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 
17 #define LOG_TAG "Operations"
18 
19 #include "HalInterfaces.h"
20 #include "IndexedShapeWrapper.h"
21 #include "OperationResolver.h"
22 #include "OperationsUtils.h"
23 
24 namespace android {
25 namespace nn {
26 namespace select_op {
27 
28 constexpr uint32_t kNumInputs = 3;
29 constexpr uint32_t kInputCondition = 0;
30 constexpr uint32_t kInputTensor1 = 1;
31 constexpr uint32_t kInputTensor2 = 2;
32 
33 constexpr uint32_t kNumOutputs = 1;
34 constexpr uint32_t kOutputTensor = 0;
35 
36 namespace {
37 
38 template <typename T>
compute(const bool8 * conditionData,const Shape & conditionShape,const T * aData,const Shape & aShape,const T * bData,const Shape & bShape,T * outputData,const Shape & outputShape)39 bool compute(const bool8* conditionData, const Shape& conditionShape, const T* aData,
40              const Shape& aShape, const T* bData, const Shape& bShape, T* outputData,
41              const Shape& outputShape) {
42     // The code assumes that condition has the same shape as all other tensors.
43     // This should be checked during preparation stage.
44     uint32_t size = getNumberOfElements(conditionShape);
45     for (uint32_t i = 0; i < size; ++i) {
46         T a = aData[i];
47         T b = bData[i];
48         if (aShape.type == OperandType::TENSOR_QUANT8_ASYMM) {
49             a = requantize(a, aShape, outputShape);
50             b = requantize(b, bShape, outputShape);
51         }
52         outputData[i] = conditionData[i] ? a : b;
53     }
54     return true;
55 }
56 
57 template <typename T>
executeTyped(IOperationExecutionContext * context)58 bool executeTyped(IOperationExecutionContext* context) {
59     return compute<T>(
60             context->getInputBuffer<bool8>(kInputCondition),
61             context->getInputShape(kInputCondition), context->getInputBuffer<T>(kInputTensor1),
62             context->getInputShape(kInputTensor1), context->getInputBuffer<T>(kInputTensor2),
63             context->getInputShape(kInputTensor2), context->getOutputBuffer<T>(kOutputTensor),
64             context->getOutputShape(kOutputTensor));
65 }
66 
67 }  // namespace
68 
validate(const IOperationValidationContext * context)69 bool validate(const IOperationValidationContext* context) {
70     NN_RET_CHECK_EQ(context->getNumInputs(), kNumInputs);
71     NN_RET_CHECK_EQ(context->getNumOutputs(), kNumOutputs);
72     OperandType inputType = context->getInputType(kInputTensor1);
73     NN_RET_CHECK(
74             inputType == OperandType::TENSOR_FLOAT16 || inputType == OperandType::TENSOR_FLOAT32 ||
75             inputType == OperandType::TENSOR_INT32 || inputType == OperandType::TENSOR_QUANT8_ASYMM)
76             << "Unsupported input operand type for select op: " << toString(inputType);
77     NN_RET_CHECK(validateInputTypes(context, {OperandType::TENSOR_BOOL8, inputType, inputType}));
78     NN_RET_CHECK(validateOutputTypes(context, {inputType}));
79     return validateHalVersion(context, HalVersion::V1_2);
80 }
81 
prepare(IOperationExecutionContext * context)82 bool prepare(IOperationExecutionContext* context) {
83     Shape inputCondition = context->getInputShape(kInputCondition);
84     Shape input1 = context->getInputShape(kInputTensor1);
85     if (inputCondition.dimensions.size() != input1.dimensions.size()) {
86         LOG(ERROR) << "Condition and input tensor dimensions are not equal";
87         return false;
88     }
89     for (int i = 0; i < inputCondition.dimensions.size(); ++i) {
90         if (inputCondition.dimensions[i] != input1.dimensions[i]) {
91             LOG(ERROR) << "Condition and input tensor dimensions are not equal";
92             return false;
93         }
94     }
95 
96     Shape input2 = context->getInputShape(kInputTensor2);
97     NN_RET_CHECK(SameShape(input1, input2));
98 
99     Shape output = context->getOutputShape(kOutputTensor);
100     NN_RET_CHECK(SetShape(input1, &output));
101     return context->setOutputShape(kOutputTensor, output);
102 }
103 
execute(IOperationExecutionContext * context)104 bool execute(IOperationExecutionContext* context) {
105     switch (context->getInputType(kInputTensor1)) {
106         case OperandType::TENSOR_FLOAT16:
107             return executeTyped<_Float16>(context);
108         case OperandType::TENSOR_FLOAT32:
109             return executeTyped<float>(context);
110         case OperandType::TENSOR_INT32:
111             return executeTyped<int32_t>(context);
112         case OperandType::TENSOR_QUANT8_ASYMM:
113             return executeTyped<uint8_t>(context);
114         default:
115             NN_RET_CHECK_FAIL() << "Unsupported tensor type for SELECT op.";
116     }
117 }
118 
119 }  // namespace select_op
120 
121 NN_REGISTER_OPERATION(SELECT, "SELECT", select_op::validate, select_op::prepare,
122                       select_op::execute);
123 
124 }  // namespace nn
125 }  // namespace android
126