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 <tensorflow/lite/kernels/internal/reference/reference_ops.h>
20
21 #include <algorithm>
22 #include <limits>
23 #include <vector>
24
25 #include "HalInterfaces.h"
26 #include "OperationResolver.h"
27 #include "OperationsUtils.h"
28 #include "Tracing.h"
29
30 namespace android {
31 namespace nn {
32 namespace reduce {
33
34 constexpr uint32_t kNumInputs = 3;
35 constexpr uint32_t kInputTensor = 0;
36 constexpr uint32_t kInputAxes = 1;
37 constexpr uint32_t kInputKeepDims = 2;
38
39 constexpr uint32_t kNumOutputs = 1;
40 constexpr uint32_t kOutputTensor = 0;
41
42 // Values from
43 // https://en.wikipedia.org/wiki/Half-precision_floating-point_format#IEEE_754_half-precision_binary_floating-point_format:_binary16
44 constexpr _Float16 kFloat16Max = 65504;
45 constexpr _Float16 kFloat16Lowest = -kFloat16Max;
46
47 namespace {
48
49 using namespace hal;
50
51 template <typename T>
compute(IOperationExecutionContext * context,T init,T func (T,T))52 inline bool compute(IOperationExecutionContext* context, T init, T func(T, T)) {
53 const Shape inputShape = context->getInputShape(kInputTensor);
54 const Shape axesShape = context->getInputShape(kInputAxes);
55 const Shape outputShape = context->getOutputShape(kOutputTensor);
56 const uint32_t inputRank = getNumberOfDimensions(inputShape);
57 const uint32_t numAxes = getNumberOfElements(axesShape);
58 std::vector<int> tempIndex(inputShape.dimensions.size());
59 std::vector<int> tempAxes(numAxes);
60 return tflite::reference_ops::ReduceGeneric<T>(
61 context->getInputBuffer<T>(kInputTensor),
62 reinterpret_cast<const int32_t*>(inputShape.dimensions.data()), inputRank,
63 context->getOutputBuffer<T>(kOutputTensor),
64 reinterpret_cast<const int32_t*>(outputShape.dimensions.data()),
65 outputShape.dimensions.size(), context->getInputBuffer<int32_t>(kInputAxes), numAxes,
66 context->getInputValue<bool8>(kInputKeepDims), tempIndex.data(), tempAxes.data(), init,
67 func);
68 }
69
70 } // namespace
71
validateProdSum(const IOperationValidationContext * context)72 bool validateProdSum(const IOperationValidationContext* context) {
73 NN_RET_CHECK_EQ(context->getNumInputs(), kNumInputs);
74 NN_RET_CHECK_EQ(context->getNumOutputs(), kNumOutputs);
75 OperandType inputType = context->getInputType(kInputTensor);
76 NN_RET_CHECK(inputType == OperandType::TENSOR_FLOAT16 ||
77 inputType == OperandType::TENSOR_FLOAT32)
78 << "Unsupported tensor type for REDUCE_PROD or REDUCE_SUM";
79 NN_RET_CHECK(
80 validateInputTypes(context, {inputType, OperandType::TENSOR_INT32, OperandType::BOOL}));
81 NN_RET_CHECK(validateOutputTypes(context, {inputType}));
82 const Shape& input = context->getInputShape(kInputTensor);
83 if (hasKnownRank(input)) {
84 NN_RET_CHECK_LE(getNumberOfDimensions(input), 4);
85 }
86 return validateHalVersion(context, HalVersion::V1_2);
87 }
88
validateMaxMin(const IOperationValidationContext * context)89 bool validateMaxMin(const IOperationValidationContext* context) {
90 NN_RET_CHECK_EQ(context->getNumInputs(), kNumInputs);
91 NN_RET_CHECK_EQ(context->getNumOutputs(), kNumOutputs);
92 OperandType inputType = context->getInputType(kInputTensor);
93 NN_RET_CHECK(inputType == OperandType::TENSOR_FLOAT16 ||
94 inputType == OperandType::TENSOR_FLOAT32 ||
95 inputType == OperandType::TENSOR_QUANT8_ASYMM ||
96 inputType == OperandType::TENSOR_QUANT8_ASYMM_SIGNED)
