret_cast<const char *>(biases.GetMemoryArea()) + |
| 1026 | biasesDataOffset)); |
| 1027 | |
| 1028 | isSupported = false; |
| 1029 | FORWARD_LAYER_SUPPORT_FUNC(__func__, |
| 1030 | IsConvolution2dSupported, |
| 1031 | data.m_Backends, |
| 1032 | isSupported, |
| 1033 | groupInputInfo, |
| 1034 | groupOutputInfo, |
| 1035 | desc, |
| 1036 | groupWeightsInfo, |
| 1037 | Optional<TensorInfo>(groupBiasesInfo)); |
| 1038 | if (!isSupported) |
| 1039 | { |
| 1040 | return false; |
| 1041 | } |
| 1042 | |
| 1043 | IConnectableLayer *convLayer = |
| 1044 | data.m_Network->AddConvolution2dLayer(desc, groupWeights, Optional<ConstTensor>(groupBiases)); |
| 1045 | if (!convLayer) |
| 1046 | { |
| 1047 | return Fail("%s: AddConvolution2dLayer failed", __func__); |
| 1048 | } |
| 1049 | |
| 1050 | splitterLayer->GetOutputSlot(group).Connect(convLayer->GetInputSlot(0)); |
| 1051 | convLayer->GetOutputSlot(0).SetTensorInfo(groupOutputInfo); |
| 1052 | |
| 1053 | convLayers[index] = convLayer; |
| 1054 | } |
| 1055 | } |
| 1056 | |
| 1057 | // |
| 1058 | // Set up Concat layer |
| 1059 | // |
| 1060 | ConcatDescriptor concatDescriptor(outputInfo.GetShape()[channelsIndex]); |
| 1061 | for (unsigned int group = 0u; group < numGroups; ++group) |
| 1062 | { |
| 1063 | for (unsigned int m = 0u; m < channelMultiplier; ++m) |
| 1064 | { |
| 1065 | auto index = group * channelMultiplier + m; |
| 1066 | concatDescriptor.SetViewOriginCoord(index, channelsIndex, index); |
| 1067 | concatDescriptor.SetConcatAxis(channelsIndex); |
| 1068 | } |
| 1069 | } |
| 1070 | |
| 1071 | isSupported = false; |
| 1072 | FORWARD_LAYER_SUPPORT_FUNC(__func__, |
| 1073 | IsConcatSupported, |
| 1074 | data.m_Backends, |
| 1075 | isSupported, |
| 1076 | std::vector<const TensorInfo*>(numGroups * channelMultiplier, &groupOutputInfo), |
| 1077 | outputInfo, |
| 1078 | concatDescriptor); |
| 1079 | if (!isSupported) |
| 1080 | { |
| 1081 | return false; |
| 1082 | } |
| 1083 | |
| 1084 | IConnectableLayer* concatLayer = data.m_Network->AddConcatLayer(concatDescriptor); |
| 1085 | if (!concatLayer) |
| 1086 | { |
| 1087 | return Fail("%s: AddConcatLayer failed", __func__); |
| 1088 | } |
| 1089 | |
| 1090 | for (unsigned int group = 0u; group < numGroups; ++group) |
| 1091 | { |
| 1092 | for (unsigned int m = 0u; m < channelMultiplier; ++m) |
| 1093 | { |
| 1094 | auto index = group * channelMultiplier + m; |
| 1095 | convLayers[index]->GetOutputSlot(0).Connect(concatLayer->GetInputSlot(index)); |
| 1096 | } |
| 1097 | } |
| 1098 | concatLayer->GetOutputSlot(0).SetTensorInfo(outputInfo); |
| 1099 | |
| 1100 | // |
| 1101 | // Set up Activation layer (if it is set) |
| 1102 | // |
| 1103 | IConnectableLayer* endLayer = ProcessActivation(outputInfo, activation, concatLayer, data); |
| 1104 | if (!endLayer) |
| 1105 | { |
| 1106 | return Fail("%s: ProcessActivation failed", __func__); |
| 1107 | } |
| 1108 | |
| 1109 | return SetupAndTrackLayerOutputSlot<HalPolicy>(operation, 0, *endLayer, model, data); |
| 1110 | } |
| 1111 | |
Aron Virginas-Tar | a2a7380 | 2019-10-09 15:30:40 +0100 | [diff] [blame] | 1112 | bool HalPolicy::ConvertInstanceNormalization(const Operation& operation, const Model& model, ConversionData& data) |
| 1113 | { |
| 1114 | ALOGV("hal_1_2::HalPolicy::ConvertInstanceNormalization()"); |
| 1115 | |
| 1116 | LayerInputHandle input = ConvertToLayerInputHandle<hal_1_2::HalPolicy>(operation, 0, model, data); |
| 1117 | if (!input.IsValid()) |
| 1118 | { |
| 1119 | return Fail("%s: Operation has an invalid input 0", __func__); |
| 1120 | } |
| 1121 | |
| 1122 | const Operand* output = GetOutputOperand<HalPolicy>(operation, 0, model); |
| 1123 | if (!output) |
| 1124 | { |
| 1125 | return Fail("%s: Operation has an invalid output", __func__); |
| 1126 | } |
| 1127 | |
| 1128 | const TensorInfo& outputInfo = GetTensorInfoForOperand(*output); |
| 1129 | if (IsDynamicTensor(outputInfo)) |
| 1130 | { |
| 1131 | return Fail("%s: Dynamic output tensors are not supported", __func__); |
| 1132 | } |
| 1133 | |
| 1134 | // Determine data type of input tensor |
| 1135 | OperandType inputType; |
| 1136 | if (!GetOperandType<hal_1_2::HalPolicy>(operation, 0, model, inputType)) |
| 1137 | { |
| 1138 | return Fail("%s: Operation has invalid inputs", __func__); |
| 1139 | } |
| 1140 | |
| 1141 | InstanceNormalizationDescriptor desc; |
| 1142 | |
| 1143 | // Read gamma, beta & epsilon |
| 1144 | if (inputType == OperandType::TENSOR_FLOAT16) |
| 1145 | { |
| 1146 | Half fp16Gamma; |
| 1147 | Half fp16Beta; |
| 1148 | Half fp16Epsilon; |
| 1149 | |
| 1150 | if (!GetInputScalar<hal_1_2::HalPolicy>(operation, 1, OperandType::FLOAT16, fp16Gamma, model, data) || |
| 1151 | !GetInputScalar<hal_1_2::HalPolicy>(operation, 2, OperandType::FLOAT16, fp16Beta, model, data) || |
| 1152 | !GetInputScalar<hal_1_2::HalPolicy>(operation, 3, OperandType::FLOAT16, fp16Epsilon, model, data)) |
| 1153 | { |
| 1154 | return Fail("%s: Operation has invalid inputs (FLOAT16)", __func__); |
| 1155 | } |
| 1156 | |
| 1157 | desc.m_Gamma = static_cast<float>(fp16Gamma); |
| 1158 | desc.m_Beta = static_cast<float>(fp16Beta); |
| 1159 | desc.m_Eps = static_cast<float>(fp16Epsilon); |
| 1160 | } |
| 1161 | else if (inputType == OperandType::TENSOR_FLOAT32) |
| 1162 | { |
| 1163 | if (!GetInputScalar<hal_1_2::HalPolicy>(operation, 1, OperandType::FLOAT32, desc.m_Gamma, model, data) || |
| 1164 | !GetInputScalar<hal_1_2::HalPolicy>(operation, 2, OperandType::FLOAT32, desc.m_Beta, model, data) || |
| 1165 | !GetInputScalar<hal_1_2::HalPolicy>(operation, 3, OperandType::FLOAT32, desc.m_Eps, model, data)) |
| 1166 | { |
| 1167 | return Fail("%s: Operation has invalid inputs (FLOAT32)", __func__); |
| 1168 | } |
| 1169 | } |
| 1170 | else |
| 1171 | { |
| 1172 | return Fail("%s: Unsupported input tensor type: %d", __func__, inputType); |
| 1173 | } |
| 1174 | |
| 1175 | desc.m_DataLayout = OptionalDataLayout<hal_1_2::HalPolicy>(operation, 4, model, data); |
| 1176 | |
| 1177 | bool isSupported = false; |
| 1178 | FORWARD_LAYER_SUPPORT_FUNC(__func__, |
| 1179 | IsInstanceNormalizationSupported, |
| 1180 | data.m_Backends, |
| 1181 | isSupported, |
| 1182 | input.GetTensorInfo(), |
| 1183 | outputInfo, |
| 1184 | desc); |
| 1185 | if (!isSupported) |
| 1186 | { |
| 1187 | return false; |
| 1188 | } |
| 1189 | |
| 1190 | IConnectableLayer* layer = data.m_Network->AddInstanceNormalizationLayer(desc); |
| 1191 | input.Connect(layer->GetInputSlot(0)); |
| 1192 | |
| 1193 | return SetupAndTrackLayerOutputSlot<hal_1_2::HalPolicy>(operation, 0, *layer, model, data); |
| 1194 | } |
| 1195 | |
Mike Kelly | 4627280 | 2019-08-14 17:00:48 +0100 | [diff] [blame] | 1196 | bool HalPolicy::ConvertL2Normalization(const Operation& operation, const Model& model, ConversionData& data) |
| 1197 | { |
| 1198 | ALOGV("hal_1_2::HalPolicy::ConvertL2Normalization()"); |
| 1199 | return ::ConvertL2Normalization<hal_1_2::HalPolicy>(operation, model, data); |
| 1200 | } |
| 1201 | |
Sadik Armagan | 15d63e2 | 2019-07-26 16:59:35 +0100 | [diff] [blame] | 1202 | bool HalPolicy::ConvertL2Pool2d(const Operation& operation, const Model& model, ConversionData& data) |
| 1203 | { |
| 1204 | ALOGV("hal_1_2::HalPolicy::ConvertL2Pool2d()"); |
Teresa Charlin | 8f6429d | 2019-10-01 13:10:15 +0100 | [diff] [blame] | 1205 | return ConvertPooling2d<hal_1_2::HalPolicy>(operation, __func__, PoolingAlgorithm::L2, model, data); |
Sadik Armagan | 15d63e2 | 2019-07-26 16:59:35 +0100 | [diff] [blame] | 1206 | } |
| 1207 | |
Mike Kelly | 4627280 | 2019-08-14 17:00:48 +0100 | [diff] [blame] | 1208 | bool HalPolicy::ConvertLocalResponseNormalization(const Operation& operation, |
| 1209 | const Model& model, |
| 1210 | ConversionData& data) |
| 1211 | { |
| 1212 | ALOGV("hal_1_2::HalPolicy::ConvertLocalResponseNormalization()"); |
| 1213 | return ::ConvertLocalResponseNormalization<hal_1_2::HalPolicy>(operation, model, data); |
| 1214 | } |
| 1215 | |
| 1216 | bool HalPolicy::ConvertLogistic(const Operation& operation, const Model& model, ConversionData& data) |
| 1217 | { |
| 1218 | ALOGV("hal_1_2::HalPolicy::ConvertLogistic()"); |
| 1219 | return ::ConvertLogistic<hal_1_2::HalPolicy>(operation, model, data); |
| 1220 | } |
| 1221 | |
Aron Virginas-Tar | 75e6779 | 2019-10-15 13:33:03 +0100 | [diff] [blame] | 1222 | bool HalPolicy::ConvertLogSoftmax(const Operation& operation, const Model& model, ConversionData& data) |
| 1223 | { |
| 1224 | ALOGV("hal_1_2::HalPolicy::ConvertLogSoftmax()"); |
| 1225 | |
| 1226 | LayerInputHandle input = ConvertToLayerInputHandle<hal_1_2::HalPolicy>(operation, 0, model, data); |
| 1227 | if (!input.IsValid()) |
| 1228 | { |
| 1229 | return Fail("%s: Failed to read input 0", __func__); |
| 1230 | } |
| 1231 | |
| 1232 | const Operand* output = GetOutputOperand<hal_1_2::HalPolicy>(operation, 0, model); |
| 1233 | if (!output) |
| 1234 | { |
| 1235 | return Fail("%s: Failed to read output", __func__); |
| 1236 | } |
| 1237 | |
| 1238 | const TensorInfo& outputInfo = GetTensorInfoForOperand(*output); |
| 1239 | if (IsDynamicTensor(outputInfo)) |
| 1240 | { |
| 1241 | return Fail("%s: Dynamic output tensors are not supported", __func__); |
| 1242 | } |
| 1243 | |
| 1244 | // Determine data type of input tensor |
| 1245 | OperandType inputType; |
| 1246 | if (!GetOperandType<hal_1_2::HalPolicy>(operation, 0, model, inputType)) |
| 1247 | { |
| 1248 | return Fail("%s: Operation has invalid inputs", __func__); |
| 1249 | } |
| 1250 | |
| 1251 | LogSoftmaxDescriptor descriptor; |
| 1252 | |
| 1253 | // Read beta |
| 1254 | if (inputType == OperandType::TENSOR_FLOAT16) |
| 1255 | { |
| 1256 | Half fp16Beta; |
| 1257 | if (!GetInputScalar<hal_1_2::HalPolicy>(operation, 1, OperandType::FLOAT16, fp16Beta, model, data)) |
| 1258 | { |
| 1259 | return Fail("%s: Failed to read input 1 (FLOAT16)", __func__); |
| 1260 | } |
| 1261 | |
| 1262 | descriptor.m_Beta = static_cast<float>(fp16Beta); |
| 1263 | } |
| 1264 | else if (inputType == OperandType::TENSOR_FLOAT32) |
| 1265 | { |
| 1266 | if (!GetInputScalar<hal_1_2::HalPolicy>(operation, 1, OperandType::FLOAT32, descriptor.m_Beta, model, data)) |
| 1267 | { |
| 1268 | return Fail("%s: Failed to read input 1 (FLOAT32)", __func__); |
| 1269 | } |
| 1270 | } |
| 1271 | else |
| 1272 | { |
| 1273 | return Fail("%s: Unsupported input tensor type: %d", __func__, inputType); |
| 1274 | } |
| 1275 | |
| 1276 | // Read axis |
| 1277 | if (!GetInputInt32<hal_1_2::HalPolicy>(operation, 2, descriptor.m_Axis, model, data)) |
| 1278 | { |
| 1279 | return Fail("%s: Failed to read input 2", __func__); |
| 1280 | } |
| 1281 | |
| 1282 | bool isSupported = false; |
| 1283 | FORWARD_LAYER_SUPPORT_FUNC(__func__, |
| 1284 | IsLogSoftmaxSupported, |
| 1285 | data.m_Backends, |
| 1286 | isSupported, |
| 1287 | input.GetTensorInfo(), |
| 1288 | outputInfo, |
| 1289 | descriptor); |
| 1290 | if (!isSupported) |
| 1291 | { |
| 1292 | return false; |
| 1293 | } |
| 1294 | |
Aron Virginas-Tar | 3e0982b | 2019-10-29 14:25:09 +0000 | [diff] [blame] | 1295 | IConnectableLayer* layer = data.m_Network->AddLogSoftmaxLayer(descriptor); |
Aron Virginas-Tar | 75e6779 | 2019-10-15 13:33:03 +0100 | [diff] [blame] | 1296 | if (!layer) |
| 1297 | { |
| 1298 | return Fail("%s: AddLogSoftmaxLayer() returned nullptr", __func__); |
| 1299 | } |
| 1300 | |
| 1301 | input.Connect(layer->GetInputSlot(0)); |
| 1302 | |
| 1303 | return SetupAndTrackLayerOutputSlot<HalPolicy>(operation, 0, *layer, model, data); |
| 1304 | } |
| 1305 | |
Sadik Armagan | 15d63e2 | 2019-07-26 16:59:35 +0100 | [diff] [blame] | 1306 | bool HalPolicy::ConvertMaxPool2d(const Operation& operation, const Model& model, ConversionData& data) |
| 1307 | { |
| 1308 | ALOGV("hal_1_2::HalPolicy::ConvertMaxPool2d()"); |
Teresa Charlin | 8f6429d | 2019-10-01 13:10:15 +0100 | [diff] [blame] | 1309 | return ConvertPooling2d<hal_1_2::HalPolicy>(operation, __func__, PoolingAlgorithm::Max, model, data); |
Sadik Armagan | 15d63e2 | 2019-07-26 16:59:35 +0100 | [diff] [blame] | 1310 | } |
| 1311 | |
Narumol Prangnawarat | 95b1ef6 | 2019-07-15 12:02:20 +0100 | [diff] [blame] | 1312 | bool HalPolicy::ConvertMaximum(const Operation& operation, const Model& model, ConversionData& data) |
| 1313 | { |
Aron Virginas-Tar | 29404fb | 2019-07-24 13:55:31 +0100 | [diff] [blame] | 1314 | ALOGV("hal_1_2::HalPolicy::ConvertMaximum()"); |
| 1315 | |
Narumol Prangnawarat | 95b1ef6 | 2019-07-15 12:02:20 +0100 | [diff] [blame] | 1316 | LayerInputHandle input0 = ConvertToLayerInputHandle<hal_1_2::HalPolicy>(operation, 0, model, data); |
| 1317 | LayerInputHandle input1 = ConvertToLayerInputHandle<hal_1_2::HalPolicy>(operation, 1, model, data); |
| 1318 | |
| 1319 | if (!input0.IsValid() || !input1.IsValid()) |
| 1320 | { |
| 1321 | return Fail("%s: Operation has invalid inputs", __func__); |
| 1322 | } |
| 1323 | |
| 1324 | const Operand* outputOperand = GetOutputOperand<hal_1_2::HalPolicy>(operation, 0, model); |
