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 "Reduce.h"
20
21 #include <algorithm>
22 #include <limits>
23 #include <vector>
24
25 #include "OperationResolver.h"
26 #include "OperationsExecutionUtils.h"
27 #include "Tracing.h"
28
29 #ifdef NN_INCLUDE_CPU_IMPLEMENTATION
30 #pragma clang diagnostic push
31 #pragma clang diagnostic ignored "-Wunused-parameter"
32 #pragma clang diagnostic ignored "-Wsign-compare"
33 #include <tensorflow/lite/kernels/internal/reference/reference_ops.h>
34 #pragma clang diagnostic pop
35 #endif // NN_INCLUDE_CPU_IMPLEMENTATION
36
37 namespace android {
38 namespace nn {
39 namespace reduce {
40
41 #ifdef NN_INCLUDE_CPU_IMPLEMENTATION
42 namespace {
43
44 template <typename T>
compute(IOperationExecutionContext * context,T init,T func (T,T))45 inline bool compute(IOperationExecutionContext* context, T init, T func(T, T)) {
46 const Shape inputShape = context->getInputShape(kInputTensor);
47 const Shape axesShape = context->getInputShape(kInputAxes);
48 const Shape outputShape = context->getOutputShape(kOutputTensor);
49 const uint32_t inputRank = getNumberOfDimensions(inputShape);
50 const uint32_t numAxes = getNumberOfElements(axesShape);
51 std::vector<int> tempIndex(inputShape.dimensions.size());
52 std::vector<int> tempAxes(numAxes);
53 return tflite::reference_ops::ReduceGeneric<T>(
54 context->getInputBuffer<T>(kInputTensor),
55 reinterpret_cast<const int32_t*>(inputShape.dimensions.data()), inputRank,
56 context->getOutputBuffer<T>(kOutputTensor),
57 reinterpret_cast<const int32_t*>(outputShape.dimensions.data()),
58 outputShape.dimensions.size(), context->getInputBuffer<int32_t>(kInputAxes), numAxes,
59 context->getInputValue<bool8>(kInputKeepDims), tempIndex.data(), tempAxes.data(), init,
60 func);
61 }
62
63 } // namespace
64
prepare(IOperationExecutionContext * context)65 bool prepare(IOperationExecutionContext* context) {
66 Shape inputShape = context->getInputShape(kInputTensor);
67 const uint32_t inputRank = getNumberOfDimensions(inputShape);
68 NN_RET_CHECK_LE(inputRank, 4u);
69
70 std::vector<bool> shouldReduce(inputRank);
71 const int32_t* axes = context->getInputBuffer<int32_t>(kInputAxes);
72 Shape axesShape = context->getInputShape(kInputAxes);
73 NN_RET_CHECK_EQ(getNumberOfDimensions(axesShape), 1u);
74 const uint32_t numAxes = getNumberOfElements(axesShape);
75 for (uint32_t i = 0; i < numAxes; ++i) {
76 int32_t axis = axes[i];
77 NN_RET_CHECK(handleNegativeAxis(inputRank, &axis));
78 shouldReduce[axis] = true;
79 }
80
81 // Input and output must have the same quantization parameters, etc.
82 Shape outputShape = inputShape;
83 outputShape.dimensions.clear();
84 bool keepDims = context->getInputValue<bool8>(kInputKeepDims);
85 for (uint32_t axis = 0; axis < inputRank; ++axis) {
86 if (shouldReduce[axis]) {
87 if (keepDims) {
88 outputShape.dimensions.push_back(1);
89 }
90 } else {
91 outputShape.dimensions.push_back(getSizeOfDimension(inputShape, axis));
92 }
93 }
94
95 // Handle the case when all dimensions are removed
96 if (outputShape.dimensions.empty()) {
97 outputShape.dimensions.push_back(1);
98 }
99
100 return context->setOutputShape(kOutputTensor, outputShape);
101 }
102
executeProd(IOperationExecutionContext * context)103 bool executeProd(IOperationExecutionContext* context) {
104 switch (context->getInputType(kInputTensor)) {
105 case OperandType::TENSOR_FLOAT16:
106 return compute<_Float16>(context, 1, [](_Float16 a, _Float16 b) -> _Float16 {
107 // Handle the zero case because 0 * inf evaluates to nan.
108 if (a == 0 || b == 0) return 0;
109 return a * b;
110 });
111 case OperandType::TENSOR_FLOAT32:
112 return compute<float>(context, 1, [](float a, float b) -> float {
113 // Handle the zero case because 0 * inf evaluates to nan.