97 << "Unsupported tensor type for REDUCE_MAX or REDUCE_MIN";
98 NN_RET_CHECK(
99 validateInputTypes(context, {inputType, OperandType::TENSOR_INT32, OperandType::BOOL}));
100 NN_RET_CHECK(validateOutputTypes(context, {inputType}));
101 auto minHalVersion = HalVersion::V1_2;
102 if (inputType == OperandType::TENSOR_QUANT8_ASYMM_SIGNED) {
103 minHalVersion = HalVersion::V1_3;
104 }
105 const Shape& input = context->getInputShape(kInputTensor);
106 if (hasKnownRank(input)) {
107 NN_RET_CHECK_LE(getNumberOfDimensions(input), 4);
108 }
109 return validateHalVersion(context, minHalVersion);
110 }
111
validateLogical(const IOperationValidationContext * context)112 bool validateLogical(const IOperationValidationContext* context) {
113 NN_RET_CHECK_EQ(context->getNumInputs(), kNumInputs);
114 NN_RET_CHECK_EQ(context->getNumOutputs(), kNumOutputs);
115 OperandType inputType = context->getInputType(kInputTensor);
116 NN_RET_CHECK(inputType == OperandType::TENSOR_BOOL8)
117 << "Unsupported tensor type for REDUCE_ANY or REDUCE_ALL";
118 NN_RET_CHECK(
119 validateInputTypes(context, {inputType, OperandType::TENSOR_INT32, OperandType::BOOL}));
120 NN_RET_CHECK(validateOutputTypes(context, {inputType}));
121 const Shape& input = context->getInputShape(kInputTensor);
122 if (hasKnownRank(input)) {
123 NN_RET_CHECK_LE(getNumberOfDimensions(input), 4);
124 }
125 return validateHalVersion(context, HalVersion::V1_2);
126 }
127
prepare(IOperationExecutionContext * context)128 bool prepare(IOperationExecutionContext* context) {
129 Shape inputShape = context->getInputShape(kInputTensor);
130 const uint32_t inputRank = getNumberOfDimensions(inputShape);
131 NN_RET_CHECK_LE(inputRank, 4);
132
133 std::vector<bool> shouldReduce(inputRank);
134 const int32_t* axes = context->getInputBuffer<int32_t>(kInputAxes);
135 Shape axesShape = context->getInputShape(kInputAxes);
136 NN_RET_CHECK_EQ(getNumberOfDimensions(axesShape), 1u);
137 const uint32_t numAxes = getNumberOfElements(axesShape);
138 for (uint32_t i = 0; i < numAxes; ++i) {
139 int32_t axis = axes[i];
140 NN_RET_CHECK(handleNegativeAxis(inputRank, &axis));
141 shouldReduce[axis] = true;
142 }
143
144 // Input and output must have the same quantization parameters, etc.
145 Shape outputShape = inputShape;
146 outputShape.dimensions.clear();
147 bool keepDims = context->getInputValue<bool8>(kInputKeepDims);
148 for (uint32_t axis = 0; axis < inputRank; ++axis) {
149 if (shouldReduce[axis]) {
150 if (keepDims) {
151 outputShape.dimensions.push_back(1);
152 }
153 } else {
154 outputShape.dimensions.push_back(getSizeOfDimension(inputShape, axis));
155 }
156 }
157
158 // Handle the case when all dimensions are removed
159 if (outputShape.dimensions.empty()) {
160 outputShape.dimensions.push_back(1);
161 }
162
163 return context->setOutputShape(kOutputTensor, outputShape);
164 }
165
executeProd(IOperationExecutionContext * context)166 bool executeProd(IOperationExecutionContext* context) {
167 switch (context->getInputType(kInputTensor)) {
168 case OperandType::TENSOR_FLOAT16:
169 return compute<_Float16>(context, 1, [](_Float16 a, _Float16 b) -> _Float16 {
170 // Handle the zero case because 0 * inf evaluates to nan.
171 if (a == 0 || b == 0) return 0;
172 return a * b;
173 });
174 case OperandType::TENSOR_FLOAT32:
175 return compute<float>(context, 1, [](float a, float b) -> float {
176 // Handle the zero case because 0 * inf evaluates to nan.