| 1325 | if (!outputOperand) |
| 1326 | { |
| 1327 | return Fail("%s: Could not read output", __func__); |
| 1328 | } |
| 1329 | |
Teresa Charlin | 8f6429d | 2019-10-01 13:10:15 +0100 | [diff] [blame] | 1330 | const TensorInfo& outInfo = GetTensorInfoForOperand(*outputOperand); |
Aron Virginas-Tar | 573a8fa | 2019-07-23 14:01:37 +0100 | [diff] [blame] | 1331 | if (IsDynamicTensor(outInfo)) |
Narumol Prangnawarat | 95b1ef6 | 2019-07-15 12:02:20 +0100 | [diff] [blame] | 1332 | { |
Aron Virginas-Tar | b7421e5 | 2019-07-26 13:14:39 +0100 | [diff] [blame] | 1333 | return Fail("%s: Dynamic output tensors are not supported", __func__); |
Narumol Prangnawarat | 95b1ef6 | 2019-07-15 12:02:20 +0100 | [diff] [blame] | 1334 | } |
| 1335 | |
Aron Virginas-Tar | d759323 | 2019-07-16 13:17:06 +0100 | [diff] [blame] | 1336 | bool isSupported = false; |
| 1337 | FORWARD_LAYER_SUPPORT_FUNC(__func__, |
| 1338 | IsMaximumSupported, |
| 1339 | data.m_Backends, |
| 1340 | isSupported, |
| 1341 | input0.GetTensorInfo(), |
| 1342 | input1.GetTensorInfo(), |
| 1343 | outInfo); |
| 1344 | |
| 1345 | if (!isSupported) |
Narumol Prangnawarat | 95b1ef6 | 2019-07-15 12:02:20 +0100 | [diff] [blame] | 1346 | { |
| 1347 | return false; |
| 1348 | } |
| 1349 | |
Teresa Charlin | 8f6429d | 2019-10-01 13:10:15 +0100 | [diff] [blame] | 1350 | IConnectableLayer* layer = data.m_Network->AddMaximumLayer(); |
Narumol Prangnawarat | 95b1ef6 | 2019-07-15 12:02:20 +0100 | [diff] [blame] | 1351 | assert(layer != nullptr); |
Derek Lamberti | 6fd4ceb | 2019-12-19 15:45:35 +0000 | [diff] [blame] | 1352 | bool isReshapeSupported = BroadcastTensor(input0, input1, layer, data); |
Sadik Armagan | 64b19b5 | 2019-08-19 09:49:58 +0100 | [diff] [blame] | 1353 | if (!isReshapeSupported) |
| 1354 | { |
| 1355 | return false; |
| 1356 | } |
Narumol Prangnawarat | 95b1ef6 | 2019-07-15 12:02:20 +0100 | [diff] [blame] | 1357 | |
Aron Virginas-Tar | b7421e5 | 2019-07-26 13:14:39 +0100 | [diff] [blame] | 1358 | return SetupAndTrackLayerOutputSlot<hal_1_2::HalPolicy>(operation, 0, *layer, model, data); |
Narumol Prangnawarat | 95b1ef6 | 2019-07-15 12:02:20 +0100 | [diff] [blame] | 1359 | } |
| 1360 | |
Mike Kelly | 4627280 | 2019-08-14 17:00:48 +0100 | [diff] [blame] | 1361 | bool HalPolicy::ConvertMean(const Operation& operation, const Model& model, ConversionData& data) |
| 1362 | { |
| 1363 | ALOGV("hal_1_2::HalPolicy::ConvertMean()"); |
| 1364 | return ::ConvertMean<hal_1_2::HalPolicy>(operation, model, data); |
| 1365 | } |
| 1366 | |
Ellen Norris-Thompson | 1cb29aa | 2019-07-11 17:27:37 +0100 | [diff] [blame] | 1367 | bool HalPolicy::ConvertMinimum(const Operation& operation, const Model& model, ConversionData& data) |
| 1368 | { |
Aron Virginas-Tar | 29404fb | 2019-07-24 13:55:31 +0100 | [diff] [blame] | 1369 | ALOGV("hal_1_2::HalPolicy::ConvertMinimum()"); |
| 1370 | |
Ellen Norris-Thompson | 1cb29aa | 2019-07-11 17:27:37 +0100 | [diff] [blame] | 1371 | LayerInputHandle input0 = ConvertToLayerInputHandle<hal_1_2::HalPolicy>(operation, 0, model, data); |
| 1372 | LayerInputHandle input1 = ConvertToLayerInputHandle<hal_1_2::HalPolicy>(operation, 1, model, data); |
| 1373 | |
| 1374 | if (!input0.IsValid() || !input1.IsValid()) |
| 1375 | { |
| 1376 | return Fail("%s: Operation has invalid inputs", __func__); |
| 1377 | } |
| 1378 | |
| 1379 | const Operand* output = GetOutputOperand<hal_1_2::HalPolicy>(operation, 0, model); |
| 1380 | if (!output) |
| 1381 | { |
| 1382 | return Fail("%s: Could not read output 0", __func__); |
| 1383 | } |
| 1384 | |
Teresa Charlin | 8f6429d | 2019-10-01 13:10:15 +0100 | [diff] [blame] | 1385 | const TensorInfo& outputInfo = GetTensorInfoForOperand(*output); |
Aron Virginas-Tar | 573a8fa | 2019-07-23 14:01:37 +0100 | [diff] [blame] | 1386 | if (IsDynamicTensor(outputInfo)) |
Ellen Norris-Thompson | 1cb29aa | 2019-07-11 17:27:37 +0100 | [diff] [blame] | 1387 | { |
Aron Virginas-Tar | b7421e5 | 2019-07-26 13:14:39 +0100 | [diff] [blame] | 1388 | return Fail("%s: Dynamic output tensors are not supported", __func__); |
Ellen Norris-Thompson | 1cb29aa | 2019-07-11 17:27:37 +0100 | [diff] [blame] | 1389 | } |
| 1390 | |
| 1391 | bool isSupported = false; |
| 1392 | FORWARD_LAYER_SUPPORT_FUNC(__func__, |
| 1393 | IsMinimumSupported, |
| 1394 | data.m_Backends, |
| 1395 | isSupported, |
| 1396 | input0.GetTensorInfo(), |
| 1397 | input1.GetTensorInfo(), |
| 1398 | outputInfo); |
| 1399 | |
| 1400 | if (!isSupported) |
| 1401 | { |
| 1402 | return false; |
| 1403 | } |
| 1404 | |
Teresa Charlin | 8f6429d | 2019-10-01 13:10:15 +0100 | [diff] [blame] | 1405 | IConnectableLayer* const layer = data.m_Network->AddMinimumLayer(); |
Ellen Norris-Thompson | 1cb29aa | 2019-07-11 17:27:37 +0100 | [diff] [blame] | 1406 | assert(layer != nullptr); |
Derek Lamberti | 6fd4ceb | 2019-12-19 15:45:35 +0000 | [diff] [blame] | 1407 | bool isReshapeSupported = BroadcastTensor(input0, input1, layer, data); |
Sadik Armagan | 64b19b5 | 2019-08-19 09:49:58 +0100 | [diff] [blame] | 1408 | if (!isReshapeSupported) |
| 1409 | { |
| 1410 | return false; |
| 1411 | } |
Ellen Norris-Thompson | 1cb29aa | 2019-07-11 17:27:37 +0100 | [diff] [blame] | 1412 | |
Aron Virginas-Tar | b7421e5 | 2019-07-26 13:14:39 +0100 | [diff] [blame] | 1413 | return SetupAndTrackLayerOutputSlot<hal_1_2::HalPolicy>(operation, 0, *layer, model, data); |
Ellen Norris-Thompson | 1cb29aa | 2019-07-11 17:27:37 +0100 | [diff] [blame] | 1414 | } |
| 1415 | |
Mike Kelly | 4627280 | 2019-08-14 17:00:48 +0100 | [diff] [blame] | 1416 | bool HalPolicy::ConvertMul(const Operation& operation, const Model& model, ConversionData& data) |
| 1417 | { |
| 1418 | ALOGV("hal_1_2::HalPolicy::ConvertMul()"); |
| 1419 | return ::ConvertMul<hal_1_2::HalPolicy>(operation, model, data); |
| 1420 | } |
| 1421 | |
Aron Virginas-Tar | c921f6b | 2019-07-25 10:14:33 +0100 | [diff] [blame] | 1422 | bool HalPolicy::ConvertPad(const Operation& operation, const Model& model, ConversionData& data) |
| 1423 | { |
| 1424 | ALOGV("hal_1_2::HalPolicy::ConvertPad()"); |
| 1425 | return ::ConvertPad<hal_1_2::HalPolicy>(operation, model, data); |
| 1426 | } |
| 1427 | |
Aron Virginas-Tar | cb8ac84 | 2019-07-05 15:47:07 +0100 | [diff] [blame] | 1428 | bool HalPolicy::ConvertPadV2(const Operation& operation, const Model& model, ConversionData& data) |
| 1429 | { |
Aron Virginas-Tar | 29404fb | 2019-07-24 13:55:31 +0100 | [diff] [blame] | 1430 | ALOGV("hal_1_2::HalPolicy::ConvertPadV2()"); |
| 1431 | |
Aron Virginas-Tar | cb8ac84 | 2019-07-05 15:47:07 +0100 | [diff] [blame] | 1432 | LayerInputHandle input = ConvertToLayerInputHandle<hal_1_2::HalPolicy>(operation, 0, model, data); |
| 1433 | if (!input.IsValid()) |
| 1434 | { |
| 1435 | return Fail("%s: Could not read input 0", __func__); |
| 1436 | } |
| 1437 | |
Aron Virginas-Tar | 366e0a6 | 2019-07-10 13:01:41 +0100 | [diff] [blame] | 1438 | const Operand* output = GetOutputOperand<hal_1_2::HalPolicy>(operation, 0, model); |
| 1439 | if (!output) |
| 1440 | { |
| 1441 | return Fail("%s: Could not read output", __func__); |
| 1442 | } |
| 1443 | |
Teresa Charlin | 8f6429d | 2019-10-01 13:10:15 +0100 | [diff] [blame] | 1444 | const TensorInfo& inputInfo = input.GetTensorInfo(); |
Aron Virginas-Tar | cb8ac84 | 2019-07-05 15:47:07 +0100 | [diff] [blame] | 1445 | unsigned int rank = inputInfo.GetNumDimensions(); |
| 1446 | |
Teresa Charlin | 8f6429d | 2019-10-01 13:10:15 +0100 | [diff] [blame] | 1447 | PadDescriptor descriptor; |
Aron Virginas-Tar | cb8ac84 | 2019-07-05 15:47:07 +0100 | [diff] [blame] | 1448 | if (!ConvertPaddings<hal_1_2::HalPolicy>(operation, model, data, rank, descriptor)) |
| 1449 | { |
| 1450 | return Fail("%s: Could not convert paddings", __func__); |
| 1451 | } |
| 1452 | |
Teresa Charlin | 8f6429d | 2019-10-01 13:10:15 +0100 | [diff] [blame] | 1453 | const TensorInfo& outputInfo = GetTensorInfoForOperand(*output); |
Aron Virginas-Tar | 573a8fa | 2019-07-23 14:01:37 +0100 | [diff] [blame] | 1454 | if (IsDynamicTensor(outputInfo)) |
Sadik Armagan | 310d8ff | 2019-07-11 10:53:38 +0100 | [diff] [blame] | 1455 | { |
Aron Virginas-Tar | b7421e5 | 2019-07-26 13:14:39 +0100 | [diff] [blame] | 1456 | return Fail("%s: Dynamic output tensors are not supported", __func__); |
Sadik Armagan | 310d8ff | 2019-07-11 10:53:38 +0100 | [diff] [blame] | 1457 | } |
| 1458 | |
Aron Virginas-Tar | cb8ac84 | 2019-07-05 15:47:07 +0100 | [diff] [blame] | 1459 | // Determine type of padding value |
| 1460 | OperandType operandType0; |
| 1461 | OperandType operandType2; |
| 1462 | |
| 1463 | if (!GetOperandType<hal_1_2::HalPolicy>(operation, 0, model, operandType0) || |
| 1464 | !GetOperandType<hal_1_2::HalPolicy>(operation, 2, model, operandType2)) |
| 1465 | { |
| 1466 | return Fail("%s: Operation has invalid inputs", __func__); |
| 1467 | } |
| 1468 | |
| 1469 | // Read value to use for padding |
| 1470 | if (operandType0 == OperandType::TENSOR_FLOAT16 && operandType2 == OperandType::FLOAT16) |
| 1471 | { |
Teresa Charlin | 8f6429d | 2019-10-01 13:10:15 +0100 | [diff] [blame] | 1472 | Half f16PadValue; |
Aron Virginas-Tar | cb8ac84 | 2019-07-05 15:47:07 +0100 | [diff] [blame] | 1473 | if (!GetInputScalar<hal_1_2::HalPolicy>(operation, 2, operandType2, f16PadValue, model, data)) |
| 1474 | { |
| 1475 | return Fail("%s: Could not read input 2 (FLOAT16)", __func__); |
| 1476 | } |
| 1477 | |
| 1478 | descriptor.m_PadValue = f16PadValue; |
| 1479 | } |
| 1480 | else if (operandType0 == OperandType::TENSOR_FLOAT32 && operandType2 == OperandType::FLOAT32) |
| 1481 | { |
| 1482 | if (!GetInputFloat32<hal_1_2::HalPolicy>(operation, 2, descriptor.m_PadValue, model, data)) |
| 1483 | { |
| 1484 | return Fail("%s: Could not read input 2 (FLOAT32)", __func__); |
| 1485 | } |
| 1486 | } |
| 1487 | else if (operandType0 == OperandType::TENSOR_QUANT8_ASYMM && operandType2 == OperandType::INT32) |
| 1488 | { |
Mike Kelly | 3c67394 | 2019-07-25 09:26:06 +0100 | [diff] [blame] | 1489 | int32_t intPadValue = 0; |
| 1490 | if (!GetInputInt32<hal_1_2::HalPolicy>(operation, 2, intPadValue, model, data)) |
Aron Virginas-Tar | cb8ac84 | 2019-07-05 15:47:07 +0100 | [diff] [blame] | 1491 | { |
| 1492 | return Fail("%s: Could not read input 2 (INT32)", __func__); |
| 1493 | } |
Mike Kelly | 3c67394 | 2019-07-25 09:26:06 +0100 | [diff] [blame] | 1494 | descriptor.m_PadValue = intPadValue; |
Aron Virginas-Tar | cb8ac84 | 2019-07-05 15:47:07 +0100 | [diff] [blame] | 1495 | } |
| 1496 | else |
| 1497 | { |
| 1498 | return Fail("%s: Operation has invalid inputs: type mismatch", __func__); |
| 1499 | } |
| 1500 | |
Ferran Balaguer | d30093c | 2019-07-09 17:04:47 +0100 | [diff] [blame] | 1501 | bool isSupported = false; |
| 1502 | FORWARD_LAYER_SUPPORT_FUNC(__func__, |
| 1503 | IsPadSupported, |
| 1504 | data.m_Backends, |
| 1505 | isSupported, |
| 1506 | inputInfo, |
| 1507 | outputInfo, |
| 1508 | descriptor); |
| 1509 | if (!isSupported) |
Aron Virginas-Tar | cb8ac84 | 2019-07-05 15:47:07 +0100 | [diff] [blame] | 1510 | { |
| 1511 | return false; |
| 1512 | } |
| 1513 | |
Teresa Charlin | 8f6429d | 2019-10-01 13:10:15 +0100 | [diff] [blame] | 1514 | IConnectableLayer* const layer = data.m_Network->AddPadLayer(descriptor); |
Aron Virginas-Tar | cb8ac84 | 2019-07-05 15:47:07 +0100 | [diff] [blame] | 1515 | assert(layer != nullptr); |
| 1516 | input.Connect(layer->GetInputSlot(0)); |
| 1517 | layer->GetOutputSlot(0).SetTensorInfo(outputInfo); |
| 1518 | |
Aron Virginas-Tar | b7421e5 | 2019-07-26 13:14:39 +0100 | [diff] [blame] | 1519 | return SetupAndTrackLayerOutputSlot<hal_1_2::HalPolicy>(operation, 0, *layer, model, data); |
Aron Virginas-Tar | cb8ac84 | 2019-07-05 15:47:07 +0100 | [diff] [blame] | 1520 | } |
| 1521 | |
Matteo Martincigh | 17ffff3 | 2019-06-27 14:12:55 +0100 | [diff] [blame] | 1522 | bool HalPolicy::ConvertPrelu(const Operation& operation, const Model& model, ConversionData& data) |
| 1523 | { |
Aron Virginas-Tar | 29404fb | 2019-07-24 13:55:31 +0100 | [diff] [blame] | 1524 | ALOGV("hal_1_2::HalPolicy::ConvertPrelu()"); |
| 1525 | |
Matteo Martincigh | 17ffff3 | 2019-06-27 14:12:55 +0100 | [diff] [blame] | 1526 | LayerInputHandle input = ConvertToLayerInputHandle<hal_1_2::HalPolicy>(operation, 0, model, data); |
| 1527 | LayerInputHandle alpha = ConvertToLayerInputHandle<hal_1_2::HalPolicy>(operation, 1, model, data); |
| 1528 | |
| 1529 | if (!input.IsValid() || !alpha.IsValid()) |
| 1530 | { |
| 1531 | return Fail("%s: Operation has invalid inputs", __func__); |
| 1532 | } |
| 1533 | |
| 1534 | const Operand* output = GetOutputOperand<hal_1_2::HalPolicy>(operation, 0, model); |
| 1535 | |
| 1536 | if (!output) |
| 1537 | { |
Matteo Martincigh | 0bd89a8 | 2019-07-02 16:53:10 +0100 | [diff] [blame] | 1538 | return Fail("%s: Could not read output", __func__); |
Matteo Martincigh | 17ffff3 | 2019-06-27 14:12:55 +0100 | [diff] [blame] | 1539 | } |
