NeuralNetworksWrapper.h revision 7b87fec6bb919d16a8cc2820d470733a2776e8fa 
1/*
2 * Copyright (C) 2017 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// Provides C++ classes to more easily use the Neural Networks API.
18
19#ifndef ANDROID_ML_NN_RUNTIME_NEURAL_NETWORKS_WRAPPER_H
20#define ANDROID_ML_NN_RUNTIME_NEURAL_NETWORKS_WRAPPER_H
21
22#include "NeuralNetworks.h"
23
24#include <math.h>
25#include <vector>
26
27namespace android {
28namespace nn {
29namespace wrapper {
30
31enum class Type {
32    FLOAT16 = ANEURALNETWORKS_FLOAT16,
33    FLOAT32 = ANEURALNETWORKS_FLOAT32,
34    INT8 = ANEURALNETWORKS_INT8,
35    UINT8 = ANEURALNETWORKS_UINT8,
36    INT16 = ANEURALNETWORKS_INT16,
37    UINT16 = ANEURALNETWORKS_UINT16,
38    INT32 = ANEURALNETWORKS_INT32,
39    UINT32 = ANEURALNETWORKS_UINT32,
40    TENSOR_FLOAT16 = ANEURALNETWORKS_TENSOR_FLOAT16,
41    TENSOR_FLOAT32 = ANEURALNETWORKS_TENSOR_FLOAT32,
42    TENSOR_INT32 = ANEURALNETWORKS_TENSOR_INT32,
43    TENSOR_QUANT8_ASYMM = ANEURALNETWORKS_TENSOR_QUANT8_ASYMM,
44};
45
46enum class ExecutePreference {
47    PREFER_LOW_POWER = ANEURALNETWORKS_PREFER_LOW_POWER,
48    PREFER_FAST_SINGLE_ANSWER = ANEURALNETWORKS_PREFER_FAST_SINGLE_ANSWER,
49    PREFER_SUSTAINED_SPEED = ANEURALNETWORKS_PREFER_SUSTAINED_SPEED
50};
51
52enum class Result {
53    NO_ERROR = ANEURALNETWORKS_NO_ERROR,
54    OUT_OF_MEMORY = ANEURALNETWORKS_OUT_OF_MEMORY,
55    INCOMPLETE = ANEURALNETWORKS_INCOMPLETE,
56    UNEXPECTED_NULL = ANEURALNETWORKS_UNEXPECTED_NULL,
57    BAD_DATA = ANEURALNETWORKS_BAD_DATA,
58};
59
60struct OperandType {
61    ANeuralNetworksOperandType operandType;
62    // uint32_t type;
63    std::vector<uint32_t> dimensions;
64
65    OperandType(Type type, const std::vector<uint32_t>& d) : dimensions(d) {
66        operandType.type = static_cast<uint32_t>(type);
67        operandType.scale = 0.0f;
68        operandType.offset = 0;
69
70        operandType.dimensions.count = static_cast<uint32_t>(dimensions.size());
71        operandType.dimensions.data = dimensions.data();
72    }
73
74    OperandType(Type type, float scale, const std::vector<uint32_t>& d) : OperandType(type, d) {
75        operandType.scale = scale;
76    }
77
78    OperandType(Type type, float f_min, float f_max, const std::vector<uint32_t>& d)
79        : OperandType(type, d) {
80        uint8_t q_min = std::numeric_limits<uint8_t>::min();
81        uint8_t q_max = std::numeric_limits<uint8_t>::max();
82        float range = q_max - q_min;
83        float scale = (f_max - f_min) / range;
84        int32_t offset =
85                    fmin(q_max, fmax(q_min, static_cast<uint8_t>(round(q_min - f_min / scale))));
86
87        operandType.scale = scale;
88        operandType.offset = offset;
89    }
90};
91
92inline Result Initialize() {
93    return static_cast<Result>(ANeuralNetworksInitialize());
94}
95
96inline void Shutdown() {
97    ANeuralNetworksShutdown();
98}
99
100class Memory {
101public:
102    // TODO Also have constructors for file descriptor, gralloc buffers, etc.
103    Memory(size_t size) {
104        mValid = ANeuralNetworksMemory_createShared(size, &mMemory) == ANEURALNETWORKS_NO_ERROR;
105    }
106    ~Memory() { ANeuralNetworksMemory_free(mMemory); }
107    Result getPointer(uint8_t** buffer) {
108        return static_cast<Result>(ANeuralNetworksMemory_getPointer(mMemory, buffer));
109    }
110
111    // Disallow copy semantics to ensure the runtime object can only be freed
112    // once. Copy semantics could be enabled if some sort of reference counting
113    // or deep-copy system for runtime objects is added later.
