1 /*
2 * Copyright (C) 2018 The Android Open Source Project
3 *
4 * Licensed under the Apache License, Version 2.0 (the "License");
5 * you may not use this file except in compliance with the License.
6 * You may obtain a copy of the License at
7 *
8 * http://www.apache.org/licenses/LICENSE-2.0
9 *
10 * Unless required by applicable law or agreed to in writing, software
11 * distributed under the License is distributed on an "AS IS" BASIS,
12 * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
13 * See the License for the specific language governing permissions and
14 * limitations under the License.
15 */
16 #include "CpuOperationUtils.h"
17 #include "IndexedShapeWrapper.h"
18 #include "OperationResolver.h"
19
20 #include <vector>
21
22 namespace android {
23 namespace nn {
24 namespace slice {
25
26 constexpr char kOperationName[] = "SLICE";
27
28 constexpr uint32_t kNumInputs = 3;
29 constexpr uint32_t kInputTensor = 0;
30 constexpr uint32_t kBeginTensor = 1;
31 constexpr uint32_t kSizeTensor = 2;
32
33 constexpr uint32_t kNumOutputs = 1;
34 constexpr uint32_t kOutputTensor = 0;
35
36 namespace {
37
38 template <typename T>
addVectors(const std::vector<T> & a,const std::vector<T> & b,std::vector<T> * res)39 void addVectors(const std::vector<T>& a, const std::vector<T>& b, std::vector<T>* res) {
40 for (int i = 0; i < res->size(); ++i) {
41 res->at(i) = a[i] + b[i];
42 }
43 }
44
45 template <typename T>
evalGeneric(const T * inputData,const Shape & inputShape,const int32_t * beginData,const Shape & beginShape,const int32_t * sizeData,const Shape & sizeShape,T * outputData,const Shape & outputShape)46 bool evalGeneric(const T* inputData, const Shape& inputShape, const int32_t* beginData,
47 const Shape& beginShape, const int32_t* sizeData, const Shape& sizeShape,
48 T* outputData, const Shape& outputShape) {
49 const int outputSize = getNumberOfElements(outputShape);
50 const IndexedShapeWrapper indexedOutput = IndexedShapeWrapper(outputShape);
51 const IndexedShapeWrapper indexedInput = IndexedShapeWrapper(inputShape);
52 std::vector<uint32_t> outputIndex(getNumberOfDimensions(outputShape), 0);
53 std::vector<uint32_t> beginIndex(getSizeOfDimension(beginShape, 0));
54 std::vector<uint32_t> inputIndex(getNumberOfDimensions(inputShape));
55
56 for (int i = 0; i < beginIndex.size(); ++i) {
57 beginIndex[i] = static_cast<uint32_t>(beginData[i]);
58 }
59
60 bool lastIndex = false;
61 uint32_t outputOffset;
62 uint32_t inputOffset;
63
64 do {
65 addVectors(outputIndex, beginIndex, &inputIndex);
66
67 NN_RET_CHECK(indexedOutput.indexToFlatIndex(outputIndex, &outputOffset));
68 NN_RET_CHECK(indexedInput.indexToFlatIndex(inputIndex, &inputOffset));
69
70 outputData[outputOffset] = inputData[inputOffset];
71 NN_RET_CHECK(indexedOutput.nextIndexInplace(&outputIndex, &lastIndex));
72 } while (!lastIndex);
73 return true;
74 }
75
76 } // namespace
77
validate(const IOperationValidationContext * context)78 bool validate(const IOperationValidationContext* context) {
79 NN_RET_CHECK_EQ(context->getNumInputs(), kNumInputs);
80 NN_RET_CHECK_EQ(context->getNumOutputs(), kNumOutputs);
81
82 const OperandType inputType = context->getInputType(kInputTensor);
83 NN_RET_CHECK(
84 inputType == OperandType::TENSOR_FLOAT16 || inputType == OperandType::TENSOR_FLOAT32 ||
85 inputType == OperandType::TENSOR_INT32 || inputType == OperandType::TENSOR_QUANT8_ASYMM)
86 << "Unsupported tensor type for operation " << kOperationName;
87 NN_RET_CHECK(validateHalVersion(context, HalVersion::V1_2));
88 return validateInputTypes(context,
89 {inputType, OperandType::TENSOR_INT32, OperandType::TENSOR_INT32}) &&
90 validateOutputTypes(context, {inputType});
91 }
92
prepare(IOperationExecutionContext * context)93 bool prepare(IOperationExecutionContext* context) {
