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