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 #define LOG_TAG "Operations"
18
19 #include "Dequantize.h"
20
21 #include "IndexedShapeWrapper.h"
22 #include "OperationResolver.h"
23 #include "OperationsExecutionUtils.h"
24
25 namespace android {
26 namespace nn {
27 namespace dequantize {
28 namespace {
29
30 template <typename InputType, typename OutputType>
compute(const InputType * inputData,const Shape & inputShape,OutputType * outputData)31 bool compute(const InputType* inputData, const Shape& inputShape, OutputType* outputData) {
32 const int numElements = getNumberOfElements(inputShape);
33 const int32_t zeroPoint = inputShape.offset;
34 const float scale = inputShape.scale;
35 for (int i = 0; i < numElements; ++i) {
36 const int32_t value = inputData[i];
37 // This dequantization formula also appears in Elementwise.cpp.
38 outputData[i] = static_cast<OutputType>(scale * (value - zeroPoint));
39 }
40 return true;
41 }
42
43 template <typename OutputType>
computePerChannel(const int8_t * inputData,const Shape & inputShape,OutputType * outputData)44 bool computePerChannel(const int8_t* inputData, const Shape& inputShape, OutputType* outputData) {
45 // First we calculate a stride which is the number of elements we need to
46 // skip to change an index along a dimension with different quantization
47 // scales.
48 const int channelDim =
49 std::get<Operand::SymmPerChannelQuantParams>(inputShape.extraParams).channelDim;
50 int stride = 1;
51 for (int i = getNumberOfDimensions(inputShape) - 1; i > channelDim; --i) {
52 stride *= getSizeOfDimension(inputShape, i);
53 }
54
55 const int numElements = getNumberOfElements(inputShape);
56 const int32_t zeroPoint = inputShape.offset;
57
58 for (int i = 0; i < numElements; ++i) {
59 // To get current index along the quantized dimension we calculate how
60 // many even |strides| we looped through and take this number modulo the
61 // size of the dimension (so that we don't have an overflow if the
62 // channelDim is not 0).
63 const int scaleIndex = (i / stride) % getSizeOfDimension(inputShape, channelDim);
64 const float scale = std::get<Operand::SymmPerChannelQuantParams>(inputShape.extraParams)
65 .scales[scaleIndex];
66 const int32_t value = inputData[i];
67 outputData[i] = static_cast<OutputType>(scale * (value - zeroPoint));
68 }
69 return true;
70 }
71
72 } // namespace
73
prepare(IOperationExecutionContext * context)74 bool prepare(IOperationExecutionContext* context) {
75 const Shape& input = context->getInputShape(kInputTensor);
76 NN_RET_CHECK_LE(getNumberOfDimensions(input), 4u);
77 Shape output = context->getOutputShape(kOutputTensor);
78 output.dimensions = input.dimensions;
79 return context->setOutputShape(kOutputTensor, output);
80 }
81
execute(IOperationExecutionContext * context)82 bool execute(IOperationExecutionContext* context) {
83 // Bypass execution in the case of zero-sized input.
84 if (getNumberOfElements(context->getOutputShape(kOutputTensor)) == 0) return true;
85
86 const OperandType inputType = context->getInputType(kInputTensor);
87 const OperandType outputType = context->getOutputType(kOutputTensor);
88
89 const Shape& inputShape = context->getInputShape(kInputTensor);
90 if (inputType == OperandType::TENSOR_QUANT8_ASYMM) {
91 const uint8_t* inputBuffer = context->getInputBuffer<uint8_t>(kInputTensor);
92 if (outputType == OperandType::TENSOR_FLOAT16) {
93 return compute(inputBuffer, inputShape,
94 context->getOutputBuffer<_Float16>(kOutputTensor));
95 } else if (outputType == OperandType::TENSOR_FLOAT32) {
96 return compute(inputBuffer, inputShape, context->getOutputBuffer<float>(kOutputTensor));
97 }
98 } else if (inputType == OperandType::TENSOR_QUANT8_SYMM) {
99 const int8_t* inputBuffer = context->getInputBuffer<int8_t>(kInputTensor);
100 if (outputType == OperandType::TENSOR_FLOAT16) {
101 return compute(inputBuffer, inputShape,
102 context->getOutputBuffer<_Float16>(kOutputTensor));
103 } else if (outputType == OperandType::TENSOR_FLOAT32) {
104 return compute(inputBuffer, inputShape, context->getOutputBuffer<float>(kOutputTensor));
105 }
106 } else if (inputType == OperandType::TENSOR_QUANT8_ASYMM_SIGNED) {
107 const int8_t* inputBuffer = context->getInputBuffer<int8_t>(kInputTensor);
108 if (outputType == OperandType::TENSOR_FLOAT16) {
109 return compute(inputBuffer, inputShape,
110 context->getOutputBuffer<_Float16>(kOutputTensor));
111 } else if (outputType == OperandType::TENSOR_FLOAT32) {
112 return compute(inputBuffer, inputShape, context->getOutputBuffer<float>(kOutputTensor));
113 }
114 } else if (inputType == OperandType::TENSOR_QUANT8_SYMM_PER_CHANNEL) {
115 const int8_t* inputBuffer = context->getInputBuffer<int8_t>(kInputTensor);
116 if (outputType == OperandType::TENSOR_FLOAT16) {
117 return computePerChannel(inputBuffer, inputShape,
118 context->getOutputBuffer<_Float16>(kOutputTensor));
119 } else if (outputType == OperandType::TENSOR_FLOAT32) {
120 return computePerChannel(inputBuffer, inputShape,
121 context->getOutputBuffer<float>(kOutputTensor));
122 }
123 }
124 NN_RET_CHECK_FAIL() << "Unsupported tensor types combination for dequantize op. (input type: "
125 << inputType << " output type: " << outputType << ")";
126 }
127
128 } // namespace dequantize
129
130 NN_REGISTER_OPERATION_DEFAULT_VALIDATION(DEQUANTIZE, dequantize::prepare, dequantize::execute,
131 .allowZeroSizedInput = true);
132
133 } // namespace nn
134 } // namespace android
135