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 <algorithm>
20 #include <vector>
21
22 #include "IndexedShapeWrapper.h"
23 #include "OperationResolver.h"
24 #include "OperationsUtils.h"
25 #include "Tracing.h"
26
27 #ifdef NN_INCLUDE_CPU_IMPLEMENTATION
28 #include <tensorflow/lite/kernels/internal/optimized/legacy_optimized_ops.h>
29 #endif // NN_INCLUDE_CPU_IMPLEMENTATION
30
31 namespace android {
32 namespace nn {
33 namespace prelu {
34
35 constexpr char kOperationName[] = "PRELU";
36
37 constexpr uint32_t kNumInputs = 2;
38 constexpr uint32_t kInputTensor = 0;
39 constexpr uint32_t kAlphaTensor = 1;
40
41 constexpr uint32_t kNumOutputs = 1;
42 constexpr uint32_t kOutputTensor = 0;
43
44 #ifdef NN_INCLUDE_CPU_IMPLEMENTATION
45 template <typename T>
eval(const std::function<T (const T &,const T &)> & func,const T * aData,const Shape & aShape,const T * bData,const Shape & bShape,T * outputData,const Shape & outputShape)46 inline bool eval(const std::function<T(const T&, const T&)>& func, const T* aData,
47 const Shape& aShape, const T* bData, const Shape& bShape, T* outputData,
48 const Shape& outputShape) {
49 IndexedShapeWrapper aShapeIndexed(aShape);
50 IndexedShapeWrapper bShapeIndexed(bShape);
51 IndexedShapeWrapper outputShapeIndexed(outputShape);
52 std::vector<uint32_t> curIndex(outputShape.dimensions.size(), 0);
53 bool lastIndex = false;
54 do {
55 uint32_t outputFlatIndex;
56 NN_RET_CHECK(outputShapeIndexed.indexToFlatIndex(curIndex, &outputFlatIndex));
57 uint32_t aFlatIndex;
58 NN_RET_CHECK(aShapeIndexed.broadcastedIndexToFlatIndex(curIndex, &aFlatIndex));
59 uint32_t bFlatIndex;
60 NN_RET_CHECK(bShapeIndexed.broadcastedIndexToFlatIndex(curIndex, &bFlatIndex));
61
62 outputData[outputFlatIndex] = func(aData[aFlatIndex], bData[bFlatIndex]);
63
64 NN_RET_CHECK(outputShapeIndexed.nextIndexInplace(&curIndex, &lastIndex));
65 } while (!lastIndex);
66 return true;
67 }
68
69 template <typename T>
evalQuant8(const T * aData,const Shape & aShape,const T * bData,const Shape & bShape,T * outputData,const Shape & outputShape)70 bool evalQuant8(const T* aData, const Shape& aShape, const T* bData, const Shape& bShape,
71 T* outputData, const Shape& outputShape) {
72 const int32_t input_offset = -aShape.offset;
73 const int32_t alpha_offset = -bShape.offset;
74 const int32_t output_offset = outputShape.offset;
75 const double input_product_scale = aShape.scale * bShape.scale;
76 const double real_multiplier_pos = aShape.scale / outputShape.scale;
77 const double real_multiplier_neg = input_product_scale / outputShape.scale;
78 int32_t output_multiplier_pos, output_shift_pos;
79 int32_t output_multiplier_neg, output_shift_neg;
80 tflite::QuantizeMultiplier(real_multiplier_pos, &output_multiplier_pos, &output_shift_pos);
81 tflite::QuantizeMultiplier(real_multiplier_neg, &output_multiplier_neg, &output_shift_neg);
82 return eval<T>(
83 [&](const T& val1, const T& val2) -> uint8_t {
84 const int32_t input = input_offset + static_cast<int32_t>(val1);
85 int32_t output_val;
86 if (input >= 0) {
87 output_val =
88 output_offset + tflite::MultiplyByQuantizedMultiplier(
89 input, output_multiplier_pos, output_shift_pos);
90 } else {
91 const int32_t alpha = alpha_offset + static_cast<int32_t>(val2);
92 output_val = output_offset +
93 tflite::MultiplyByQuantizedMultiplier(
94 input * alpha, output_multiplier_neg, output_shift_neg);
95 }
96 return saturateCast<T>(output_val);
97 },
98 aData, aShape, bData, bShape, outputData, outputShape);
99 }
100 #endif // NN_INCLUDE_CPU_IMPLEMENTATION