177 if (a == 0 || b == 0) return 0;
178 return a * b;
179 });
180 default:
181 NN_RET_CHECK_FAIL() << "Unsupported tensor type for operation REDUCE_PROD";
182 }
183 }
184
executeSum(IOperationExecutionContext * context)185 bool executeSum(IOperationExecutionContext* context) {
186 switch (context->getInputType(kInputTensor)) {
187 case OperandType::TENSOR_FLOAT16:
188 return compute<_Float16>(context, 0, [](_Float16 a, _Float16 b) { return a + b; });
189 case OperandType::TENSOR_FLOAT32:
190 return compute<float>(context, 0, [](float a, float b) { return a + b; });
191 default:
192 NN_RET_CHECK_FAIL() << "Unsupported tensor type for operation REDUCE_SUM";
193 }
194 }
195
executeMax(IOperationExecutionContext * context)196 bool executeMax(IOperationExecutionContext* context) {
197 switch (context->getInputType(kInputTensor)) {
198 case OperandType::TENSOR_FLOAT16:
199 return compute<_Float16>(context, kFloat16Lowest,
200 [](_Float16 a, _Float16 b) { return std::max(a, b); });
201 case OperandType::TENSOR_FLOAT32:
202 return compute<float>(context, std::numeric_limits<float>::lowest(),
203 [](float a, float b) { return std::max(a, b); });
204 case OperandType::TENSOR_QUANT8_ASYMM:
205 return compute<uint8_t>(context, std::numeric_limits<uint8_t>::lowest(),
206 [](uint8_t a, uint8_t b) { return std::max(a, b); });
207 case OperandType::TENSOR_QUANT8_ASYMM_SIGNED:
208 return compute<int8_t>(context, std::numeric_limits<int8_t>::lowest(),
209 [](int8_t a, int8_t b) { return std::max(a, b); });
210 default:
211 NN_RET_CHECK_FAIL() << "Unsupported tensor type for operation REDUCE_MAX";
212 }
213 }
214
executeMin(IOperationExecutionContext * context)215 bool executeMin(IOperationExecutionContext* context) {
216 switch (context->getInputType(kInputTensor)) {
217 case OperandType::TENSOR_FLOAT16:
218 return compute<_Float16>(context, kFloat16Max,
219 [](_Float16 a, _Float16 b) { return std::min(a, b); });
220 case OperandType::TENSOR_FLOAT32:
221 return compute<float>(context, std::numeric_limits<float>::max(),
222 [](float a, float b) { return std::min(a, b); });
223 case OperandType::TENSOR_QUANT8_ASYMM:
224 return compute<uint8_t>(context, std::numeric_limits<uint8_t>::max(),
225 [](uint8_t a, uint8_t b) { return std::min(a, b); });
226 case OperandType::TENSOR_QUANT8_ASYMM_SIGNED:
227 return compute<int8_t>(context, std::numeric_limits<int8_t>::max(),
228 [](int8_t a, int8_t b) { return std::min(a, b); });
229 default:
230 NN_RET_CHECK_FAIL() << "Unsupported tensor type for operation REDUCE_MIN";
231 }
232 }
233
executeAny(IOperationExecutionContext * context)234 bool executeAny(IOperationExecutionContext* context) {
235 switch (context->getInputType(kInputTensor)) {
236 case OperandType::TENSOR_BOOL8:
237 return compute<bool8>(context, false,
238 [](bool8 a, bool8 b) { return static_cast<bool8>(a || b); });
239 default:
240 NN_RET_CHECK_FAIL() << "Unsupported tensor type for operation REDUCE_ANY";
241 }
242 }
243
executeAll(IOperationExecutionContext * context)244 bool executeAll(IOperationExecutionContext* context) {
245 switch (context->getInputType(kInputTensor)) {
246 case OperandType::TENSOR_BOOL8:
247 return compute<bool8>(context, true,
248 [](bool8 a, bool8 b) { return static_cast<bool8>(a && b); });
249 default:
250 NN_RET_CHECK_FAIL() << "Unsupported tensor type for operation REDUCE_ALL";
251 }
252 }
253
254 } // namespace reduce
255
256 NN_REGISTER_OPERATION(REDUCE_PROD, "REDUCE_PROD", reduce::validateProdSum, reduce::prepare,
257 reduce::executeProd);
258 NN_REGISTER_OPERATION(REDUCE_SUM, "REDUCE_SUM", reduce::validateProdSum, reduce::prepare,
259 reduce::executeSum);
260 NN_REGISTER_OPERATION(REDUCE_MAX, "REDUCE_MAX", reduce::validateMaxMin, reduce::prepare,
261 reduce::executeMax);
262 NN_REGISTER_OPERATION(REDUCE_MIN, "REDUCE_MIN", reduce::validateMaxMin, reduce::prepare,
263 reduce::executeMin);
264 NN_REGISTER_OPERATION(REDUCE_ANY, "REDUCE_ANY", reduce::validateLogical, reduce::prepare,
265 reduce::executeAny);
266 NN_REGISTER_OPERATION(REDUCE_ALL, "REDUCE_ALL", reduce::validateLogical, reduce::prepare,
267 reduce::executeAll);
268
269 } // namespace nn
270 } // namespace android
271