| 1540 | |
Teresa Charlin | 8f6429d | 2019-10-01 13:10:15 +0100 | [diff] [blame] | 1541 | const TensorInfo& inputInfo = input.GetTensorInfo(); |
| 1542 | const TensorInfo& alphaInfo = alpha.GetTensorInfo(); |
| 1543 | const TensorInfo& outputInfo = GetTensorInfoForOperand(*output); |
Aron Virginas-Tar | f03fcf0 | 2019-07-09 17:44:24 +0100 | [diff] [blame] | 1544 | |
Aron Virginas-Tar | 573a8fa | 2019-07-23 14:01:37 +0100 | [diff] [blame] | 1545 | if (IsDynamicTensor(outputInfo)) |
Aron Virginas-Tar | f03fcf0 | 2019-07-09 17:44:24 +0100 | [diff] [blame] | 1546 | { |
Aron Virginas-Tar | b7421e5 | 2019-07-26 13:14:39 +0100 | [diff] [blame] | 1547 | return Fail("%s: Dynamic output tensors are not supported", __func__); |
Aron Virginas-Tar | f03fcf0 | 2019-07-09 17:44:24 +0100 | [diff] [blame] | 1548 | } |
Matteo Martincigh | 17ffff3 | 2019-06-27 14:12:55 +0100 | [diff] [blame] | 1549 | |
Ferran Balaguer | d30093c | 2019-07-09 17:04:47 +0100 | [diff] [blame] | 1550 | bool isSupported = false; |
| 1551 | FORWARD_LAYER_SUPPORT_FUNC(__func__, |
| 1552 | IsPreluSupported, |
| 1553 | data.m_Backends, |
| 1554 | isSupported, |
| 1555 | inputInfo, |
| 1556 | alphaInfo, |
| 1557 | outputInfo); |
| 1558 | if (!isSupported) |
Matteo Martincigh | 17ffff3 | 2019-06-27 14:12:55 +0100 | [diff] [blame] | 1559 | { |
| 1560 | return false; |
| 1561 | } |
| 1562 | |
Teresa Charlin | 8f6429d | 2019-10-01 13:10:15 +0100 | [diff] [blame] | 1563 | IConnectableLayer* const layer = data.m_Network->AddPreluLayer(); |
Matteo Martincigh | 17ffff3 | 2019-06-27 14:12:55 +0100 | [diff] [blame] | 1564 | |
| 1565 | if (!layer) |
| 1566 | { |
| 1567 | return Fail("%s: AddPreluLayer failed", __func__); |
| 1568 | } |
| 1569 | |
Derek Lamberti | 6fd4ceb | 2019-12-19 15:45:35 +0000 | [diff] [blame] | 1570 | bool isReshapeSupported = BroadcastTensor(input, alpha, layer, data); |
Sadik Armagan | 64b19b5 | 2019-08-19 09:49:58 +0100 | [diff] [blame] | 1571 | if (!isReshapeSupported) |
| 1572 | { |
| 1573 | return false; |
| 1574 | } |
Matteo Martincigh | 17ffff3 | 2019-06-27 14:12:55 +0100 | [diff] [blame] | 1575 | |
Aron Virginas-Tar | b7421e5 | 2019-07-26 13:14:39 +0100 | [diff] [blame] | 1576 | return SetupAndTrackLayerOutputSlot<hal_1_2::HalPolicy>(operation, 0, *layer, model, data); |
Matteo Martincigh | 17ffff3 | 2019-06-27 14:12:55 +0100 | [diff] [blame] | 1577 | } |
| 1578 | |
Sadik Armagan | 5a476a8 | 2019-07-30 09:43:18 +0100 | [diff] [blame] | 1579 | bool HalPolicy::ConvertQuantize(const Operation& operation, const Model& model, ConversionData& data) |
| 1580 | { |
| 1581 | ALOGV("hal_1_2::HalPolicy::ConvertQuantize()"); |
| 1582 | |
| 1583 | LayerInputHandle input = ConvertToLayerInputHandle<hal_1_2::HalPolicy>(operation, 0, model, data); |
| 1584 | if (!input.IsValid()) |
| 1585 | { |
| 1586 | return Fail("%s: Operation has invalid input", __func__); |
| 1587 | } |
| 1588 | |
| 1589 | const Operand* const outputOperand = GetOutputOperand<hal_1_2::HalPolicy>(operation, 0, model); |
| 1590 | if (!outputOperand) |
| 1591 | { |
| 1592 | return Fail("%s: Operation has invalid outputs", __func__); |
| 1593 | } |
| 1594 | |
Teresa Charlin | 8f6429d | 2019-10-01 13:10:15 +0100 | [diff] [blame] | 1595 | const TensorInfo& outputInfo = GetTensorInfoForOperand(*outputOperand); |
Sadik Armagan | 5a476a8 | 2019-07-30 09:43:18 +0100 | [diff] [blame] | 1596 | if (IsDynamicTensor(outputInfo)) |
| 1597 | { |
| 1598 | return Fail("%s: Dynamic output tensors are not supported", __func__); |
| 1599 | } |
| 1600 | |
| 1601 | bool isSupported = false; |
| 1602 | FORWARD_LAYER_SUPPORT_FUNC(__func__, |
| 1603 | IsQuantizeSupported, |
| 1604 | data.m_Backends, |
| 1605 | isSupported, |
| 1606 | input.GetTensorInfo(), |
| 1607 | outputInfo); |
| 1608 | if (!isSupported) |
| 1609 | { |
| 1610 | return false; |
| 1611 | } |
| 1612 | |
Teresa Charlin | 8f6429d | 2019-10-01 13:10:15 +0100 | [diff] [blame] | 1613 | IConnectableLayer* const layer = data.m_Network->AddQuantizeLayer(); |
Sadik Armagan | 5a476a8 | 2019-07-30 09:43:18 +0100 | [diff] [blame] | 1614 | assert(layer != nullptr); |
| 1615 | input.Connect(layer->GetInputSlot(0)); |
| 1616 | |
| 1617 | return SetupAndTrackLayerOutputSlot<hal_1_2::HalPolicy>(operation, 0, *layer, model, data); |
| 1618 | } |
| 1619 | |
Ellen Norris-Thompson | 7efb46d | 2019-07-24 17:39:19 +0100 | [diff] [blame] | 1620 | bool HalPolicy::ConvertQuantizedLstm(const Operation& operation, const Model& model, ConversionData& data) |
| 1621 | { |
| 1622 | ALOGV("hal_1_2::HalPolicy::ConvertQuantizedLstm()"); |
| 1623 | |
| 1624 | //Inputs: |
| 1625 | // 0: The input: A 2-D tensor of type ANEURALNETWORKS_TENSOR_QUANT8_ASYMM and shape [numBatches, inputSize] |
| 1626 | // specifying the input to the LSTM cell. Tensor is quantized with a fixed quantization range of -1, 127/128. |
| 1627 | LayerInputHandle input = ConvertToLayerInputHandle<hal_1_2::HalPolicy>(operation, 0, model, data); |
| 1628 | if (!input.IsValid()) |
| 1629 | { |
| 1630 | return Fail("%s: Could not read input 0: input", __func__); |
| 1631 | } |
| 1632 | |
| 1633 | //13: The previous cell state: A 2-D tensor of type ANEURALNETWORKS_TENSOR_QUANT16_SYMM and shape |
| 1634 | // [numBatches, outputSize] specifying the cell state from the previous time step of the LSTM cell. |
| 1635 | // It is quantized using a quantization range of -2^4, 2^4 * 32767/32768. |
| 1636 | LayerInputHandle previousCellStateIn = ConvertToLayerInputHandle<hal_1_2::HalPolicy>(operation, 13, model, data); |
| 1637 | if (!previousCellStateIn.IsValid()) |
| 1638 | { |
| 1639 | return Fail("%s: Could not read input 13: previousCellStateIn", __func__); |
| 1640 | } |
| 1641 | |
| 1642 | // 14: The previous output state: A 2-D tensor of type ANEURALNETWORKS_TENSOR_QUANT8_ASYMM and shape |
| 1643 | // [numBathes, outputSize] specifying the output of the LSTM cell from previous time-step. Tensor |
| 1644 | // is quantized with a fixed quantization range of -1, 127/128. |
| 1645 | LayerInputHandle previousOutputIn = ConvertToLayerInputHandle<hal_1_2::HalPolicy>(operation, 14, model, data); |
| 1646 | if (!previousOutputIn.IsValid()) |
| 1647 | { |
| 1648 | return Fail("%s: Could not read input 14: previousOutputIn", __func__); |
| 1649 | } |
| 1650 | |
| 1651 | // Get the input tensors: |
| 1652 | // 1: The input-to-input weights. A 2-D tensor of type ANEURALNETWORKS_TENSOR_QUANT8_ASYMM and shape |
| 1653 | // [outputSize, inputSize] specifying input-to-input part of weights for fully-connected layer inside the |
| 1654 | // LSTM cell. Quantization zero point and scale must be the same across all the weights. |
| 1655 | const ConstTensorPin inputToInputWeightsPin = |
Ellen Norris-Thompson | a3d7fad | 2019-08-05 14:20:32 +0100 | [diff] [blame] | 1656 | ConvertOperationInputToConstTensorPin<hal_1_2::HalPolicy>(operation, 1, model, data); |
Ellen Norris-Thompson | 7efb46d | 2019-07-24 17:39:19 +0100 | [diff] [blame] | 1657 | |
| 1658 | // 2: The input-to-forget weights. A 2-D tensor of type ANEURALNETWORKS_TENSOR_QUANT8_ASYMM and shape |
| 1659 | // [outputSize, inputSize] specifying input-to-forget part of weights for fully-connected layer inside the |
| 1660 | // LSTM cell. Quantization zero point and scale must be the same across all the weights. |
| 1661 | const ConstTensorPin inputToForgetWeightsPin = |
Ellen Norris-Thompson | a3d7fad | 2019-08-05 14:20:32 +0100 | [diff] [blame] | 1662 | ConvertOperationInputToConstTensorPin<hal_1_2::HalPolicy>(operation, 2, model, data); |
Ellen Norris-Thompson | 7efb46d | 2019-07-24 17:39:19 +0100 | [diff] [blame] | 1663 | |
| 1664 | // 3: The input-to-cell weights. A 2-D tensor of type ANEURALNETWORKS_TENSOR_QUANT8_ASYMM and shape |
| 1665 | // [outputSize, inputSize] specifying input-to-cell part of weights for fully-connected layer inside the |
| 1666 | // LSTM cell. Quantization zero point and scale must be the same across all the weights. |
| 1667 | const ConstTensorPin inputToCellWeightsPin = |
Ellen Norris-Thompson | a3d7fad | 2019-08-05 14:20:32 +0100 | [diff] [blame] | 1668 | ConvertOperationInputToConstTensorPin<hal_1_2::HalPolicy>(operation, 3, model, data); |
Ellen Norris-Thompson | 7efb46d | 2019-07-24 17:39:19 +0100 | [diff] [blame] | 1669 | |
| 1670 | // 4: The input-to-output weights. A 2-D tensor of type ANEURALNETWORKS_TENSOR_QUANT8_ASYMM and shape |
| 1671 | // [outputSize, inputSize] specifying input-to-output part of weights for fully-connected layer inside the |
| 1672 | // LSTM cell. Quantization zero point and scale must be the same across all the weights. |
| 1673 | const ConstTensorPin inputToOutputWeightsPin = |
Ellen Norris-Thompson | a3d7fad | 2019-08-05 14:20:32 +0100 | [diff] [blame] | 1674 | ConvertOperationInputToConstTensorPin<hal_1_2::HalPolicy>(operation, 4, model, data); |
Ellen Norris-Thompson | 7efb46d | 2019-07-24 17:39:19 +0100 | [diff] [blame] | 1675 | |
| 1676 | // 5: The recurrent-to-input weights. A 2-D tensor of type ANEURALNETWORKS_TENSOR_QUANT8_ASYMM and shape |
| 1677 | // [outputSize, outputSize] specifying recurrent-to-input part of weights for fully-connected layer inside |
| 1678 | // the LSTM cell. Quantization zero point and scale must be the same across all the weights. |
| 1679 | const ConstTensorPin recurrentToInputWeightsPin = |
Ellen Norris-Thompson | a3d7fad | 2019-08-05 14:20:32 +0100 | [diff] [blame] | 1680 | ConvertOperationInputToConstTensorPin<hal_1_2::HalPolicy>(operation, 5, model, data); |
Ellen Norris-Thompson | 7efb46d | 2019-07-24 17:39:19 +0100 | [diff] [blame] | 1681 | |
| 1682 | // 6: The recurrent-to-forget weights. A 2-D tensor of type ANEURALNETWORKS_TENSOR_QUANT8_ASYMM and shape |
| 1683 | // [outputSize, outputSize] specifying recurrent-to-forget part of weights for fully-connected layer inside |
| 1684 | // the LSTM cell. Quantization zero point and scale must be the same across all the weights. |
| 1685 | const ConstTensorPin recurrentToForgetWeightsPin = |
Ellen Norris-Thompson | a3d7fad | 2019-08-05 14:20:32 +0100 | [diff] [blame] | 1686 | ConvertOperationInputToConstTensorPin<hal_1_2::HalPolicy>(operation, 6, model, data); |
Ellen Norris-Thompson | 7efb46d | 2019-07-24 17:39:19 +0100 | [diff] [blame] | 1687 | |
| 1688 | // 7: The recurrent-to-cell weights. A 2-D tensor of type ANEURALNETWORKS_TENSOR_QUANT8_ASYMM and shape |
| 1689 | // [outputSize, outputSize] specifying recurrent-to-cell part of weights for fully-connected layer inside |
| 1690 | // the LSTM cell. Quantization zero point and scale must be the same across all the weights. |
| 1691 | const ConstTensorPin recurrentToCellWeightsPin = |
Ellen Norris-Thompson | a3d7fad | 2019-08-05 14:20:32 +0100 | [diff] [blame] | 1692 | ConvertOperationInputToConstTensorPin<hal_1_2::HalPolicy>(operation, 7, model, data); |
Ellen Norris-Thompson | 7efb46d | 2019-07-24 17:39:19 +0100 | [diff] [blame] | 1693 | |
| 1694 | // 8: The recurrent-to-output weights. A 2-D tensor of type ANEURALNETWORKS_TENSOR_QUANT8_ASYMM and shape |
| 1695 | // [outputSize, outputSize] specifying recurrent-to-output part of weights for fully-connected layer inside |
| 1696 | // the LSTM cell. Quantization zero point and scale must be the same across all the weights. |
| 1697 | const ConstTensorPin recurrentToOutputWeightsPin = |
Ellen Norris-Thompson | a3d7fad | 2019-08-05 14:20:32 +0100 | [diff] [blame] | 1698 | ConvertOperationInputToConstTensorPin<hal_1_2::HalPolicy>(operation, 8, model, data); |
Ellen Norris-Thompson | 7efb46d | 2019-07-24 17:39:19 +0100 | [diff] [blame] | 1699 | |
| 1700 | // 9: The input gate bias. A 1-D tensor of type ANEURALNETWORKS_TENSOR_INT32 and shape [outputSize] specifying the |
| 1701 | // bias for the fully-connected layer inside the LSTM cell. Bias is quantized with scale being a product |
| 1702 | // of input and weights scales and zeroPoint equal to 0. |
| 1703 | const ConstTensorPin inputGateBiasPin = |
Ellen Norris-Thompson | a3d7fad | 2019-08-05 14:20:32 +0100 | [diff] [blame] | 1704 | ConvertOperationInputToConstTensorPin<hal_1_2::HalPolicy>(operation, 9, model, data); |
Ellen Norris-Thompson | 7efb46d | 2019-07-24 17:39:19 +0100 | [diff] [blame] | 1705 | |
| 1706 | // 10: The forget gate bias. A 1-D tensor of type ANEURALNETWORKS_TENSOR_INT32 and shape [outputSize] specifying |
| 1707 | // the bias for the fully-connected layer inside the LSTM cell. Bias is quantized with scale being a product |
| 1708 | // of input and weights scales and zeroPoint equal to 0. |
| 1709 | const ConstTensorPin forgetGateBiasPin = |
Ellen Norris-Thompson | a3d7fad | 2019-08-05 14:20:32 +0100 | [diff] [blame] | 1710 | ConvertOperationInputToConstTensorPin<hal_1_2::HalPolicy>(operation, 10, model, data); |
Ellen Norris-Thompson | 7efb46d | 2019-07-24 17:39:19 +0100 | [diff] [blame] | 1711 | |
| 1712 | // 11:The cell bias. A 1-D tensor of type ANEURALNETWORKS_TENSOR_INT32 and shape [outputSize] specifying the bias |