101
validate(const IOperationValidationContext * context)102 Result<Version> validate(const IOperationValidationContext* context) {
103 NN_RET_CHECK_EQ(context->getNumInputs(), kNumInputs);
104 NN_RET_CHECK_EQ(context->getNumOutputs(), kNumOutputs);
105 auto inputType = context->getInputType(kInputTensor);
106 NN_RET_CHECK(inputType == OperandType::TENSOR_FLOAT16 ||
107 inputType == OperandType::TENSOR_FLOAT32 ||
108 inputType == OperandType::TENSOR_QUANT8_ASYMM ||
109 inputType == OperandType::TENSOR_QUANT8_ASYMM_SIGNED)
110 << "Unsupported tensor type for operation " << kOperationName;
111 NN_RET_CHECK(validateInputTypes(context, {inputType, inputType}));
112 NN_RET_CHECK(validateOutputTypes(context, {inputType}));
113 if (inputType == OperandType::TENSOR_QUANT8_ASYMM_SIGNED) {
114 return Version::ANDROID_R;
115 } else {
116 return Version::ANDROID_Q;
117 }
118 }
119
120 #ifdef NN_INCLUDE_CPU_IMPLEMENTATION
prepare(IOperationExecutionContext * context)121 bool prepare(IOperationExecutionContext* context) {
122 Shape input = context->getInputShape(kInputTensor);
123 Shape alpha = context->getInputShape(kAlphaTensor);
124 NN_RET_CHECK(input.type == alpha.type);
125 Shape output = context->getOutputShape(kOutputTensor);
126 NN_RET_CHECK(calculateBroadcastedShape(input, alpha, &output));
127 return context->setOutputShape(kOutputTensor, output);
128 }
129
execute(IOperationExecutionContext * context)130 bool execute(IOperationExecutionContext* context) {
131 switch (context->getInputType(kInputTensor)) {
132 case OperandType::TENSOR_FLOAT16:
133 return eval<_Float16>(
134 [](const _Float16& val1, const _Float16& val2) -> _Float16 {
135 return val1 >= 0.0f ? val1 : val1 * val2;
136 },
137 context->getInputBuffer<_Float16>(kInputTensor),
138 context->getInputShape(kInputTensor),
139 context->getInputBuffer<_Float16>(kAlphaTensor),
140 context->getInputShape(kAlphaTensor),
141 context->getOutputBuffer<_Float16>(kOutputTensor),
142 context->getOutputShape(kOutputTensor));
143 case OperandType::TENSOR_FLOAT32:
144 return eval<float>(
145 [](const float& val1, const float& val2) -> float {
146 return val1 >= 0.0f ? val1 : val1 * val2;
147 },
148 context->getInputBuffer<float>(kInputTensor),
149 context->getInputShape(kInputTensor),
150 context->getInputBuffer<float>(kAlphaTensor),
151 context->getInputShape(kAlphaTensor),
152 context->getOutputBuffer<float>(kOutputTensor),
153 context->getOutputShape(kOutputTensor));
154 case OperandType::TENSOR_QUANT8_ASYMM: {
155 return evalQuant8(context->getInputBuffer<uint8_t>(kInputTensor),
156 context->getInputShape(kInputTensor),
157 context->getInputBuffer<uint8_t>(kAlphaTensor),
158 context->getInputShape(kAlphaTensor),
159 context->getOutputBuffer<uint8_t>(kOutputTensor),
160 context->getOutputShape(kOutputTensor));
161 }
162 case OperandType::TENSOR_QUANT8_ASYMM_SIGNED: {
163 return evalQuant8(context->getInputBuffer<int8_t>(kInputTensor),
164 context->getInputShape(kInputTensor),
165 context->getInputBuffer<int8_t>(kAlphaTensor),
166 context->getInputShape(kAlphaTensor),
167 context->getOutputBuffer<int8_t>(kOutputTensor),
168 context->getOutputShape(kOutputTensor));
169 }
170 default:
171 NN_RET_CHECK_FAIL() << "Unsupported tensor type for operation " << kOperationName;
172 }
173 }
174 #endif // NN_INCLUDE_CPU_IMPLEMENTATION
175
176 } // namespace prelu
177
178 NN_REGISTER_OPERATION(PRELU, prelu::kOperationName, prelu::validate, prelu::prepare,
179 prelu::execute);
180
181 } // namespace nn
182 } // namespace android
183