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