114    Memory(const Memory&) = delete;
115    Memory& operator=(const Memory&) = delete;
116
117    // Move semantics to remove access to the runtime object from the wrapper
118    // object that is being moved. This ensures the runtime object will be
119    // freed only once.
120    Memory(Memory&& other) {
121        *this = std::move(other);
122    }
123    Memory& operator=(Memory&& other) {
124        if (this != &other) {
125            mMemory = other.mMemory;
126            mValid = other.mValid;
127            other.mMemory = nullptr;
128            other.mValid = false;
129        }
130        return *this;
131    }
132
133    ANeuralNetworksMemory* get() const { return mMemory; }
134    bool isValid() const { return mValid; }
135
136private:
137    ANeuralNetworksMemory* mMemory = nullptr;
138    bool mValid = true;
139};
140
141class Model {
142public:
143    Model() {
144        // TODO handle the value returned by this call
145        ANeuralNetworksModel_create(&mModel);
146    }
147    ~Model() { ANeuralNetworksModel_free(mModel); }
148
149    // Disallow copy semantics to ensure the runtime object can only be freed
150    // once. Copy semantics could be enabled if some sort of reference counting
151    // or deep-copy system for runtime objects is added later.
152    Model(const Model&) = delete;
153    Model& operator=(const Model&) = delete;
154
155    // Move semantics to remove access to the runtime object from the wrapper
156    // object that is being moved. This ensures the runtime object will be
157    // freed only once.
158    Model(Model&& other) {
159        *this = std::move(other);
160    }
161    Model& operator=(Model&& other) {
162        if (this != &other) {
163            mModel = other.mModel;
164            mNextOperandId = other.mNextOperandId;
165            mValid = other.mValid;
166            other.mModel = nullptr;
167            other.mNextOperandId = 0;
168            other.mValid = false;
169        }
170        return *this;
171    }
172
173    uint32_t addOperand(const OperandType* type) {
174        if (ANeuralNetworksModel_addOperand(mModel, &(type->operandType)) !=
175            ANEURALNETWORKS_NO_ERROR) {
176            mValid = false;
177        }
178        return mNextOperandId++;
179    }
180
181    void setOperandValue(uint32_t index, const void* buffer, size_t length) {
182        if (ANeuralNetworksModel_setOperandValue(mModel, index, buffer, length) !=
183            ANEURALNETWORKS_NO_ERROR) {
184            mValid = false;
185        }
186    }
187
188    void setOperandValueFromMemory(uint32_t index, const Memory* memory, uint32_t offset,
189                                   size_t length) {
190        if (ANeuralNetworksModel_setOperandValueFromMemory(mModel, index, memory->get(), offset,
191                                                           length) != ANEURALNETWORKS_NO_ERROR) {
192            mValid = false;
193        }
194    }
195
196    void addOperation(ANeuralNetworksOperationType type, const std::vector<uint32_t>& inputs,
197                      const std::vector<uint32_t>& outputs) {
198        ANeuralNetworksIntList in, out;
199        Set(&in, inputs);
200        Set(&out, outputs);
201        if (ANeuralNetworksModel_addOperation(mModel, type, &in, &out) !=
202            ANEURALNETWORKS_NO_ERROR) {
203            mValid = false;
204        }
205    }
206    void setInputsAndOutputs(const std::vector<uint32_t>& inputs,
207                             const std::vector<uint32_t>& outputs) {
208        ANeuralNetworksIntList in, out;
209        Set(&in, inputs);
210        Set(&out, outputs);
211        if (ANeuralNetworksModel_setInputsAndOutputs(mModel, &in, &out) !=
212            ANEURALNETWORKS_NO_ERROR) {
213            mValid = false;
214        }
215    }
216    ANeuralNetworksModel* getHandle() const { return mModel; }
217    bool isValid() const { return mValid; }
218
219private:
220    /**
221     * WARNING list won't be valid once vec is destroyed or modified.
222     */
223    void Set(ANeuralNetworksIntList* list, const std::vector<uint32_t>& vec) {
224        list->count = static_cast<uint32_t>(vec.size());
225        list->data = vec.data();
226    }
227
228    ANeuralNetworksModel* mModel = nullptr;
229    // We keep track of the operand ID as a convenience to the caller.