| 1713 | // for the fully-connected layer inside the LSTM cell. Bias is quantized with scale being a product of input |
| 1714 | // and weights scales and zeroPoint equal to 0. |
| 1715 | const ConstTensorPin cellBiasPin = |
Ellen Norris-Thompson | a3d7fad | 2019-08-05 14:20:32 +0100 | [diff] [blame] | 1716 | ConvertOperationInputToConstTensorPin<hal_1_2::HalPolicy>(operation, 11, model, data); |
Ellen Norris-Thompson | 7efb46d | 2019-07-24 17:39:19 +0100 | [diff] [blame] | 1717 | |
| 1718 | // 12:The output gate bias. A 1-D tensor of type ANEURALNETWORKS_TENSOR_INT32 and shape [outputSize] specifying |
| 1719 | // the bias for the fully-connected layer inside the LSTM cell. Bias is quantized with scale being a product |
| 1720 | // of input and weights scales and zeroPoint equal to 0. |
| 1721 | const ConstTensorPin outputGateBiasPin = |
Ellen Norris-Thompson | a3d7fad | 2019-08-05 14:20:32 +0100 | [diff] [blame] | 1722 | ConvertOperationInputToConstTensorPin<hal_1_2::HalPolicy>(operation, 12, model, data); |
Ellen Norris-Thompson | 7efb46d | 2019-07-24 17:39:19 +0100 | [diff] [blame] | 1723 | |
| 1724 | if (!inputToInputWeightsPin.IsValid() || |
| 1725 | !inputToForgetWeightsPin.IsValid() || |
| 1726 | !inputToCellWeightsPin.IsValid() || |
| 1727 | !inputToOutputWeightsPin.IsValid() || |
| 1728 | !recurrentToInputWeightsPin.IsValid() || |
| 1729 | !recurrentToForgetWeightsPin.IsValid() || |
| 1730 | !recurrentToCellWeightsPin.IsValid() || |
| 1731 | !recurrentToOutputWeightsPin.IsValid() || |
| 1732 | !inputGateBiasPin.IsValid() || |
| 1733 | !forgetGateBiasPin.IsValid() || |
| 1734 | !cellBiasPin.IsValid() || |
| 1735 | !outputGateBiasPin.IsValid()) |
| 1736 | { |
| 1737 | return Fail("%s: Operation has invalid tensor inputs", __func__); |
| 1738 | } |
| 1739 | |
| 1740 | // Outputs: |
| 1741 | // 0: The cell state: A 2-D tensor of type ANEURALNETWORKS_TENSOR_QUANT16_SYMM and shape [numBatches, outputSize] |
| 1742 | // which contains a cell state from the current time step. Tensor is quantized using a quantization range |
| 1743 | // of -2^4, 2^4 * 32767/32768. |
| 1744 | const Operand* cellStateOut = GetOutputOperand<hal_1_2::HalPolicy>(operation, 0, model); |
| 1745 | if (!cellStateOut) |
| 1746 | { |
| 1747 | return Fail("%s: Could not read output 0: cellStateOut", __func__); |
| 1748 | } |
| 1749 | |
| 1750 | // 1: The output: A 2-D tensor of type ANEURALNETWORKS_TENSOR_QUANT8_ASYMM and shape [numBathes, outputSize] which |
| 1751 | // contains the output value. Tensor is quantized with a fixed quantization range of -1, 127/128. |
| 1752 | const Operand* output = GetOutputOperand<hal_1_2::HalPolicy>(operation, 1, model); |
| 1753 | if (!output) |
| 1754 | { |
| 1755 | return Fail("%s: Could not read output 1: output", __func__); |
| 1756 | } |
| 1757 | |
| 1758 | // Inputs |
Teresa Charlin | 8f6429d | 2019-10-01 13:10:15 +0100 | [diff] [blame] | 1759 | const TensorInfo& inputInfo = input.GetTensorInfo(); |
| 1760 | const TensorInfo& previousCellStateInInfo = previousCellStateIn.GetTensorInfo(); |
| 1761 | const TensorInfo& previousOutputInInfo = previousOutputIn.GetTensorInfo(); |
Ellen Norris-Thompson | 7efb46d | 2019-07-24 17:39:19 +0100 | [diff] [blame] | 1762 | |
| 1763 | // Outputs |
Teresa Charlin | 8f6429d | 2019-10-01 13:10:15 +0100 | [diff] [blame] | 1764 | const TensorInfo& cellStateOutInfo = GetTensorInfoForOperand(*cellStateOut); |
| 1765 | const TensorInfo& outputInfo = GetTensorInfoForOperand(*output); |
Ellen Norris-Thompson | 7efb46d | 2019-07-24 17:39:19 +0100 | [diff] [blame] | 1766 | |
| 1767 | // Dynamic tensors currently not supported |
| 1768 | if (IsDynamicTensor(cellStateOutInfo) || IsDynamicTensor(outputInfo)) |
| 1769 | { |
| 1770 | return Fail("%s: Dynamic output tensors are not supported", __func__); |
| 1771 | } |
| 1772 | |
Teresa Charlin | 8f6429d | 2019-10-01 13:10:15 +0100 | [diff] [blame] | 1773 | QuantizedLstmInputParams params; |
Ellen Norris-Thompson | 7efb46d | 2019-07-24 17:39:19 +0100 | [diff] [blame] | 1774 | |
| 1775 | params.m_InputToInputWeights = inputToInputWeightsPin.GetConstTensorPtr(); |
| 1776 | params.m_InputToForgetWeights = inputToForgetWeightsPin.GetConstTensorPtr(); |
| 1777 | params.m_InputToCellWeights = inputToCellWeightsPin.GetConstTensorPtr(); |
| 1778 | params.m_InputToOutputWeights = inputToOutputWeightsPin.GetConstTensorPtr(); |
| 1779 | params.m_RecurrentToInputWeights = recurrentToInputWeightsPin.GetConstTensorPtr(); |
| 1780 | params.m_RecurrentToForgetWeights = recurrentToForgetWeightsPin.GetConstTensorPtr(); |
| 1781 | params.m_RecurrentToCellWeights = recurrentToCellWeightsPin.GetConstTensorPtr(); |
| 1782 | params.m_RecurrentToOutputWeights = recurrentToOutputWeightsPin.GetConstTensorPtr(); |
| 1783 | params.m_InputGateBias = inputGateBiasPin.GetConstTensorPtr(); |
| 1784 | params.m_ForgetGateBias = forgetGateBiasPin.GetConstTensorPtr(); |
| 1785 | params.m_CellBias = cellBiasPin.GetConstTensorPtr(); |
| 1786 | params.m_OutputGateBias = outputGateBiasPin.GetConstTensorPtr(); |
| 1787 | |
Teresa Charlin | 8f6429d | 2019-10-01 13:10:15 +0100 | [diff] [blame] | 1788 | QuantizedLstmInputParamsInfo paramsInfo; |
Ellen Norris-Thompson | 7efb46d | 2019-07-24 17:39:19 +0100 | [diff] [blame] | 1789 | paramsInfo.m_InputToInputWeights = &(params.m_InputToInputWeights->GetInfo()); |
| 1790 | paramsInfo.m_InputToForgetWeights = &(params.m_InputToForgetWeights->GetInfo()); |
| 1791 | paramsInfo.m_InputToCellWeights = &(params.m_InputToCellWeights->GetInfo()); |
| 1792 | paramsInfo.m_InputToOutputWeights = &(params.m_InputToOutputWeights->GetInfo()); |
| 1793 | paramsInfo.m_RecurrentToInputWeights = &(params.m_RecurrentToInputWeights->GetInfo()); |
| 1794 | paramsInfo.m_RecurrentToForgetWeights = &(params.m_RecurrentToForgetWeights->GetInfo()); |
| 1795 | paramsInfo.m_RecurrentToCellWeights = &(params.m_RecurrentToCellWeights->GetInfo()); |
| 1796 | paramsInfo.m_RecurrentToOutputWeights = &(params.m_RecurrentToOutputWeights->GetInfo()); |
| 1797 | paramsInfo.m_InputGateBias = &(params.m_InputGateBias->GetInfo()); |
| 1798 | paramsInfo.m_ForgetGateBias = &(params.m_ForgetGateBias->GetInfo()); |
| 1799 | paramsInfo.m_CellBias = &(params.m_CellBias->GetInfo()); |
| 1800 | paramsInfo.m_OutputGateBias = &(params.m_OutputGateBias->GetInfo()); |
| 1801 | |
| 1802 | bool isSupported = false; |
| 1803 | FORWARD_LAYER_SUPPORT_FUNC(__func__, |
| 1804 | IsQuantizedLstmSupported, |
| 1805 | data.m_Backends, |
| 1806 | isSupported, |
| 1807 | inputInfo, |
| 1808 | previousCellStateInInfo, |
| 1809 | previousOutputInInfo, |
| 1810 | cellStateOutInfo, |
| 1811 | outputInfo, |
| 1812 | paramsInfo); |
| 1813 | |
| 1814 | if (!isSupported) |
| 1815 | { |
| 1816 | return false; |
| 1817 | } |
| 1818 | |
Teresa Charlin | 8f6429d | 2019-10-01 13:10:15 +0100 | [diff] [blame] | 1819 | IConnectableLayer* const layer = data.m_Network->AddQuantizedLstmLayer(params, "QuantizedLstm"); |
Ellen Norris-Thompson | 7efb46d | 2019-07-24 17:39:19 +0100 | [diff] [blame] | 1820 | input.Connect(layer->GetInputSlot(0)); |
Ellen Norris-Thompson | a3d7fad | 2019-08-05 14:20:32 +0100 | [diff] [blame] | 1821 | previousCellStateIn.Connect(layer->GetInputSlot(1)); |
| 1822 | previousOutputIn.Connect(layer->GetInputSlot(2)); |
Ellen Norris-Thompson | 7efb46d | 2019-07-24 17:39:19 +0100 | [diff] [blame] | 1823 | |
| 1824 | return (SetupAndTrackLayerOutputSlot<hal_1_2::HalPolicy>(operation, 0, *layer, 0, model, data) && |
| 1825 | SetupAndTrackLayerOutputSlot<hal_1_2::HalPolicy>(operation, 1, *layer, 1, model, data)); |
| 1826 | } |
| 1827 | |
Sadik Armagan | 6111316 | 2019-07-25 09:09:40 +0100 | [diff] [blame] | 1828 | bool HalPolicy::ConvertReLu(const Operation& operation, const Model& model, ConversionData& data) |
| 1829 | { |
| 1830 | ALOGV("hal_1_2::HalPolicy::ConvertReLu()"); |
| 1831 | return ::ConvertReLu<hal_1_2::HalPolicy>(operation, model, data); |
| 1832 | } |
| 1833 | |
| 1834 | bool HalPolicy::ConvertReLu1(const Operation& operation, const Model& model, ConversionData& data) |
| 1835 | { |
| 1836 | ALOGV("hal_1_2::HalPolicy::ConvertReLu1()"); |
| 1837 | return ::ConvertReLu1<hal_1_2::HalPolicy>(operation, model, data); |
| 1838 | } |
| 1839 | |
| 1840 | bool HalPolicy::ConvertReLu6(const Operation& operation, const Model& model, ConversionData& data) |
| 1841 | { |
| 1842 | ALOGV("hal_1_2::HalPolicy::ConvertReLu6()"); |
| 1843 | return ::ConvertReLu6<hal_1_2::HalPolicy>(operation, model, data); |
| 1844 | } |
| 1845 | |
Mike Kelly | 4627280 | 2019-08-14 17:00:48 +0100 | [diff] [blame] | 1846 | bool HalPolicy::ConvertReshape(const Operation& operation, const Model& model, ConversionData& data) |
| 1847 | { |
| 1848 | ALOGV("hal_1_2::HalPolicy::ConvertReshape()"); |
| 1849 | return ::ConvertReshape<hal_1_2::HalPolicy>(operation, model, data); |
| 1850 | } |
| 1851 | |
Aron Virginas-Tar | fb2fa29 | 2019-07-04 11:59:48 +0100 | [diff] [blame] | 1852 | bool HalPolicy::ConvertResize(const Operation& operation, |
| 1853 | const Model& model, |
| 1854 | ConversionData& data, |
Teresa Charlin | 8f6429d | 2019-10-01 13:10:15 +0100 | [diff] [blame] | 1855 | ResizeMethod resizeMethod) |
Aron Virginas-Tar | 7a6d11b | 2019-07-03 15:27:08 +0100 | [diff] [blame] | 1856 | { |
Aron Virginas-Tar | 29404fb | 2019-07-24 13:55:31 +0100 | [diff] [blame] | 1857 | ALOGV("hal_1_2::HalPolicy::ConvertResize()"); |
Aron Virginas-Tar | 7d2ccfd | 2019-10-29 14:03:51 +0000 | [diff] [blame] | 1858 | ALOGV("resizeMethod = %s", GetResizeMethodAsCString(resizeMethod)); |
Aron Virginas-Tar | 29404fb | 2019-07-24 13:55:31 +0100 | [diff] [blame] | 1859 | |
| 1860 | LayerInputHandle input = ConvertToLayerInputHandle<hal_1_2::HalPolicy>(operation, 0, model, data); |
Aron Virginas-Tar | 7a6d11b | 2019-07-03 15:27:08 +0100 | [diff] [blame] | 1861 | if (!input.IsValid()) |
| 1862 | { |
| 1863 | return Fail("%s: Could not read input 0", __func__); |
| 1864 | } |
| 1865 | |
| 1866 | const Operand* output = GetOutputOperand<hal_1_2::HalPolicy>(operation, 0, model); |
| 1867 | if (!output) |
| 1868 | { |
| 1869 | return Fail("%s: Could not read output 0", __func__); |
| 1870 | } |
| 1871 | |
Teresa Charlin | 8f6429d | 2019-10-01 13:10:15 +0100 | [diff] [blame] | 1872 | const TensorInfo& inputInfo = input.GetTensorInfo(); |
| 1873 | const TensorInfo& outputInfo = GetTensorInfoForOperand(*output); |
Aron Virginas-Tar | b7421e5 | 2019-07-26 13:14:39 +0100 | [diff] [blame] | 1874 | |
| 1875 | if (IsDynamicTensor(outputInfo)) |
| 1876 | { |
| 1877 | return Fail("%s: Dynamic output tensors are not supported", __func__); |
| 1878 | } |
Aron Virginas-Tar | 7a6d11b | 2019-07-03 15:27:08 +0100 | [diff] [blame] | 1879 | |
Teresa Charlin | 8f6429d | 2019-10-01 13:10:15 +0100 | [diff] [blame] | 1880 | ResizeDescriptor descriptor; |
Aron Virginas-Tar | fb2fa29 | 2019-07-04 11:59:48 +0100 | [diff] [blame] | 1881 | descriptor.m_Method = resizeMethod; |
Aron Virginas-Tar | 7a6d11b | 2019-07-03 15:27:08 +0100 | [diff] [blame] | 1882 | descriptor.m_DataLayout = OptionalDataLayout<hal_1_2::HalPolicy>(operation, 3, model, data); |
| 1883 | |
| 1884 | OperandType operandType1; |
| 1885 | OperandType operandType2; |
| 1886 | |
| 1887 | if (!GetOperandType<hal_1_2::HalPolicy>(operation, 1, model, operandType1) || |
| 1888 | !GetOperandType<hal_1_2::HalPolicy>(operation, 2, model, operandType2)) |
| 1889 | { |
| 1890 | return Fail("%s: Operation has invalid inputs", __func__); |
| 1891 | } |
| 1892 | |
| 1893 | if (operandType1 != operandType2) |
| 1894 | { |
| 1895 | return Fail("%s: Operation has invalid inputs. Type of input 1 and 2 should be the same", __func__); |
| 1896 | } |
| 1897 | |
| 1898 | if (operandType1 == OperandType::INT32) |
| 1899 | { |
| 1900 | // Case 1: resizing by shape |
| 1901 | int32_t targetWidth = 0; |
| 1902 | int32_t targetHeight = 0; |
| 1903 | |
| 1904 | if (!GetInputInt32<hal_1_2::HalPolicy>(operation, 1, targetWidth, model, data) || |
| 1905 | !GetInputInt32<hal_1_2::HalPolicy>(operation, 2, targetHeight, model, data)) |
| 1906 | { |
| 1907 | return Fail("%s: Operation has invalid inputs for resizing by shape", __func__); |
| 1908 | } |
| 1909 | |
| 1910 | if (targetWidth < 0 || targetHeight < 0) |
| 1911 | { |
| 1912 | return Fail("%s: Operation has invalid inputs for resizing by shape. " |
| 1913 | "Target width/height cannot be < 0", __func__); |
| 1914 | } |
| 1915 | |
| 1916 | descriptor.m_TargetWidth = static_cast<uint32_t>(targetWidth); |
Teresa Charlin | 9843c01 | 2019-07-19 12:18:35 +0100 | [diff] [blame] | 1917 | descriptor.m_TargetHeight = static_cast<uint32_t>(targetHeight); |
Aron Virginas-Tar | 7a6d11b | 2019-07-03 15:27:08 +0100 | [diff] [blame] | 1918 | } |
| 1919 | else if (operandType1 == OperandType::FLOAT32) |
| 1920 | { |
| 1921 | // Case 2: resizing by scale |
| 1922 | float widthScale = 1.0f; |
| 1923 | float heightScale = 1.0f; |
| 1924 | |