114 if (a == 0 || b == 0) return 0;
115 return a * b;
116 });
117 default:
118 NN_RET_CHECK_FAIL() << "Unsupported tensor type for operation REDUCE_PROD";
119 }
120 }
121
executeSum(IOperationExecutionContext * context)122 bool executeSum(IOperationExecutionContext* context) {
123 switch (context->getInputType(kInputTensor)) {
124 case OperandType::TENSOR_FLOAT16:
125 return compute<_Float16>(context, 0, [](_Float16 a, _Float16 b) { return a + b; });
126 case OperandType::TENSOR_FLOAT32:
127 return compute<float>(context, 0, [](float a, float b) { return a + b; });
128 default:
129 NN_RET_CHECK_FAIL() << "Unsupported tensor type for operation REDUCE_SUM";
130 }
131 }
132
executeMax(IOperationExecutionContext * context)133 bool executeMax(IOperationExecutionContext* context) {
134 switch (context->getInputType(kInputTensor)) {
135 case OperandType::TENSOR_FLOAT16:
136 return compute<_Float16>(context, kFloat16Lowest,
137 [](_Float16 a, _Float16 b) { return std::max(a, b); });
138 case OperandType::TENSOR_FLOAT32:
139 return compute<float>(context, std::numeric_limits<float>::lowest(),
140 [](float a, float b) { return std::max(a, b); });
141 case OperandType::TENSOR_QUANT8_ASYMM:
142 return compute<uint8_t>(context, std::numeric_limits<uint8_t>::lowest(),
143 [](uint8_t a, uint8_t b) { return std::max(a, b); });
144 case OperandType::TENSOR_QUANT8_ASYMM_SIGNED:
145 return compute<int8_t>(context, std::numeric_limits<int8_t>::lowest(),
146 [](int8_t a, int8_t b) { return std::max(a, b); });
147 default:
148 NN_RET_CHECK_FAIL() << "Unsupported tensor type for operation REDUCE_MAX";
149 }
150 }
151
executeMin(IOperationExecutionContext * context)152 bool executeMin(IOperationExecutionContext* context) {
153 switch (context->getInputType(kInputTensor)) {
154 case OperandType::TENSOR_FLOAT16:
155 return compute<_Float16>(context, kFloat16Max,
156 [](_Float16 a, _Float16 b) { return std::min(a, b); });
157 case OperandType::TENSOR_FLOAT32:
158 return compute<float>(context, std::numeric_limits<float>::max(),
159 [](float a, float b) { return std::min(a, b); });
160 case OperandType::TENSOR_QUANT8_ASYMM:
161 return compute<uint8_t>(context, std::numeric_limits<uint8_t>::max(),
162 [](uint8_t a, uint8_t b) { return std::min(a, b); });
163 case OperandType::TENSOR_QUANT8_ASYMM_SIGNED:
164 return compute<int8_t>(context, std::numeric_limits<int8_t>::max(),
165 [](int8_t a, int8_t b) { return std::min(a, b); });
166 default:
167 NN_RET_CHECK_FAIL() << "Unsupported tensor type for operation REDUCE_MIN";
168 }
169 }
170
executeAny(IOperationExecutionContext * context)171 bool executeAny(IOperationExecutionContext* context) {
172 switch (context->getInputType(kInputTensor)) {
173 case OperandType::TENSOR_BOOL8:
174 return compute<bool8>(context, false,
175 [](bool8 a, bool8 b) { return static_cast<bool8>(a || b); });
176 default:
177 NN_RET_CHECK_FAIL() << "Unsupported tensor type for operation REDUCE_ANY";
178 }
179 }
180
executeAll(IOperationExecutionContext * context)181 bool executeAll(IOperationExecutionContext* context) {
182 switch (context->getInputType(kInputTensor)) {
183 case OperandType::TENSOR_BOOL8:
184 return compute<bool8>(context, true,
185 [](bool8 a, bool8 b) { return static_cast<bool8>(a && b); });
186 default:
187 NN_RET_CHECK_FAIL() << "Unsupported tensor type for operation REDUCE_ALL";
188 }
189 }
190 #endif // NN_INCLUDE_CPU_IMPLEMENTATION
191
192 } // namespace reduce
193
194 NN_REGISTER_OPERATION_DEFAULT_VALIDATION(REDUCE_PROD, reduce::prepare, reduce::executeProd);
195 NN_REGISTER_OPERATION_DEFAULT_VALIDATION(REDUCE_SUM, reduce::prepare, reduce::executeSum);
196 NN_REGISTER_OPERATION_DEFAULT_VALIDATION(REDUCE_MAX, reduce::prepare, reduce::executeMax);
197 NN_REGISTER_OPERATION_DEFAULT_VALIDATION(REDUCE_MIN, reduce::prepare, reduce::executeMin);
198 NN_REGISTER_OPERATION_DEFAULT_VALIDATION(REDUCE_ANY, reduce::prepare, reduce::executeAny);
199 NN_REGISTER_OPERATION_DEFAULT_VALIDATION(REDUCE_ALL, reduce::prepare, reduce::executeAll);
200
201 } // namespace nn
202 } // namespace android
203