1 /*
2 * Copyright (C) 2019 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 #include "OperationsUtils.h"
18 #define LOG_TAG "Operations"
19
20 #include "HalInterfaces.h"
21 #include "IndexedShapeWrapper.h"
22 #include "OperationResolver.h"
23
24 namespace android {
25 namespace nn {
26 namespace dequantize {
27
28 constexpr uint32_t kNumInputs = 1;
29 constexpr uint32_t kInputTensor = 0;
30
31 constexpr uint32_t kNumOutputs = 1;
32 constexpr uint32_t kOutputTensor = 0;
33
34 namespace {
35
36 using namespace hal;
37
38 template <typename InputType, typename OutputType>
compute(const InputType * inputData,const Shape & inputShape,OutputType * outputData)39 bool compute(const InputType* inputData, const Shape& inputShape, OutputType* outputData) {
40 const int numElements = getNumberOfElements(inputShape);
41 const int32_t zeroPoint = inputShape.offset;
42 const float scale = inputShape.scale;
43 for (int i = 0; i < numElements; ++i) {
44 const int32_t value = inputData[i];
45 outputData[i] = static_cast<OutputType>(scale * (value - zeroPoint));
46 }
47 return true;
48 }
49
50 template <typename OutputType>
computePerChannel(const int8_t * inputData,const Shape & inputShape,OutputType * outputData)51 bool computePerChannel(const int8_t* inputData, const Shape& inputShape, OutputType* outputData) {
52 // First we calculate a stride which is the number of elements we need to
53 // skip to change an index along a dimension with different quantization
54 // scales.
55 const int channelDim = inputShape.extraParams.channelQuant().channelDim;
56 int stride = 1;
57 for (int i = getNumberOfDimensions(inputShape) - 1; i > channelDim; --i) {
58 stride *= getSizeOfDimension(inputShape, i);
59 }
60
61 const int numElements = getNumberOfElements(inputShape);
62 const int32_t zeroPoint = inputShape.offset;
63
64 for (int i = 0; i < numElements; ++i) {
65 // To get current index along the quantized dimension we calculate how
66 // many even |strides| we looped through and take this number modulo the
67 // size of the dimension (so that we don't have an overflow if the
68 // channelDim is not 0).
69 const int scaleIndex = (i / stride) % getSizeOfDimension(inputShape, channelDim);
70 const float scale = inputShape.extraParams.channelQuant().scales[scaleIndex];
71 const int32_t value = inputData[i];
72 outputData[i] = static_cast<OutputType>(scale * (value - zeroPoint));
73 }
74 return true;
75 }
76
77 } // namespace
78
validate(const IOperationValidationContext * context)79 bool validate(const IOperationValidationContext* context) {
80 NN_RET_CHECK_EQ(context->getNumInputs(), kNumInputs);
81 NN_RET_CHECK_EQ(context->getNumOutputs(), kNumOutputs);
82
83 const OperandType inputType = context->getInputType(kInputTensor);
84 const OperandType outputType = context->getOutputType(kOutputTensor);
85
86 const Shape& input = context->getInputShape(kInputTensor);
87 if (hasKnownRank(input)) {
88 NN_RET_CHECK_LE(getNumberOfDimensions(input), 4);
89 }
90
91 if (inputType == OperandType::TENSOR_QUANT8_ASYMM &&
92 outputType == OperandType::TENSOR_FLOAT32) {
93 return validateHalVersion(context, HalVersion::V1_0);
94 }
95
96 NN_RET_CHECK(inputType == OperandType::TENSOR_QUANT8_ASYMM ||
97 inputType == OperandType::TENSOR_QUANT8_ASYMM_SIGNED ||
98 inputType == OperandType::TENSOR_QUANT8_SYMM ||
99 inputType == OperandType::TENSOR_QUANT8_SYMM_PER_CHANNEL)
100 << "Unsupported input operand type for DEQUANTIZE op: " << toString(inputType);
101 NN_RET_CHECK(outputType == OperandType::TENSOR_FLOAT16 ||
102 outputType == OperandType::TENSOR_FLOAT32)
103 << "Unsupported output operand type for DEQUANTIZE op: " << toString(outputType);
104 return validateHalVersion(context, HalVersion::V1_2);
105 }
106
prepare(IOperationExecutionContext * context)107 bool prepare(IOperationExecutionContext* context) {
108 const Shape& input = context->getInputShape(kInputTensor);
109 NN_RET_CHECK_LE(getNumberOfDimensions(input), 4);
110 Shape output = context->getOutputShape(kOutputTensor);
111 output.dimensions = input.dimensions;
112 return context->setOutputShape(kOutputTensor, output);
113 }
114
execute(IOperationExecutionContext * context)115 bool execute(IOperationExecutionContext* context) {
116 // Bypass execution in the case of zero-sized input.