| 1925 | if (!GetInputFloat32<hal_1_2::HalPolicy>(operation, 1, widthScale, model, data) || |
| 1926 | !GetInputFloat32<hal_1_2::HalPolicy>(operation, 2, heightScale, model, data)) |
| 1927 | { |
| 1928 | return Fail("%s: Operation has invalid inputs for resizing by scale", __func__); |
| 1929 | } |
| 1930 | |
Teresa Charlin | 8f6429d | 2019-10-01 13:10:15 +0100 | [diff] [blame] | 1931 | const TensorShape& inputShape = inputInfo.GetShape(); |
Aron Virginas-Tar | 7a6d11b | 2019-07-03 15:27:08 +0100 | [diff] [blame] | 1932 | armnnUtils::DataLayoutIndexed dataLayoutIndexed(descriptor.m_DataLayout); |
| 1933 | |
| 1934 | float width = inputShape[dataLayoutIndexed.GetWidthIndex()]; |
| 1935 | float height = inputShape[dataLayoutIndexed.GetHeightIndex()]; |
| 1936 | |
| 1937 | descriptor.m_TargetWidth = std::floor(width * widthScale); |
| 1938 | descriptor.m_TargetHeight = std::floor(height * heightScale); |
| 1939 | } |
Keith Davis | d410b60 | 2019-12-19 12:31:30 +0000 | [diff] [blame] | 1940 | else if (operandType1 == OperandType::FLOAT16) |
| 1941 | { |
| 1942 | Half widthScale; |
| 1943 | Half heightScale; |
| 1944 | |
| 1945 | if (!GetInputScalar<HalPolicy>(operation, 1, HalPolicy::OperandType::FLOAT16, widthScale, model, data) || |
| 1946 | !GetInputScalar<HalPolicy>(operation, 2, HalPolicy::OperandType::FLOAT16, heightScale, model, data)) |
| 1947 | { |
| 1948 | return Fail("%s: Operation has invalid inputs for resizing by scale", __func__); |
| 1949 | } |
| 1950 | |
| 1951 | const TensorShape& inputShape = inputInfo.GetShape(); |
| 1952 | armnnUtils::DataLayoutIndexed dataLayoutIndexed(descriptor.m_DataLayout); |
| 1953 | |
| 1954 | Half width = static_cast<Half>(inputShape[dataLayoutIndexed.GetWidthIndex()]); |
| 1955 | Half height = static_cast<Half>(inputShape[dataLayoutIndexed.GetHeightIndex()]); |
| 1956 | |
| 1957 | descriptor.m_TargetWidth = std::floor(width * widthScale); |
| 1958 | descriptor.m_TargetHeight = std::floor(height * heightScale); |
| 1959 | } |
Aron Virginas-Tar | 7a6d11b | 2019-07-03 15:27:08 +0100 | [diff] [blame] | 1960 | else |
| 1961 | { |
Keith Davis | d410b60 | 2019-12-19 12:31:30 +0000 | [diff] [blame] | 1962 | return Fail("%s: Operand has invalid data type for resizing by scale", __func__); |
Aron Virginas-Tar | 7a6d11b | 2019-07-03 15:27:08 +0100 | [diff] [blame] | 1963 | } |
| 1964 | |
Ferran Balaguer | d30093c | 2019-07-09 17:04:47 +0100 | [diff] [blame] | 1965 | bool isSupported = false; |
| 1966 | FORWARD_LAYER_SUPPORT_FUNC(__func__, |
| 1967 | IsResizeSupported, |
| 1968 | data.m_Backends, |
| 1969 | isSupported, |
| 1970 | inputInfo, |
| 1971 | outputInfo, |
| 1972 | descriptor); |
Aron Virginas-Tar | be5d356 | 2019-07-16 11:32:29 +0100 | [diff] [blame] | 1973 | |
Ferran Balaguer | d30093c | 2019-07-09 17:04:47 +0100 | [diff] [blame] | 1974 | if (!isSupported) |
Aron Virginas-Tar | 7a6d11b | 2019-07-03 15:27:08 +0100 | [diff] [blame] | 1975 | { |
| 1976 | return false; |
| 1977 | } |
| 1978 | |
Teresa Charlin | 8f6429d | 2019-10-01 13:10:15 +0100 | [diff] [blame] | 1979 | IConnectableLayer* layer = data.m_Network->AddResizeLayer(descriptor); |
Aron Virginas-Tar | 7a6d11b | 2019-07-03 15:27:08 +0100 | [diff] [blame] | 1980 | |
| 1981 | assert(layer != nullptr); |
| 1982 | |
Aron Virginas-Tar | 7a6d11b | 2019-07-03 15:27:08 +0100 | [diff] [blame] | 1983 | input.Connect(layer->GetInputSlot(0)); |
| 1984 | |
Aron Virginas-Tar | b7421e5 | 2019-07-26 13:14:39 +0100 | [diff] [blame] | 1985 | return SetupAndTrackLayerOutputSlot<hal_1_2::HalPolicy>(operation, 0, *layer, model, data); |
Aron Virginas-Tar | 7a6d11b | 2019-07-03 15:27:08 +0100 | [diff] [blame] | 1986 | } |
| 1987 | |
Finn Williams | d74c505 | 2019-07-30 17:06:00 +0100 | [diff] [blame] | 1988 | bool HalPolicy::ConvertSpaceToBatchNd(const Operation& operation, const Model& model, ConversionData& data) |
| 1989 | { |
| 1990 | ALOGV("hal_1_2::HalPolicy::ConvertSpaceToBatchNd()"); |
| 1991 | return ::ConvertSpaceToBatchNd<hal_1_2::HalPolicy>(operation, model, data); |
| 1992 | } |
| 1993 | |
Keith Davis | a6bc52f | 2019-06-26 09:39:49 +0100 | [diff] [blame] | 1994 | bool HalPolicy::ConvertSpaceToDepth(const Operation& operation, const Model& model, ConversionData& data) |
| 1995 | { |
Aron Virginas-Tar | 29404fb | 2019-07-24 13:55:31 +0100 | [diff] [blame] | 1996 | ALOGV("hal_1_2::HalPolicy::ConvertSpaceToDepth()"); |
Keith Davis | a6bc52f | 2019-06-26 09:39:49 +0100 | [diff] [blame] | 1997 | |
Aron Virginas-Tar | 29404fb | 2019-07-24 13:55:31 +0100 | [diff] [blame] | 1998 | LayerInputHandle input = ConvertToLayerInputHandle<hal_1_2::HalPolicy>(operation, 0, model, data); |
Keith Davis | a6bc52f | 2019-06-26 09:39:49 +0100 | [diff] [blame] | 1999 | if (!input.IsValid() ) |
| 2000 | { |
| 2001 | return Fail("%s: Operation has invalid inputs", __func__); |
| 2002 | } |
| 2003 | |
Teresa Charlin | 8f6429d | 2019-10-01 13:10:15 +0100 | [diff] [blame] | 2004 | const TensorInfo& inputInfo = input.GetTensorInfo(); |
Keith Davis | a6bc52f | 2019-06-26 09:39:49 +0100 | [diff] [blame] | 2005 | unsigned int rank = inputInfo.GetNumDimensions(); |
Keith Davis | a6bc52f | 2019-06-26 09:39:49 +0100 | [diff] [blame] | 2006 | if (rank != 4) |
| 2007 | { |
| 2008 | return Fail("%s: Only inputs with rank 4 are supported", __func__); |
| 2009 | } |
| 2010 | |
Aron Virginas-Tar | b7421e5 | 2019-07-26 13:14:39 +0100 | [diff] [blame] | 2011 | const Operand* output = GetOutputOperand<hal_1_2::HalPolicy>(operation, 0, model); |
| 2012 | if (!output) |
| 2013 | { |
| 2014 | return Fail("%s: Could not read output 0", __func__); |
| 2015 | } |
| 2016 | |
Teresa Charlin | 8f6429d | 2019-10-01 13:10:15 +0100 | [diff] [blame] | 2017 | const TensorInfo& outputInfo = GetTensorInfoForOperand(*output); |
Aron Virginas-Tar | b7421e5 | 2019-07-26 13:14:39 +0100 | [diff] [blame] | 2018 | if (IsDynamicTensor(outputInfo)) |
| 2019 | { |
| 2020 | return Fail("%s: Dynamic output tensors are not supported", __func__); |
| 2021 | } |
| 2022 | |
Teresa Charlin | 8f6429d | 2019-10-01 13:10:15 +0100 | [diff] [blame] | 2023 | SpaceToDepthDescriptor desc; |
Keith Davis | a6bc52f | 2019-06-26 09:39:49 +0100 | [diff] [blame] | 2024 | |
| 2025 | GetInputScalar<hal_1_2::HalPolicy>(operation, 1, OperandType::INT32, desc.m_BlockSize, model, data); |
| 2026 | |
| 2027 | if (desc.m_BlockSize <= 1) |
| 2028 | { |
| 2029 | return Fail("%s: Block size must be at least 1 in all dimensions"); |
| 2030 | } |
| 2031 | |
| 2032 | desc.m_DataLayout = OptionalDataLayout<hal_1_2::HalPolicy>(operation, 2, model, data); |
| 2033 | |
Ferran Balaguer | d30093c | 2019-07-09 17:04:47 +0100 | [diff] [blame] | 2034 | bool isSupported = false; |
| 2035 | FORWARD_LAYER_SUPPORT_FUNC(__func__, |
| 2036 | IsSpaceToDepthSupported, |
| 2037 | data.m_Backends, |
| 2038 | isSupported, |
| 2039 | inputInfo, |
| 2040 | outputInfo, |
| 2041 | desc); |
| 2042 | if (!isSupported) |
Keith Davis | a6bc52f | 2019-06-26 09:39:49 +0100 | [diff] [blame] | 2043 | { |
| 2044 | return false; |
| 2045 | } |
| 2046 | |
Teresa Charlin | 8f6429d | 2019-10-01 13:10:15 +0100 | [diff] [blame] | 2047 | IConnectableLayer* const layer = data.m_Network->AddSpaceToDepthLayer(desc); |
Keith Davis | a6bc52f | 2019-06-26 09:39:49 +0100 | [diff] [blame] | 2048 | assert(layer != nullptr); |
| 2049 | input.Connect(layer->GetInputSlot(0)); |
| 2050 | |
Aron Virginas-Tar | b7421e5 | 2019-07-26 13:14:39 +0100 | [diff] [blame] | 2051 | return SetupAndTrackLayerOutputSlot<hal_1_2::HalPolicy>(operation, 0, *layer, model, data); |
Keith Davis | a6bc52f | 2019-06-26 09:39:49 +0100 | [diff] [blame] | 2052 | } |
| 2053 | |
Francis Murtagh | 074c25a | 2019-07-22 16:40:57 +0100 | [diff] [blame] | 2054 | bool HalPolicy::ConvertSoftmax(const Operation& operation, const Model& model, ConversionData& data) |
| 2055 | { |
Aron Virginas-Tar | 29404fb | 2019-07-24 13:55:31 +0100 | [diff] [blame] | 2056 | ALOGV("hal_1_2::HalPolicy::ConvertSoftmax()"); |
| 2057 | |
Francis Murtagh | 074c25a | 2019-07-22 16:40:57 +0100 | [diff] [blame] | 2058 | LayerInputHandle input = ConvertToLayerInputHandle<hal_1_2::HalPolicy>(operation, 0, model, data); |
| 2059 | if (!input.IsValid()) |
| 2060 | { |
| 2061 | return Fail("%s: Operation has invalid inputs", __func__); |
| 2062 | } |
| 2063 | |
| 2064 | const Operand* outputOperand = GetOutputOperand<hal_1_2::HalPolicy>(operation, 0, model); |
| 2065 | if (!outputOperand) |
| 2066 | { |
| 2067 | return Fail("%s: Operation has no outputs", __func__); |
| 2068 | } |
| 2069 | |
Teresa Charlin | 8f6429d | 2019-10-01 13:10:15 +0100 | [diff] [blame] | 2070 | const TensorInfo& outputInfo = GetTensorInfoForOperand(*outputOperand); |
Aron Virginas-Tar | 573a8fa | 2019-07-23 14:01:37 +0100 | [diff] [blame] | 2071 | if (IsDynamicTensor(outputInfo)) |
Francis Murtagh | 074c25a | 2019-07-22 16:40:57 +0100 | [diff] [blame] | 2072 | { |
Aron Virginas-Tar | b7421e5 | 2019-07-26 13:14:39 +0100 | [diff] [blame] | 2073 | return Fail("%s: Dynamic output tensors are not supported", __func__); |
Francis Murtagh | 074c25a | 2019-07-22 16:40:57 +0100 | [diff] [blame] | 2074 | } |
| 2075 | |
Teresa Charlin | 8f6429d | 2019-10-01 13:10:15 +0100 | [diff] [blame] | 2076 | SoftmaxDescriptor desc; |
Francis Murtagh | 074c25a | 2019-07-22 16:40:57 +0100 | [diff] [blame] | 2077 | if (!GetInputFloat32<hal_1_2::HalPolicy>(operation, 1, desc.m_Beta, model, data)) |
| 2078 | { |
| 2079 | return Fail("%s: Operation has invalid inputs", __func__); |
| 2080 | } |
| 2081 | |
| 2082 | if (operation.inputs.size() > 2 && !GetInputScalar<hal_1_2::HalPolicy>(operation, |
| 2083 | 2, |
| 2084 | HalPolicy::OperandType::INT32, |
| 2085 | desc.m_Axis, |
| 2086 | model, |
| 2087 | data)) |
| 2088 | { |
| 2089 | return Fail("%s: Operation has invalid inputs", __func__); |
| 2090 | } |
| 2091 | |
Narumol Prangnawarat | 52dc527 | 2019-08-06 17:34:26 +0100 | [diff] [blame] | 2092 | if (input.GetTensorInfo().GetNumDimensions() > 2 || |
| 2093 | !(desc.m_Axis == 1 || |
| 2094 | (desc.m_Axis < 0 && static_cast<int>(input.GetTensorInfo().GetNumDimensions()) + desc.m_Axis == 1))) |
| 2095 | { |
| 2096 | return Fail("%s: Unsupported input greater than 2D or axis != 1", __func__); |
| 2097 | } |
| 2098 | |
Francis Murtagh | 074c25a | 2019-07-22 16:40:57 +0100 | [diff] [blame] | 2099 | bool isSupported = false; |
| 2100 | FORWARD_LAYER_SUPPORT_FUNC(__func__, |
| 2101 | IsSoftmaxSupported, |
| 2102 | data.m_Backends, |
| 2103 | isSupported, |
| 2104 | input.GetTensorInfo(), |
| 2105 | outputInfo, |
| 2106 | desc); |
| 2107 | if (!isSupported) |
| 2108 | { |
| 2109 | return false; |
| 2110 | } |
| 2111 | |
Teresa Charlin | 8f6429d | 2019-10-01 13:10:15 +0100 | [diff] [blame] | 2112 | IConnectableLayer* layer = data.m_Network->AddSoftmaxLayer(desc); |
Francis Murtagh | 074c25a | 2019-07-22 16:40:57 +0100 | [diff] [blame] | 2113 | assert(layer != nullptr); |
| 2114 | input.Connect(layer->GetInputSlot(0)); |
| 2115 | |
Aron Virginas-Tar | b7421e5 | 2019-07-26 13:14:39 +0100 | [diff] [blame] | 2116 | return SetupAndTrackLayerOutputSlot<hal_1_2::HalPolicy>(operation, 0, *layer, model, data); |
Francis Murtagh | 074c25a | 2019-07-22 16:40:57 +0100 | [diff] [blame] | 2117 | } |
| 2118 | |
Mike Kelly | 0a87936 | 2019-07-29 16:56:31 +0100 | [diff] [blame] | 2119 | bool HalPolicy::ConvertSub(const Operation& operation, const Model& model, ConversionData& data) |
| 2120 | { |
| 2121 | ALOGV("hal_1_2::HalPolicy::ConvertSub()"); |
| 2122 | return ::ConvertSub<hal_1_2::HalPolicy>(operation, model, data); |
| 2123 | } |
| 2124 | |
Sadik Armagan | 6111316 | 2019-07-25 09:09:40 +0100 | [diff] [blame] | 2125 | bool HalPolicy::ConvertTanH(const Operation& operation, const Model& model, ConversionData& data) |
| 2126 | { |
| 2127 | ALOGV("hal_1_2::HalPolicy::ConvertTanH()"); |
| 2128 | return ::ConvertTanH<hal_1_2::HalPolicy>(operation, model, data); |
| 2129 | } |
| 2130 | |
Pablo Tello | 972603f | 2019-11-28 15:21:41 +0000 | [diff] [blame] | 2131 | template<typename HalPolicy, |
| 2132 | typename HalOperation = typename HalPolicy::Operation, |
| 2133 | typename HalModel = typename HalPolicy::Model> |
| 2134 | bool SetupAndTrackLayerOutputSlotAndOverrideTensorInfo(const HalOperation& operation, |
| 2135 | uint32_t operationOutputIndex, |
| 2136 | armnn::IConnectableLayer& layer, |
| 2137 | uint32_t layerOutputIndex, |
| 2138 | const HalModel& model, |
| 2139 | ConversionData& data, |
| 2140 | const armnn::TensorInfo tensor_info) |
| 2141 | { |
| 2142 | using HalOperand = typename HalPolicy::Operand; |
| 2143 | |
| 2144 | const HalOperand* outputOperand = GetOutputOperand<HalPolicy>(operation, operationOutputIndex, model); |
| 2145 | if ((outputOperand == nullptr) || (operationOutputIndex >= layer.GetNumOutputSlots())) |
| 2146 | { |
| 2147 | return false; |
| 2148 | } |
| 2149 | |
| 2150 | armnn::IOutputSlot& outputSlot = layer.GetOutputSlot(layerOutputIndex); |
| 2151 | |