94 const Shape& inputShape = context->getInputShape(kInputTensor);
95 const int32_t n_dims = getNumberOfDimensions(inputShape);
96 NN_RET_CHECK(n_dims > 0);
97
98 const Shape& beginShape = context->getInputShape(kBeginTensor);
99 NN_RET_CHECK_EQ(getNumberOfDimensions(beginShape), 1);
100 NN_RET_CHECK_EQ(getSizeOfDimension(beginShape, 0), n_dims);
101
102 const Shape& sizeShape = context->getInputShape(kSizeTensor);
103 NN_RET_CHECK_EQ(getNumberOfDimensions(sizeShape), 1);
104 NN_RET_CHECK_EQ(getSizeOfDimension(sizeShape, 0), n_dims);
105
106 const int32_t* beginData = context->getInputBuffer<int32_t>(kBeginTensor);
107 const int32_t* sizeData = context->getInputBuffer<int32_t>(kSizeTensor);
108
109 Shape outputShape = context->getOutputShape(kOutputTensor);
110 outputShape.dimensions.resize(n_dims);
111 for (int i = 0; i < n_dims; ++i) {
112 const int32_t sliceBegin = beginData[i];
113 int32_t sliceSize = sizeData[i];
114 if (sliceSize == -1) {
115 sliceSize = getSizeOfDimension(inputShape, i) - sliceBegin;
116 }
117 NN_RET_CHECK_LE(beginData[i], getSizeOfDimension(inputShape, i));
118 NN_RET_CHECK_GE(sliceSize, 0);
119 NN_RET_CHECK_LE(sliceBegin + sliceSize, getSizeOfDimension(inputShape, i));
120 outputShape.dimensions[i] = sliceSize;
121 }
122 return context->setOutputShape(kOutputTensor, outputShape);
123 }
124
execute(IOperationExecutionContext * context)125 bool execute(IOperationExecutionContext* context) {
126 // Bypass execution in the case of zero-sized input.
127 if (getNumberOfElements(context->getOutputShape(kOutputTensor)) == 0) return true;
128 switch (context->getInputType(kInputTensor)) {
129 case OperandType::TENSOR_FLOAT16:
130 return evalGeneric(context->getInputBuffer<_Float16>(kInputTensor),
131 context->getInputShape(kInputTensor),
132 context->getInputBuffer<int32_t>(kBeginTensor),
133 context->getInputShape(kBeginTensor),
134 context->getInputBuffer<int32_t>(kSizeTensor),
135 context->getInputShape(kSizeTensor),
136 context->getOutputBuffer<_Float16>(kOutputTensor),
137 context->getOutputShape(kOutputTensor));
138 case OperandType::TENSOR_FLOAT32:
139 return evalGeneric(context->getInputBuffer<float>(kInputTensor),
140 context->getInputShape(kInputTensor),
141 context->getInputBuffer<int32_t>(kBeginTensor),
142 context->getInputShape(kBeginTensor),
143 context->getInputBuffer<int32_t>(kSizeTensor),
144 context->getInputShape(kSizeTensor),
145 context->getOutputBuffer<float>(kOutputTensor),
146 context->getOutputShape(kOutputTensor));
147 case OperandType::TENSOR_INT32:
148 return evalGeneric(context->getInputBuffer<int32_t>(kInputTensor),
149 context->getInputShape(kInputTensor),
150 context->getInputBuffer<int32_t>(kBeginTensor),
151 context->getInputShape(kBeginTensor),
152 context->getInputBuffer<int32_t>(kSizeTensor),
153 context->getInputShape(kSizeTensor),
154 context->getOutputBuffer<int32_t>(kOutputTensor),
155 context->getOutputShape(kOutputTensor));
156 case OperandType::TENSOR_QUANT8_ASYMM:
157 return evalGeneric(context->getInputBuffer<uint8_t>(kInputTensor),
158 context->getInputShape(kInputTensor),
159 context->getInputBuffer<int32_t>(kBeginTensor),
160 context->getInputShape(kBeginTensor),
161 context->getInputBuffer<int32_t>(kSizeTensor),
162 context->getInputShape(kSizeTensor),
163 context->getOutputBuffer<uint8_t>(kOutputTensor),
164 context->getOutputShape(kOutputTensor));
165 default:
166 NN_RET_CHECK_FAIL() << "Unsupported tensor type for operation " << kOperationName;
167 }
168 }
169
170 } // namespace slice
171
172 NN_REGISTER_OPERATION(SLICE, slice::kOperationName, slice::validate, slice::prepare, slice::execute,
173 .allowZeroSizedInput = true);
174
175 } // namespace nn
176 } // namespace android
177