117 if (getNumberOfElements(context->getOutputShape(kOutputTensor)) == 0) return true;
118
119 const OperandType inputType = context->getInputType(kInputTensor);
120 const OperandType outputType = context->getOutputType(kOutputTensor);
121
122 const Shape& inputShape = context->getInputShape(kInputTensor);
123 if (inputType == OperandType::TENSOR_QUANT8_ASYMM) {
124 const uint8_t* inputBuffer = context->getInputBuffer<uint8_t>(kInputTensor);
125 if (outputType == OperandType::TENSOR_FLOAT16) {
126 return compute(inputBuffer, inputShape,
127 context->getOutputBuffer<_Float16>(kOutputTensor));
128 } else if (outputType == OperandType::TENSOR_FLOAT32) {
129 return compute(inputBuffer, inputShape, context->getOutputBuffer<float>(kOutputTensor));
130 }
131 } else if (inputType == OperandType::TENSOR_QUANT8_SYMM) {
132 const int8_t* inputBuffer = context->getInputBuffer<int8_t>(kInputTensor);
133 if (outputType == OperandType::TENSOR_FLOAT16) {
134 return compute(inputBuffer, inputShape,
135 context->getOutputBuffer<_Float16>(kOutputTensor));
136 } else if (outputType == OperandType::TENSOR_FLOAT32) {
137 return compute(inputBuffer, inputShape, context->getOutputBuffer<float>(kOutputTensor));
138 }
139 } else if (inputType == OperandType::TENSOR_QUANT8_ASYMM_SIGNED) {
140 const int8_t* inputBuffer = context->getInputBuffer<int8_t>(kInputTensor);
141 if (outputType == OperandType::TENSOR_FLOAT16) {
142 return compute(inputBuffer, inputShape,
143 context->getOutputBuffer<_Float16>(kOutputTensor));
144 } else if (outputType == OperandType::TENSOR_FLOAT32) {
145 return compute(inputBuffer, inputShape, context->getOutputBuffer<float>(kOutputTensor));
146 }
147 } else if (inputType == OperandType::TENSOR_QUANT8_SYMM_PER_CHANNEL) {
148 const int8_t* inputBuffer = context->getInputBuffer<int8_t>(kInputTensor);
149 if (outputType == OperandType::TENSOR_FLOAT16) {
150 return computePerChannel(inputBuffer, inputShape,
151 context->getOutputBuffer<_Float16>(kOutputTensor));
152 } else if (outputType == OperandType::TENSOR_FLOAT32) {
153 return computePerChannel(inputBuffer, inputShape,
154 context->getOutputBuffer<float>(kOutputTensor));
155 }
156 }
157 NN_RET_CHECK_FAIL() << "Unsupported tensor types combination for dequantize op. (input type: "
158 << toString(inputType) << " output type: " << toString(outputType) << ")";
159 }
160
161 } // namespace dequantize
162
163 NN_REGISTER_OPERATION(DEQUANTIZE, "DEQUANTIZE", dequantize::validate, dequantize::prepare,
164 dequantize::execute, .allowZeroSizedInput = true);
165
166 } // namespace nn
167 } // namespace android
168