| 2152 | const uint32_t operandIndex = operation.outputs[operationOutputIndex]; |
| 2153 | data.m_OutputSlotForOperand[operandIndex] = &outputSlot; |
| 2154 | |
| 2155 | outputSlot.SetTensorInfo(tensor_info); |
| 2156 | |
| 2157 | return true; |
| 2158 | } |
| 2159 | |
| 2160 | |
Ferran Balaguer | b2397fd | 2019-07-25 12:12:39 +0100 | [diff] [blame] | 2161 | bool HalPolicy::ConvertLstm(const Operation& operation, const Model& model, ConversionData& data) |
| 2162 | { |
Pablo Tello | fb45e2f | 2019-10-18 16:51:57 +0100 | [diff] [blame] | 2163 | ALOGV("hal_1_2::HalPolicy::ConvertLstm()"); |
| 2164 | |
Ferran Balaguer | b2397fd | 2019-07-25 12:12:39 +0100 | [diff] [blame] | 2165 | // Inputs: |
| 2166 | // 00: The input: A 2-D tensor of ANEURALNETWORKS_TENSOR_FLOAT32, of shape [batch_size, input_size], where |
| 2167 | // “batch_size” corresponds to the batching dimension, and “input_size” is the size of the input. |
| 2168 | LayerInputHandle input = ConvertToLayerInputHandle<hal_1_2::HalPolicy>(operation, 0, model, data); |
| 2169 | if (!input.IsValid()) |
| 2170 | { |
| 2171 | return Fail("%s: Could not read input 0: input", __func__); |
| 2172 | } |
| 2173 | // 18: The output state: A 2-D tensor of ANEURALNETWORKS_TENSOR_FLOAT32, of shape [batch_size, output_size]. |
| 2174 | LayerInputHandle outputStateIn = ConvertToLayerInputHandle<hal_1_2::HalPolicy>(operation, 18, model, data); |
| 2175 | if (!outputStateIn.IsValid()) |
| 2176 | { |
| 2177 | return Fail("%s: Could not read input 18: outputStateIn", __func__); |
| 2178 | } |
| 2179 | // 19: The cell state: A 2-D tensor of ANEURALNETWORKS_TENSOR_FLOAT32, of shape [batch_size, num_units]. |
| 2180 | LayerInputHandle cellStateIn = ConvertToLayerInputHandle<hal_1_2::HalPolicy>(operation, 19, model, data); |
| 2181 | if (!cellStateIn.IsValid()) |
| 2182 | { |
| 2183 | return Fail("%s: Could not read input 19: cellStateIn", __func__); |
| 2184 | } |
| 2185 | |
| 2186 | // Get the mandatory input tensors: |
| 2187 | // 02: The input-to-forget weights: A 2-D tensor of ANEURALNETWORKS_TENSOR_FLOAT32, of shape |
| 2188 | // [num_units, input_size]. |
| 2189 | const ConstTensorPin inputToForgetWeightsPin = |
Pablo Tello | fb45e2f | 2019-10-18 16:51:57 +0100 | [diff] [blame] | 2190 | (DequantizeAndMakeConstTensorPin<hal_1_2::HalPolicy>(operation, model, data, 2)); |
Ferran Balaguer | b2397fd | 2019-07-25 12:12:39 +0100 | [diff] [blame] | 2191 | // 03: The input-to-cell weights: A 2-D tensor of ANEURALNETWORKS_TENSOR_FLOAT32, of shape |
| 2192 | // [num_units, input_size]. |
| 2193 | const ConstTensorPin inputToCellWeightsPin = |
Pablo Tello | fb45e2f | 2019-10-18 16:51:57 +0100 | [diff] [blame] | 2194 | (DequantizeAndMakeConstTensorPin<hal_1_2::HalPolicy>(operation, model, data, 3)); |
Ferran Balaguer | b2397fd | 2019-07-25 12:12:39 +0100 | [diff] [blame] | 2195 | // 04: The input-to-output weights: A 2-D tensor of ANEURALNETWORKS_TENSOR_FLOAT32, of shape |
| 2196 | // [num_units, input_size]. |
| 2197 | const ConstTensorPin inputToOutputWeightsPin = |
Pablo Tello | fb45e2f | 2019-10-18 16:51:57 +0100 | [diff] [blame] | 2198 | (DequantizeAndMakeConstTensorPin<hal_1_2::HalPolicy>(operation, model, data, 4)); |
Ferran Balaguer | b2397fd | 2019-07-25 12:12:39 +0100 | [diff] [blame] | 2199 | // 06: The recurrent-to-forget weights: A 2-D tensor of ANEURALNETWORKS_TENSOR_FLOAT32, of shape |
| 2200 | // [num_units, output_size]. |
| 2201 | const ConstTensorPin recurrentToForgetWeightsPin = |
Pablo Tello | fb45e2f | 2019-10-18 16:51:57 +0100 | [diff] [blame] | 2202 | (DequantizeAndMakeConstTensorPin<hal_1_2::HalPolicy>(operation, model, data, 6)); |
Ferran Balaguer | b2397fd | 2019-07-25 12:12:39 +0100 | [diff] [blame] | 2203 | // 07: The recurrent-to-cell weights: A 2-D tensor of ANEURALNETWORKS_TENSOR_FLOAT32, of shape |
| 2204 | // [num_units, output_size]. |
| 2205 | const ConstTensorPin recurrentToCellWeightsPin = |
Pablo Tello | fb45e2f | 2019-10-18 16:51:57 +0100 | [diff] [blame] | 2206 | (DequantizeAndMakeConstTensorPin<hal_1_2::HalPolicy>(operation, model, data, 7)); |
Ferran Balaguer | b2397fd | 2019-07-25 12:12:39 +0100 | [diff] [blame] | 2207 | // 08: The recurrent-to-output weights: A 2-D tensor of ANEURALNETWORKS_TENSOR_FLOAT32, of shape |
| 2208 | // [num_units, output_size]. |
| 2209 | const ConstTensorPin recurrentToOutputWeightsPin = |
Pablo Tello | fb45e2f | 2019-10-18 16:51:57 +0100 | [diff] [blame] | 2210 | (DequantizeAndMakeConstTensorPin<hal_1_2::HalPolicy>(operation, model, data, 8)); |
Ferran Balaguer | b2397fd | 2019-07-25 12:12:39 +0100 | [diff] [blame] | 2211 | // 13: The forget gate bias: A 1-D tensor of ANEURALNETWORKS_TENSOR_FLOAT32, of shape [num_units]. |
| 2212 | const ConstTensorPin forgetGateBiasPin = |
| 2213 | ConvertOperationInputToConstTensorPin<hal_1_2::HalPolicy>(operation, 13, model, data); |
| 2214 | // 14: The cell bias: A 1-D tensor of ANEURALNETWORKS_TENSOR_FLOAT32, of shape [num_units]. |
| 2215 | const ConstTensorPin cellBiasPin = |
| 2216 | ConvertOperationInputToConstTensorPin<hal_1_2::HalPolicy>(operation, 14, model, data); |
| 2217 | // 15: The output gate bias: A 1-D tensor of ANEURALNETWORKS_TENSOR_FLOAT32, of shape [num_units]. |
| 2218 | const ConstTensorPin outputGateBiasPin = |
| 2219 | ConvertOperationInputToConstTensorPin<hal_1_2::HalPolicy>(operation, 15, model, data); |
| 2220 | |
| 2221 | if (!inputToForgetWeightsPin.IsValid() || |
| 2222 | !inputToCellWeightsPin.IsValid() || |
| 2223 | !inputToOutputWeightsPin.IsValid() || |
| 2224 | !recurrentToForgetWeightsPin.IsValid() || |
| 2225 | !recurrentToCellWeightsPin.IsValid() || |
| 2226 | !recurrentToOutputWeightsPin.IsValid() || |
| 2227 | !forgetGateBiasPin.IsValid() || |
| 2228 | !cellBiasPin.IsValid() || |
| 2229 | !outputGateBiasPin.IsValid()) |
| 2230 | { |
| 2231 | return Fail("%s: Operation has invalid tensor inputs", __func__); |
| 2232 | } |
| 2233 | |
| 2234 | // Get the optional input tensors: |
| 2235 | // 01: The input-to-input weights: Optional. A 2-D tensor of ANEURALNETWORKS_TENSOR_FLOAT32, of shape |
| 2236 | // [num_units, input_size], where “num_units” corresponds to the number of cell units. |
| 2237 | const ConstTensorPin inputToInputWeightsPin = |
Pablo Tello | fb45e2f | 2019-10-18 16:51:57 +0100 | [diff] [blame] | 2238 | (DequantizeAndMakeConstTensorPin<hal_1_2::HalPolicy>(operation, model, data, 1, true)); |
Ferran Balaguer | b2397fd | 2019-07-25 12:12:39 +0100 | [diff] [blame] | 2239 | // 05: The recurrent-to-input weights: Optional. A 2-D tensor of ANEURALNETWORKS_TENSOR_FLOAT32, of shape |
| 2240 | // [num_units, output_size], where “output_size” corresponds to either the number of cell units (i.e., |
| 2241 | // “num_units”), or the second dimension of the “projection_weights”, if defined. |
| 2242 | const ConstTensorPin recurrentToInputWeightsPin = |
Pablo Tello | fb45e2f | 2019-10-18 16:51:57 +0100 | [diff] [blame] | 2243 | (DequantizeAndMakeConstTensorPin<hal_1_2::HalPolicy>(operation, model, data, 5, true)); |
Ferran Balaguer | b2397fd | 2019-07-25 12:12:39 +0100 | [diff] [blame] | 2244 | // 09: The cell-to-input weights: Optional. A 1-D tensor of ANEURALNETWORKS_TENSOR_FLOAT32, of shape [num_units]. |
| 2245 | const ConstTensorPin cellToInputWeightsPin = |
Pablo Tello | fb45e2f | 2019-10-18 16:51:57 +0100 | [diff] [blame] | 2246 | (DequantizeAndMakeConstTensorPin<hal_1_2::HalPolicy>(operation, model, data, 9, true)); |
Ferran Balaguer | b2397fd | 2019-07-25 12:12:39 +0100 | [diff] [blame] | 2247 | // 10: The cell-to-forget weights: Optional. A 1-D tensor of ANEURALNETWORKS_TENSOR_FLOAT32, of shape [num_units]. |
| 2248 | const ConstTensorPin cellToForgetWeightsPin = |
Pablo Tello | fb45e2f | 2019-10-18 16:51:57 +0100 | [diff] [blame] | 2249 | (DequantizeAndMakeConstTensorPin<hal_1_2::HalPolicy>(operation, model, data, 10, true)); |
Ferran Balaguer | b2397fd | 2019-07-25 12:12:39 +0100 | [diff] [blame] | 2250 | // 11: The cell-to-output weights: Optional. A 1-D tensor of ANEURALNETWORKS_TENSOR_FLOAT32, of shape [num_units]. |
| 2251 | const ConstTensorPin cellToOutputWeightsPin = |
Pablo Tello | fb45e2f | 2019-10-18 16:51:57 +0100 | [diff] [blame] | 2252 | (DequantizeAndMakeConstTensorPin<hal_1_2::HalPolicy>(operation, model, data, 11, true)); |
Ferran Balaguer | b2397fd | 2019-07-25 12:12:39 +0100 | [diff] [blame] | 2253 | // 12: The input gate bias: Optional. A 1-D tensor of ANEURALNETWORKS_TENSOR_FLOAT32, of shape [num_units]. |
| 2254 | const ConstTensorPin inputGateBiasPin = |
| 2255 | ConvertOperationInputToConstTensorPin<hal_1_2::HalPolicy>(operation, |
| 2256 | 12, |
| 2257 | model, |
| 2258 | data, |
| 2259 | g_DontPermute, |
| 2260 | nullptr, |
| 2261 | true); |
| 2262 | |
| 2263 | // 16: The projection weights: Optional. A 2-D tensor of ANEURALNETWORKS_TENSOR_FLOAT32, of shape |
| 2264 | // [output_size, num_units]. |
| 2265 | const ConstTensorPin projectionWeightsPin = |
Pablo Tello | fb45e2f | 2019-10-18 16:51:57 +0100 | [diff] [blame] | 2266 | (DequantizeAndMakeConstTensorPin<hal_1_2::HalPolicy>(operation, model, data, 16, true)); |
Ferran Balaguer | b2397fd | 2019-07-25 12:12:39 +0100 | [diff] [blame] | 2267 | // 17: The projection bias: Optional. A 1-D tensor of ANEURALNETWORKS_TENSOR_FLOAT32, of shape [output_size]. |
| 2268 | const ConstTensorPin projectionBiasPin = |
| 2269 | ConvertOperationInputToConstTensorPin<hal_1_2::HalPolicy>(operation, |
| 2270 | 17, |
| 2271 | model, |
| 2272 | data, |
| 2273 | g_DontPermute, |
| 2274 | nullptr, |
| 2275 | true); |
| 2276 | |
| 2277 | if ((!inputToInputWeightsPin.IsValid() && !inputToInputWeightsPin.IsOptional()) || |
| 2278 | (!recurrentToInputWeightsPin.IsValid() && !recurrentToInputWeightsPin.IsOptional()) || |
| 2279 | (!cellToInputWeightsPin.IsValid() && !cellToInputWeightsPin.IsOptional()) || |
| 2280 | (!cellToForgetWeightsPin.IsValid() && !cellToForgetWeightsPin.IsOptional()) || |
| 2281 | (!cellToOutputWeightsPin.IsValid() && !cellToOutputWeightsPin.IsOptional()) || |
| 2282 | (!inputGateBiasPin.IsValid() && !inputGateBiasPin.IsOptional()) || |
| 2283 | (!projectionWeightsPin.IsValid() && !projectionWeightsPin.IsOptional()) || |
| 2284 | (!projectionBiasPin.IsValid() && !projectionBiasPin.IsOptional())) |
| 2285 | { |
| 2286 | return Fail("%s: Operation has invalid tensor inputs", __func__); |
| 2287 | } |
| 2288 | |
| 2289 | // Get the mandatory input scalars (actually 1-D tensors of size 1): |
| 2290 | // 20: The activation function: A value indicating the activation function: |
| 2291 | // 0: None; 1: Relu; 3: Relu6; 4: Tanh; 6: Sigmoid. |
| 2292 | // 21: The clipping threshold: for the cell state, such that values are bound within [-cell_clip, cell_clip]. |
| 2293 | // If set to 0.0 then clipping is disabled. |
| 2294 | // 22: The clipping threshold: for the output from the projection layer, such that values are bound within |
| 2295 | // [-proj_clip, proj_clip]. If set to 0.0 then clipping is disabled. |
| 2296 | ActivationFn activation; |
| 2297 | float cellClip; |
| 2298 | float projClip; |
| 2299 | if (!GetInputActivationFunctionFromTensor<hal_1_2::HalPolicy>(operation, 20, activation, model, data) || |
| 2300 | !GetInputScalar<hal_1_2::HalPolicy>(operation, 21, OperandType::FLOAT32, cellClip, model, data) || |
| 2301 | !GetInputScalar<hal_1_2::HalPolicy>(operation, 22, OperandType::FLOAT32, projClip, model, data)) |
| 2302 | { |
| 2303 | return Fail("%s: Operation has invalid scalar inputs", __func__); |
| 2304 | } |
| 2305 | |
| 2306 | // Get the normalization tensors |
| 2307 | // 23: The input layer normalization weights. A 1-D tensor of shape [num_units]. |
| 2308 | // Used to rescale normalized inputs to activation at input gate. |
Pablo Tello | fb45e2f | 2019-10-18 16:51:57 +0100 | [diff] [blame] | 2309 | const ConstTensorPin inputLayerNormWeightsPin |
| 2310 | (DequantizeAndMakeConstTensorPin<hal_1_2::HalPolicy>(operation, model, data, 23, true)); |
Ferran Balaguer | b2397fd | 2019-07-25 12:12:39 +0100 | [diff] [blame] | 2311 | |
| 2312 | // 24: The forget layer normalization weights. A 1-D tensor of shape [num_units]. |
| 2313 | // Used to rescale normalized inputs to activation at forget gate. |
| 2314 | const ConstTensorPin forgetLayerNormWeightsPin = |
| 2315 | ConvertOperationInputToConstTensorPin<hal_1_2::HalPolicy>(operation, |
| 2316 | 24, |
| 2317 | model, |
| 2318 | data, |
| 2319 | g_DontPermute, |
| 2320 | nullptr, |
| 2321 | true); |
| 2322 | |
| 2323 | // 25: The cell layer normalization weights. A 1-D tensor of shape [num_units]. |
| 2324 | // Used to rescale normalized inputs to activation at cell gate. |
| 2325 | const ConstTensorPin cellLayerNormWeightsPin = |
| 2326 | ConvertOperationInputToConstTensorPin<hal_1_2::HalPolicy>(operation, |
| 2327 | 25, |
| 2328 | model, |
| 2329 | data, |
| 2330 | g_DontPermute, |
| 2331 | nullptr, |
| 2332 | true); |
| 2333 | |
| 2334 | // 26: The output layer normalization weights. A 1-D tensor of shape [num_units]. |