230    uint32_t mNextOperandId = 0;
231    bool mValid = true;
232};
233
234class Event {
235public:
236    ~Event() { ANeuralNetworksEvent_free(mEvent); }
237
238    // Disallow copy semantics to ensure the runtime object can only be freed
239    // once. Copy semantics could be enabled if some sort of reference counting
240    // or deep-copy system for runtime objects is added later.
241    Event(const Event&) = delete;
242    Event& operator=(const Event&) = delete;
243
244    // Move semantics to remove access to the runtime object from the wrapper
245    // object that is being moved. This ensures the runtime object will be
246    // freed only once.
247    Event(Event&& other) {
248        *this = std::move(other);
249    }
250    Event& operator=(Event&& other) {
251        if (this != &other) {
252            mEvent = other.mEvent;
253            other.mEvent = nullptr;
254        }
255        return *this;
256    }
257
258    Result wait() { return static_cast<Result>(ANeuralNetworksEvent_wait(mEvent)); }
259    void set(ANeuralNetworksEvent* newEvent) {
260        ANeuralNetworksEvent_free(mEvent);
261        mEvent = newEvent;
262    }
263
264private:
265    ANeuralNetworksEvent* mEvent = nullptr;
266};
267
268class Request {
269public:
270    Request(const Model* model) {
271        int result = ANeuralNetworksRequest_create(model->getHandle(), &mRequest);
272        if (result != 0) {
273            // TODO Handle the error
274        }
275    }
276
277    ~Request() { ANeuralNetworksRequest_free(mRequest); }
278
279    // Disallow copy semantics to ensure the runtime object can only be freed
280    // once. Copy semantics could be enabled if some sort of reference counting
281    // or deep-copy system for runtime objects is added later.
282    Request(const Request&) = delete;
283    Request& operator=(const Request&) = delete;
284
285    // Move semantics to remove access to the runtime object from the wrapper
286    // object that is being moved. This ensures the runtime object will be
287    // freed only once.
288    Request(Request&& other) {
289        *this = std::move(other);
290    }
291    Request& operator=(Request&& other) {
292        if (this != &other) {
293            mRequest = other.mRequest;
294            other.mRequest = nullptr;
295        }
296        return *this;
297    }
298
299    Result setPreference(ExecutePreference preference) {
300        return static_cast<Result>(ANeuralNetworksRequest_setPreference(
301                    mRequest, static_cast<uint32_t>(preference)));
302    }
303
304    Result setInput(uint32_t index, const void* buffer, size_t length,
305                    const ANeuralNetworksOperandType* type = nullptr) {
306        return static_cast<Result>(
307                    ANeuralNetworksRequest_setInput(mRequest, index, type, buffer, length));
308    }
309
310    Result setInputFromMemory(uint32_t index, const Memory* memory, uint32_t offset,
311                              uint32_t length, const ANeuralNetworksOperandType* type = nullptr) {
312        return static_cast<Result>(ANeuralNetworksRequest_setInputFromMemory(
313                    mRequest, index, type, memory->get(), offset, length));
314    }
315
316    Result setOutput(uint32_t index, void* buffer, size_t length,
317                     const ANeuralNetworksOperandType* type = nullptr) {
318        return static_cast<Result>(
319                    ANeuralNetworksRequest_setOutput(mRequest, index, type, buffer, length));
320    }
321
322    Result setOutputFromMemory(uint32_t index, const Memory* memory, uint32_t offset,
323                               uint32_t length, const ANeuralNetworksOperandType* type = nullptr) {
324        return static_cast<Result>(ANeuralNetworksRequest_setOutputFromMemory(
325                    mRequest, index, type, memory->get(), offset, length));
326    }
327
328    Result startCompute(Event* event) {
329        ANeuralNetworksEvent* ev = nullptr;
330        Result result = static_cast<Result>(ANeuralNetworksRequest_startCompute(mRequest, &ev));
331        event->set(ev);
332        return result;
333    }
334
335    Result compute() {
336        ANeuralNetworksEvent* event = nullptr;
337        Result result = static_cast<Result>(ANeuralNetworksRequest_startCompute(mRequest, &event));
338        if (result != Result::NO_ERROR) {
339            return result;
340        }
341        // TODO how to manage the lifetime of events when multiple waiters is not
342        // clear.
343        return static_cast<Result>(ANeuralNetworksEvent_wait(event));
344    }
345
346private:
347    ANeuralNetworksRequest* mRequest = nullptr;
348};
349
350}  // namespace wrapper
351}  // namespace nn
352}  // namespace android
353
354#endif  //  ANDROID_ML_NN_RUNTIME_NEURAL_NETWORKS_WRAPPER_H
355