| 2335 | // Used to rescale normalized inputs to activation at output gate. |
| 2336 | const ConstTensorPin outputLayerNormWeightsPin = |
| 2337 | ConvertOperationInputToConstTensorPin<hal_1_2::HalPolicy>(operation, |
| 2338 | 26, |
| 2339 | model, |
| 2340 | data, |
| 2341 | g_DontPermute, |
| 2342 | nullptr, |
| 2343 | true); |
| 2344 | |
| 2345 | // Outputs: |
| 2346 | // 00: The scratch buffer: A 2-D tensor of ANEURALNETWORKS_TENSOR_FLOAT32, of shape [batch_size, num_units * 4] |
| 2347 | // with CIFG, or [batch_size, num_units * 3] without CIFG. |
| 2348 | const Operand* scratchBuffer = GetOutputOperand<hal_1_2::HalPolicy>(operation, 0, model); |
| 2349 | if (!scratchBuffer) |
| 2350 | { |
| 2351 | return Fail("%s: Could not read output 0: scratchBuffer", __func__); |
| 2352 | } |
| 2353 | // 01: The output state (out): A 2-D tensor of ANEURALNETWORKS_TENSOR_FLOAT32, of shape [batch_size, output_size]. |
| 2354 | const Operand* outputStateOut = GetOutputOperand<hal_1_2::HalPolicy>(operation, 1, model); |
| 2355 | if (!outputStateOut) |
| 2356 | { |
| 2357 | return Fail("%s: Could not read output 1: outputStateOut", __func__); |
| 2358 | } |
| 2359 | // 02: The cell state (out): A 2-D tensor of ANEURALNETWORKS_TENSOR_FLOAT32, of shape [batch_size, num_units]. |
| 2360 | const Operand* cellStateOut = GetOutputOperand<hal_1_2::HalPolicy>(operation, 2, model); |
| 2361 | if (!cellStateOut) |
| 2362 | { |
| 2363 | return Fail("%s: Could not read output 2: cellStateOut", __func__); |
| 2364 | } |
| 2365 | // 03: The output: A 2-D tensor of ANEURALNETWORKS_TENSOR_FLOAT32, of shape [batch_size, output_size]. This is |
| 2366 | // effectively the same as the current “output state (out)” value. |
| 2367 | const Operand* output = GetOutputOperand<hal_1_2::HalPolicy>(operation, 3, model); |
| 2368 | if (!output) |
| 2369 | { |
| 2370 | return Fail("%s: Could not read output 3: output", __func__); |
| 2371 | } |
| 2372 | |
| 2373 | // set the params structure for the AddLstmLayer call |
Teresa Charlin | 8f6429d | 2019-10-01 13:10:15 +0100 | [diff] [blame] | 2374 | LstmInputParams params; |
Ferran Balaguer | b2397fd | 2019-07-25 12:12:39 +0100 | [diff] [blame] | 2375 | params.m_InputToInputWeights = inputToInputWeightsPin.GetConstTensorPtr(); |
| 2376 | params.m_InputToForgetWeights = inputToForgetWeightsPin.GetConstTensorPtr(); |
| 2377 | params.m_InputToCellWeights = inputToCellWeightsPin.GetConstTensorPtr(); |
| 2378 | params.m_InputToOutputWeights = inputToOutputWeightsPin.GetConstTensorPtr(); |
| 2379 | params.m_RecurrentToInputWeights = recurrentToInputWeightsPin.GetConstTensorPtr(); |
| 2380 | params.m_RecurrentToForgetWeights = recurrentToForgetWeightsPin.GetConstTensorPtr(); |
| 2381 | params.m_RecurrentToCellWeights = recurrentToCellWeightsPin.GetConstTensorPtr(); |
| 2382 | params.m_RecurrentToOutputWeights = recurrentToOutputWeightsPin.GetConstTensorPtr(); |
| 2383 | params.m_CellToInputWeights = cellToInputWeightsPin.GetConstTensorPtr(); |
| 2384 | params.m_CellToForgetWeights = cellToForgetWeightsPin.GetConstTensorPtr(); |
| 2385 | params.m_CellToOutputWeights = cellToOutputWeightsPin.GetConstTensorPtr(); |
| 2386 | params.m_InputGateBias = inputGateBiasPin.GetConstTensorPtr(); |
| 2387 | params.m_ForgetGateBias = forgetGateBiasPin.GetConstTensorPtr(); |
| 2388 | params.m_CellBias = cellBiasPin.GetConstTensorPtr(); |
| 2389 | params.m_OutputGateBias = outputGateBiasPin.GetConstTensorPtr(); |
| 2390 | params.m_ProjectionWeights = projectionWeightsPin.GetConstTensorPtr(); |
| 2391 | params.m_ProjectionBias = projectionBiasPin.GetConstTensorPtr(); |
| 2392 | params.m_InputLayerNormWeights = inputLayerNormWeightsPin.GetConstTensorPtr(); |
| 2393 | params.m_ForgetLayerNormWeights = forgetLayerNormWeightsPin.GetConstTensorPtr(); |
| 2394 | params.m_CellLayerNormWeights = cellLayerNormWeightsPin.GetConstTensorPtr(); |
| 2395 | params.m_OutputLayerNormWeights = outputLayerNormWeightsPin.GetConstTensorPtr(); |
| 2396 | |
| 2397 | // set the layer descriptor |
Teresa Charlin | 8f6429d | 2019-10-01 13:10:15 +0100 | [diff] [blame] | 2398 | LstmDescriptor desc; |
Ferran Balaguer | b2397fd | 2019-07-25 12:12:39 +0100 | [diff] [blame] | 2399 | desc.m_ActivationFunc = activation; |
| 2400 | desc.m_ClippingThresCell = cellClip; |
| 2401 | desc.m_ClippingThresProj = projClip; |
| 2402 | desc.m_CifgEnabled = (params.m_InputToInputWeights == nullptr || |
| 2403 | params.m_RecurrentToInputWeights == nullptr || |
| 2404 | params.m_InputGateBias == nullptr); |
| 2405 | desc.m_PeepholeEnabled = (params.m_CellToForgetWeights != nullptr || |
| 2406 | params.m_CellToOutputWeights != nullptr); |
| 2407 | desc.m_ProjectionEnabled = (params.m_ProjectionWeights != nullptr); |
| 2408 | desc.m_LayerNormEnabled = (params.m_InputLayerNormWeights != nullptr || |
| 2409 | params.m_ForgetLayerNormWeights != nullptr || |
| 2410 | params.m_CellLayerNormWeights != nullptr || |
| 2411 | params.m_OutputLayerNormWeights != nullptr); |
| 2412 | |
| 2413 | // validate the optional input groups |
| 2414 | if (desc.m_CifgEnabled && |
| 2415 | (params.m_InputToInputWeights != nullptr || |
| 2416 | params.m_RecurrentToInputWeights != nullptr || |
| 2417 | params.m_InputGateBias != nullptr)) |
| 2418 | { |
| 2419 | return Fail("%s: All, or none, of input-to-input weights, recurrent-to-input weights," |
| 2420 | " and input gate bias must be provided", __func__); |
| 2421 | } |
| 2422 | |
| 2423 | if (!desc.m_ProjectionEnabled && params.m_ProjectionBias != nullptr) |
| 2424 | { |
| 2425 | return Fail("%s: projection bias should not be provided without projection weights", __func__); |
| 2426 | } |
| 2427 | |
| 2428 | if (desc.m_PeepholeEnabled && |
| 2429 | (params.m_CellToForgetWeights == nullptr || |
| 2430 | params.m_CellToOutputWeights == nullptr || |
| 2431 | (!desc.m_CifgEnabled && params.m_CellToInputWeights == nullptr))) |
| 2432 | { |
| 2433 | return Fail("%s: All, or none, of cell-to-forget weights and cell-to-output weights must be provided" |
| 2434 | " and, if CIFG is not enabled, cell-to-input weights must also be provided", __func__); |
| 2435 | } |
| 2436 | |
| 2437 | if (desc.m_LayerNormEnabled && |
| 2438 | (params.m_ForgetLayerNormWeights == nullptr || |
| 2439 | params.m_CellLayerNormWeights == nullptr || |
| 2440 | params.m_OutputLayerNormWeights == nullptr || |
| 2441 | (!desc.m_CifgEnabled && params.m_InputLayerNormWeights == nullptr))) |
| 2442 | { |
| 2443 | return Fail("%s: All, or none, of forget-norm weights, cell-norm weights and output-norm weights must be" |
| 2444 | " provided and, if CIFG is not enabled, input-norm weights must also be provided", __func__); |
| 2445 | } |
| 2446 | |
| 2447 | // Check if the layer is supported |
| 2448 | // Inputs |
Teresa Charlin | 8f6429d | 2019-10-01 13:10:15 +0100 | [diff] [blame] | 2449 | const TensorInfo& inputInfo = input.GetTensorInfo(); |
| 2450 | const TensorInfo& outputStateInInfo = outputStateIn.GetTensorInfo(); |
| 2451 | const TensorInfo& cellStateInInfo = cellStateIn.GetTensorInfo(); |
Ferran Balaguer | b2397fd | 2019-07-25 12:12:39 +0100 | [diff] [blame] | 2452 | |
| 2453 | // Outputs |
Teresa Charlin | 8f6429d | 2019-10-01 13:10:15 +0100 | [diff] [blame] | 2454 | const TensorInfo& scratchBufferInfo = GetTensorInfoForOperand(*scratchBuffer); |
| 2455 | const TensorInfo& outputStateOutInfo = GetTensorInfoForOperand(*outputStateOut); |
| 2456 | const TensorInfo& cellStateOutInfo = GetTensorInfoForOperand(*cellStateOut); |
| 2457 | const TensorInfo& outputInfo = GetTensorInfoForOperand(*output); |
Ferran Balaguer | b2397fd | 2019-07-25 12:12:39 +0100 | [diff] [blame] | 2458 | |
Pablo Tello | 972603f | 2019-11-28 15:21:41 +0000 | [diff] [blame] | 2459 | // Check if the scratch buffer shape was initialized, |
| 2460 | // In some cases the shape could be (0,0) which requires the driver |
| 2461 | // to infer the shape and set it up accordingly. |
| 2462 | // The code below does that. |
| 2463 | TensorInfo fixSbInfo = scratchBufferInfo; |
| 2464 | if (IsDynamicTensor(scratchBufferInfo)) |
| 2465 | { |
| 2466 | auto & s = fixSbInfo.GetShape(); |
| 2467 | s[0] = outputStateInInfo.GetShape()[0]; |
| 2468 | if (desc.m_CifgEnabled) |
| 2469 | { |
| 2470 | // 2D tensor with dimensions [num_units * 3, batch_size] with CIFG |
| 2471 | s[1] = cellStateOutInfo.GetShape()[1]*3; |
| 2472 | } |
| 2473 | else |
| 2474 | { |
| 2475 | // scratch_buffer [num_units * 4, batch_size] without CIFG |
| 2476 | s[1] = cellStateOutInfo.GetShape()[1]*4; |
| 2477 | } |
| 2478 | } |
| 2479 | |
| 2480 | if (IsDynamicTensor(outputStateOutInfo) || |
Ferran Balaguer | a4a629a | 2019-07-30 10:16:13 +0100 | [diff] [blame] | 2481 | IsDynamicTensor(cellStateOutInfo) || |
| 2482 | IsDynamicTensor(outputInfo)) |
| 2483 | { |
Pablo Tello | fb45e2f | 2019-10-18 16:51:57 +0100 | [diff] [blame] | 2484 | return Fail("%s: Dynamic output tensors are not supported %d %d %d %d", __func__, |
| 2485 | IsDynamicTensor(scratchBufferInfo), IsDynamicTensor(outputStateOutInfo), |
| 2486 | IsDynamicTensor(cellStateOutInfo), IsDynamicTensor(outputInfo)); |
Ferran Balaguer | a4a629a | 2019-07-30 10:16:13 +0100 | [diff] [blame] | 2487 | } |
| 2488 | |
Ferran Balaguer | b2397fd | 2019-07-25 12:12:39 +0100 | [diff] [blame] | 2489 | // Basic parameters |
Teresa Charlin | 8f6429d | 2019-10-01 13:10:15 +0100 | [diff] [blame] | 2490 | LstmInputParamsInfo paramsInfo; |
Ferran Balaguer | b2397fd | 2019-07-25 12:12:39 +0100 | [diff] [blame] | 2491 | paramsInfo.m_InputToForgetWeights = &(params.m_InputToForgetWeights->GetInfo()); |
| 2492 | paramsInfo.m_InputToCellWeights = &(params.m_InputToCellWeights->GetInfo()); |
| 2493 | paramsInfo.m_InputToOutputWeights = &(params.m_InputToOutputWeights->GetInfo()); |
| 2494 | paramsInfo.m_RecurrentToForgetWeights = &(params.m_RecurrentToForgetWeights->GetInfo()); |
| 2495 | paramsInfo.m_RecurrentToCellWeights = &(params.m_RecurrentToCellWeights->GetInfo()); |
| 2496 | paramsInfo.m_RecurrentToOutputWeights = &(params.m_RecurrentToOutputWeights->GetInfo()); |
| 2497 | paramsInfo.m_ForgetGateBias = &(params.m_ForgetGateBias->GetInfo()); |
| 2498 | paramsInfo.m_CellBias = &(params.m_CellBias->GetInfo()); |
| 2499 | paramsInfo.m_OutputGateBias = &(params.m_OutputGateBias->GetInfo()); |
| 2500 | |
| 2501 | // Optional parameters |
| 2502 | if(!desc.m_CifgEnabled) |
| 2503 | { |
| 2504 | paramsInfo.m_InputToInputWeights = &(params.m_InputToInputWeights->GetInfo()); |
| 2505 | paramsInfo.m_RecurrentToInputWeights = &(params.m_RecurrentToInputWeights->GetInfo()); |
| 2506 | if (params.m_CellToInputWeights != nullptr) |
| 2507 | { |
| 2508 | paramsInfo.m_CellToInputWeights = &(params.m_CellToInputWeights->GetInfo()); |
| 2509 | } |
| 2510 | paramsInfo.m_InputGateBias = &(params.m_InputGateBias->GetInfo()); |
| 2511 | } |
| 2512 | |
| 2513 | if(desc.m_ProjectionEnabled) |
| 2514 | { |
| 2515 | paramsInfo.m_ProjectionWeights = &(params.m_ProjectionWeights->GetInfo()); |
| 2516 | if (params.m_ProjectionBias != nullptr) |
| 2517 | { |
| 2518 | paramsInfo.m_ProjectionBias = &(params.m_ProjectionBias->GetInfo()); |
| 2519 | } |
| 2520 | } |
| 2521 | |
| 2522 | if(desc.m_PeepholeEnabled) |
| 2523 | { |
| 2524 | paramsInfo.m_CellToForgetWeights = &(params.m_CellToForgetWeights->GetInfo()); |
| 2525 | paramsInfo.m_CellToOutputWeights = &(params.m_CellToOutputWeights->GetInfo()); |
| 2526 | } |
| 2527 | |
| 2528 | if (desc.m_LayerNormEnabled) |
| 2529 | { |
| 2530 | if(!desc.m_CifgEnabled) |
| 2531 | { |
| 2532 | paramsInfo.m_InputLayerNormWeights = &(params.m_InputLayerNormWeights->GetInfo()); |
| 2533 | } |
| 2534 | paramsInfo.m_ForgetLayerNormWeights = &(params.m_ForgetLayerNormWeights->GetInfo()); |
| 2535 | paramsInfo.m_CellLayerNormWeights = &(params.m_CellLayerNormWeights->GetInfo()); |
| 2536 | paramsInfo.m_OutputLayerNormWeights = &(params.m_OutputLayerNormWeights->GetInfo()); |
| 2537 | } |
| 2538 | |
| 2539 | bool isSupported = false; |
| 2540 | FORWARD_LAYER_SUPPORT_FUNC(__func__, |
| 2541 | IsLstmSupported, |
| 2542 | data.m_Backends, |
| 2543 | isSupported, |
| 2544 | inputInfo, |
| 2545 | outputStateInInfo, |
| 2546 | cellStateInInfo, |
Pablo Tello | 972603f | 2019-11-28 15:21:41 +0000 | [diff] [blame] | 2547 | fixSbInfo, |
Ferran Balaguer | b2397fd | 2019-07-25 12:12:39 +0100 | [diff] [blame] | 2548 | outputStateOutInfo, |
| 2549 | cellStateOutInfo, |
| 2550 | outputInfo, |
| 2551 | desc, |
| 2552 | paramsInfo); |
| 2553 | if (!isSupported) |
| 2554 | { |
| 2555 | return false; |
| 2556 | } |
| 2557 | |
| 2558 | // Add the layer |
Teresa Charlin | 8f6429d | 2019-10-01 13:10:15 +0100 | [diff] [blame] | 2559 | IConnectableLayer* layer = data.m_Network->AddLstmLayer(desc, params, "Lstm"); |
Ferran Balaguer | b2397fd | 2019-07-25 12:12:39 +0100 | [diff] [blame] | 2560 | |
| 2561 | input.Connect(layer->GetInputSlot(0)); |
| 2562 | outputStateIn.Connect(layer->GetInputSlot(1)); |
| 2563 | cellStateIn.Connect(layer->GetInputSlot(2)); |
| 2564 | |
Pablo Tello | 972603f | 2019-11-28 15:21:41 +0000 | [diff] [blame] | 2565 | |
| 2566 | return ( |
| 2567 | (IsDynamicTensor(scratchBufferInfo)? |
| 2568 | SetupAndTrackLayerOutputSlotAndOverrideTensorInfo<hal_1_2::HalPolicy>( |
| 2569 | operation, 0, *layer, 0, model, data,fixSbInfo): |
| 2570 | SetupAndTrackLayerOutputSlot<hal_1_2::HalPolicy>( |
| 2571 | operation, 0, *layer, 0, model, data)) && |
Ferran Balaguer | b2397fd | 2019-07-25 12:12:39 +0100 | [diff] [blame] | 2572 | SetupAndTrackLayerOutputSlot<hal_1_2::HalPolicy>(operation, 1, *layer, 1, model, data) && |
| 2573 | SetupAndTrackLayerOutputSlot<hal_1_2::HalPolicy>(operation, 2, *layer, 2, model, data) && |
| 2574 | SetupAndTrackLayerOutputSlot<hal_1_2::HalPolicy>(operation, 3, *layer, 3, model, data)); |
| 2575 | } |
| 2576 | |
Sadik Armagan | 701d9a0 | 2019-09-04 15:16:18 +0100 | [diff] [blame] | 2577 | bool HalPolicy::ConvertSqrt(const Operation& operation, const Model& model, ConversionData& data) |
| 2578 | { |
| 2579 | ALOGV("hal_1_2::HalPolicy::ConvertSqrt()"); |
Teresa Charlin | 8f6429d | 2019-10-01 13:10:15 +0100 | [diff] [blame] | 2580 | ActivationDescriptor desc; |
| 2581 | desc.m_Function = ActivationFunction::Sqrt; |
Sadik Armagan | 701d9a0 | 2019-09-04 15:16:18 +0100 | [diff] [blame] | 2582 | |
| 2583 | return ::ConvertToActivation<hal_1_2::HalPolicy>(operation, __func__, desc, model, data); |
| 2584 | } |
| 2585 | |
Mike Kelly | 4627280 | 2019-08-14 17:00:48 +0100 | [diff] [blame] | 2586 | bool HalPolicy::ConvertSqueeze(const Operation& operation, const Model& model, ConversionData& data) |
| 2587 | { |
Sadik Armagan | 701d9a0 | 2019-09-04 15:16:18 +0100 | [diff] [blame] | 2588 | ALOGV("hal_1_2::HalPolicy::ConvertSqueeze()"); |
Mike Kelly | 4627280 | 2019-08-14 17:00:48 +0100 | [diff] [blame] | 2589 | return ::ConvertSqueeze<hal_1_2::HalPolicy>(operation, model, data); |
| 2590 | } |
| 2591 | |
| 2592 | bool HalPolicy::ConvertStridedSlice(const Operation& operation, const Model& model, ConversionData& data) |
| 2593 | { |
Sadik Armagan | 701d9a0 | 2019-09-04 15:16:18 +0100 | [diff] [blame] | 2594 | ALOGV("hal_1_2::HalPolicy::ConvertStridedSlice()"); |
Mike Kelly | 4627280 | 2019-08-14 17:00:48 +0100 | [diff] [blame] | 2595 | return ::ConvertStridedSlice<hal_1_2::HalPolicy>(operation, model, data); |
| 2596 | } |
| 2597 | |
| 2598 | bool HalPolicy::ConvertTranspose(const Operation& operation, const Model& model, ConversionData& data) |
| 2599 | { |
Sadik Armagan | 701d9a0 | 2019-09-04 15:16:18 +0100 | [diff] [blame] | 2600 | ALOGV("hal_1_2::HalPolicy::ConvertTranspose()"); |
Mike Kelly | 4627280 | 2019-08-14 17:00:48 +0100 | [diff] [blame] | 2601 | return ::ConvertTranspose<hal_1_2::HalPolicy>(operation, model, data); |
| 2602 | } |
| 2603 | |
Aron Virginas-Tar | 8b99168 | 2019-07-31 12:54:59 +0100 | [diff] [blame] | 2604 | bool HalPolicy::ConvertTransposeConv2d(const Operation& operation, const Model& model, ConversionData& data) |
David Monahan | 613b49c | 2019-06-27 11:37:47 +0100 | [diff] [blame] | 2605 | { |
| 2606 | LayerInputHandle input = ConvertToLayerInputHandle<hal_1_2::HalPolicy>(operation, 0, model, data); |
| 2607 | |
| 2608 | if (!input.IsValid()) |
| 2609 | { |
| 2610 | return Fail("%s: Operation has invalid inputs", __func__); |
| 2611 | } |
| 2612 | |
| 2613 | const Operand* output = GetOutputOperand<hal_1_2::HalPolicy>(operation, 0, model); |
| 2614 | |
| 2615 | if (!output) |
| 2616 | { |
| 2617 | return Fail("%s: Could not read output 0", __func__); |
| 2618 | } |
| 2619 | |
Teresa Charlin | 8f6429d | 2019-10-01 13:10:15 +0100 | [diff] [blame] | 2620 | const TensorInfo& inputInfo = input.GetTensorInfo(); |
| 2621 | const TensorInfo& outputInfo = GetTensorInfoForOperand(*output); |
David Monahan | 613b49c | 2019-06-27 11:37:47 +0100 | [diff] [blame] | 2622 | if (IsDynamicTensor(outputInfo)) |
| 2623 | { |
| 2624 | return Fail("%s: Dynamic output tensors are not supported", __func__); |
| 2625 | } |
| 2626 | |
| 2627 | // ArmNN does not currently support non-fixed weights or bias |
| 2628 | // Find the shape of the weights tensor. In AndroidNN this will be [ 1, H, W, I * M ] |
| 2629 | const Operand* weightsOperand = GetInputOperand<hal_1_2::HalPolicy>(operation, 1, model); |
| 2630 | |
| 2631 | if (weightsOperand == nullptr) |
| 2632 | { |
| 2633 | return Fail("%s: Operand is invalid", __func__); |
| 2634 | } |
Teresa Charlin | 8f6429d | 2019-10-01 13:10:15 +0100 | [diff] [blame] | 2635 | TransposeConvolution2dDescriptor desc; |
| 2636 | desc.m_DataLayout = DataLayout::NHWC; |
David Monahan | 613b49c | 2019-06-27 11:37:47 +0100 | [diff] [blame] | 2637 | |
| 2638 | // Determine whether padding is implicit or explicit |
| 2639 | bool implicitPadding = operation.inputs.size() == 9; |
| 2640 | |
| 2641 | if (implicitPadding ) |
| 2642 | { |
| 2643 | desc.m_DataLayout = OptionalDataLayout<hal_1_2::HalPolicy>(operation, 8, model, data); |
| 2644 | } |
| 2645 | else |
| 2646 | { |
| 2647 | desc.m_DataLayout = OptionalDataLayout<hal_1_2::HalPolicy>(operation, 10, model, data); |
| 2648 | } |
| 2649 | |
| 2650 | armnnUtils::DataLayoutIndexed dataLayoutIndexed(desc.m_DataLayout); |
| 2651 | unsigned int widthIndex = dataLayoutIndexed.GetWidthIndex(); |
| 2652 | unsigned int heightIndex = dataLayoutIndexed.GetHeightIndex(); |
| 2653 | |
Teresa Charlin | 8f6429d | 2019-10-01 13:10:15 +0100 | [diff] [blame] | 2654 | const PermutationVector OHWIToOIHW = {0, 2, 3, 1}; |
David Monahan | 613b49c | 2019-06-27 11:37:47 +0100 | [diff] [blame] | 2655 | |
| 2656 | // The shape of the weight is [depth_out, filter_height, filter_width, depth_in]. |
| 2657 | // We have to permute it to OIHW if the data layout is NCHW. |
Teresa Charlin | 8f6429d | 2019-10-01 13:10:15 +0100 | [diff] [blame] | 2658 | const ConstTensorPin weightsPin = (desc.m_DataLayout == DataLayout::NCHW) ? |
David Monahan | 613b49c | 2019-06-27 11:37:47 +0100 | [diff] [blame] | 2659 | ConvertOperationInputToConstTensorPin<hal_1_2::HalPolicy>(operation, 1, model, data, OHWIToOIHW) : |
| 2660 | ConvertOperationInputToConstTensorPin<hal_1_2::HalPolicy>(operation, 1, model, data); |
| 2661 | |
| 2662 | // Bias is a 1D tensor |
| 2663 | const ConstTensorPin biasPin = |
| 2664 | ConvertOperationInputToConstTensorPin<hal_1_2::HalPolicy>(operation, 2, model, data); |
| 2665 | |
| 2666 | if (!weightsPin.IsValid()) |
| 2667 | { |
| 2668 | return Fail("%s: Operation has invalid weights", __func__); |
| 2669 | } |
| 2670 | |
| 2671 | if (!biasPin.IsValid()) |
| 2672 | { |
| 2673 | return Fail("%s: Operation has invalid biases", __func__); |
| 2674 | } |
| 2675 | |
Teresa Charlin | 8f6429d | 2019-10-01 13:10:15 +0100 | [diff] [blame] | 2676 | ConstTensor weights = weightsPin.GetConstTensor(); |
| 2677 | ConstTensor bias = biasPin.GetConstTensor(); |
David Monahan | 613b49c | 2019-06-27 11:37:47 +0100 | [diff] [blame] | 2678 | SanitizeBiasQuantizationScale(bias.GetInfo(), weights.GetInfo(), inputInfo); |
| 2679 | |
| 2680 | ActivationFn activation; |
| 2681 | |
| 2682 | if (implicitPadding) |
| 2683 | { |
Sadik Armagan | 3e3003e | 2019-08-13 12:54:34 +0100 | [diff] [blame] | 2684 | int32_t strideX{0}; |
| 2685 | int32_t strideY{0}; |
| 2686 | int32_t padLeft{0}; |
| 2687 | int32_t padRight{0}; |
| 2688 | int32_t padTop{0}; |
| 2689 | int32_t padBottom{0}; |
| 2690 | |
David Monahan | 613b49c | 2019-06-27 11:37:47 +0100 | [diff] [blame] | 2691 | android::nn::PaddingScheme paddingScheme; |
| 2692 | if (!GetInputPaddingScheme<hal_1_2::HalPolicy>(operation, 4, paddingScheme, model, data) || |
Sadik Armagan | 3e3003e | 2019-08-13 12:54:34 +0100 | [diff] [blame] | 2693 | !GetInputScalar<hal_1_2::HalPolicy>(operation, 5, OperandType::INT32, strideX, model, data) || |
| 2694 | !GetInputScalar<hal_1_2::HalPolicy>(operation, 6, OperandType::INT32, strideY, model, data) || |
David Monahan | 613b49c | 2019-06-27 11:37:47 +0100 | [diff] [blame] | 2695 | !GetInputActivationFunction<hal_1_2::HalPolicy>(operation, 7, activation, model, data)) |
| 2696 | { |
| 2697 | return Fail("%s: Operation has invalid inputs (implicit padding)", __func__); |
| 2698 | } |
| 2699 | |
| 2700 | const uint32_t kernelX = weights.GetShape()[widthIndex]; |
| 2701 | const uint32_t kernelY = weights.GetShape()[heightIndex]; |
Mike Kelly | 26123db | 2020-01-15 10:02:33 +0000 | [diff] [blame] | 2702 | const uint32_t outputX = outputInfo.GetShape()[widthIndex]; |
| 2703 | const uint32_t outputY = outputInfo.GetShape()[heightIndex]; |
David Monahan | 613b49c | 2019-06-27 11:37:47 +0100 | [diff] [blame] | 2704 | |
Mike Kelly | 26123db | 2020-01-15 10:02:33 +0000 | [diff] [blame] | 2705 | CalcPaddingTransposeConv(outputX, kernelX, strideX, padLeft, padRight, paddingScheme); |
| 2706 | CalcPaddingTransposeConv(outputY, kernelY, strideY, padTop, padBottom, paddingScheme); |
Narumol Prangnawarat | c8bdb39 | 2019-08-01 15:51:44 +0100 | [diff] [blame] | 2707 | |
| 2708 | // NOTE: The Android NN API allows for negative padding values in TransposeConv2d, |
| 2709 | // but Arm NN only supports values >= 0 |
| 2710 | if (padLeft < 0 || padRight < 0 || padTop < 0 || padBottom < 0) |
| 2711 | { |
| 2712 | return Fail("%s: Negative padding values are not supported", __func__); |
| 2713 | } |
| 2714 | |
Sadik Armagan | 3e3003e | 2019-08-13 12:54:34 +0100 | [diff] [blame] | 2715 | desc.m_StrideX = boost::numeric_cast<uint32_t>(strideX); |
| 2716 | desc.m_StrideY = boost::numeric_cast<uint32_t>(strideY); |
Narumol Prangnawarat | c8bdb39 | 2019-08-01 15:51:44 +0100 | [diff] [blame] | 2717 | desc.m_PadLeft = boost::numeric_cast<uint32_t>(padLeft); |
| 2718 | desc.m_PadRight = boost::numeric_cast<uint32_t>(padRight); |
| 2719 | desc.m_PadTop = boost::numeric_cast<uint32_t>(padTop); |
| 2720 | desc.m_PadBottom = boost::numeric_cast<uint32_t>(padBottom); |
David Monahan | 613b49c | 2019-06-27 11:37:47 +0100 | [diff] [blame] | 2721 | } |
| 2722 | else if (operation.inputs.size() == 11) |
| 2723 | { |
| 2724 | // explicit padding |
| 2725 | if (!GetInputScalar<hal_1_2::HalPolicy>(operation, 3, OperandType::INT32, desc.m_PadLeft, model, data) || |
| 2726 | !GetInputScalar<hal_1_2::HalPolicy>(operation, 4, OperandType::INT32, desc.m_PadRight, model, data) || |
| 2727 | !GetInputScalar<hal_1_2::HalPolicy>(operation, 5, OperandType::INT32, desc.m_PadTop, model, data) || |
| 2728 | !GetInputScalar<hal_1_2::HalPolicy>(operation, 6, OperandType::INT32, desc.m_PadBottom, model, data) || |
| 2729 | !GetInputScalar<hal_1_2::HalPolicy>(operation, 7, OperandType::INT32, desc.m_StrideX, model, data) || |
| 2730 | !GetInputScalar<hal_1_2::HalPolicy>(operation, 8, OperandType::INT32, desc.m_StrideY, model, data) || |
| 2731 | !GetInputActivationFunction<hal_1_2::HalPolicy>(operation, 9, activation, model, data)) |
| 2732 | { |
| 2733 | return Fail("%s: Operation has invalid inputs (explicit padding)", __func__); |
| 2734 | } |
| 2735 | } |
| 2736 | else |
| 2737 | { |
| 2738 | return Fail("%s: Unsupported number of operation inputs", __func__); |
| 2739 | } |
| 2740 | |
| 2741 | desc.m_BiasEnabled = true; |
Teresa Charlin | 8f6429d | 2019-10-01 13:10:15 +0100 | [diff] [blame] | 2742 | Optional<TensorInfo> biases(bias.GetInfo()); |
David Monahan | 613b49c | 2019-06-27 11:37:47 +0100 | [diff] [blame] | 2743 | |
| 2744 | bool isSupported = false; |
| 2745 | FORWARD_LAYER_SUPPORT_FUNC(__func__, |
| 2746 | IsTransposeConvolution2dSupported, |
| 2747 | data.m_Backends, |
| 2748 | isSupported, |
| 2749 | inputInfo, |
| 2750 | outputInfo, |
| 2751 | desc, |
| 2752 | weights.GetInfo(), |
| 2753 | biases); |
| 2754 | if (!isSupported) |
| 2755 | { |
| 2756 | return false; |
| 2757 | } |
| 2758 | |
Teresa Charlin | 8f6429d | 2019-10-01 13:10:15 +0100 | [diff] [blame] | 2759 | IConnectableLayer* startLayer = |
| 2760 | data.m_Network->AddTransposeConvolution2dLayer(desc, weights, Optional<ConstTensor>(bias)); |
David Monahan | 613b49c | 2019-06-27 11:37:47 +0100 | [diff] [blame] | 2761 | if (!startLayer) |
| 2762 | { |
| 2763 | return Fail("%s: AddTransposeConvolution2dLayer failed", __func__); |
| 2764 | } |
| 2765 | |
Teresa Charlin | 8f6429d | 2019-10-01 13:10:15 +0100 | [diff] [blame] | 2766 | IConnectableLayer* endLayer = ProcessActivation(outputInfo, activation, startLayer, data); |
David Monahan | 613b49c | 2019-06-27 11:37:47 +0100 | [diff] [blame] | 2767 | if (!endLayer) |
| 2768 | { |
| 2769 | return Fail("%s: ProcessActivation failed", __func__); |
| 2770 | } |
| 2771 | |
| 2772 | input.Connect(startLayer->GetInputSlot(0)); |
| 2773 | |
| 2774 | return SetupAndTrackLayerOutputSlot<hal_1_2::HalPolicy>(operation, 0, *endLayer, model, data); |
| 2775 | } |
| 2776 | |
Mike Kelly | b5fdf38 | 2019-06-11 16:35:25 +0100 | [diff] [blame] | 2777 | } // namespace hal_1_2 |
Matteo Martincigh | 17ffff3 | 2019-06-27 14:12:55 +0100 | [diff] [blame] | 2778 | } // namespace armnn_driver |