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 "Quantize.h"
20
21 #include <algorithm>
22 #include <cmath>
23
24 #include "IndexedShapeWrapper.h"
25 #include "OperationResolver.h"
26 #include "OperationsExecutionUtils.h"
27 #include "Tracing.h"
28
29 namespace android {
30 namespace nn {
31 namespace quantize {
32 namespace {
33
34 // The quantization formula also appears in Elementwise.cpp.
35 template <typename T>
quantizeToQuant8(const T * inputData,uint8_t * outputData,const Shape & outputShape)36 bool quantizeToQuant8(const T* inputData, uint8_t* outputData, const Shape& outputShape) {
37 NNTRACE_COMP("quantizeToQuant8");
38 uint32_t size = getNumberOfElements(outputShape);
39 for (uint32_t i = 0; i < size; ++i) {
40 outputData[i] = static_cast<uint8_t>(std::max<float>(
41 0.0f, std::min<float>(255.0f, outputShape.offset + std::round(inputData[i] /
42 outputShape.scale))));
43 }
44 return true;
45 }
46
47 // The quantization formula also appears in Elementwise.cpp.
48 template <typename T>
quantizeToQuant8Signed(const T * inputData,int8_t * outputData,const Shape & outputShape)49 bool quantizeToQuant8Signed(const T* inputData, int8_t* outputData, const Shape& outputShape) {
50 NNTRACE_COMP("quantizeToQuant8Signed");
51 uint32_t size = getNumberOfElements(outputShape);
52 for (uint32_t i = 0; i < size; ++i) {
53 outputData[i] = static_cast<int8_t>(std::max<float>(
54 -128.0f,
55 std::min<float>(127.0f, outputShape.offset +
56 std::round(inputData[i] / outputShape.scale))));
57 }
58 return true;
59 }
60
61 } // namespace
62
prepare(IOperationExecutionContext * context)63 bool prepare(IOperationExecutionContext* context) {
64 const Shape& input = context->getInputShape(kInputTensor);
65 Shape output = context->getOutputShape(kOutputTensor);
66 output.dimensions = input.dimensions;
67 return context->setOutputShape(kOutputTensor, output);
68 }
69
execute(IOperationExecutionContext * context)70 bool execute(IOperationExecutionContext* context) {
71 // Bypass execution in the case of zero-sized input.
72 if (getNumberOfElements(context->getOutputShape(kOutputTensor)) == 0) return true;
73
74 const OperandType inputType = context->getInputType(kInputTensor);
75 const OperandType outputType = context->getOutputType(kOutputTensor);
76 if (inputType == OperandType::TENSOR_FLOAT32) {
77 if (outputType == OperandType::TENSOR_QUANT8_ASYMM) {
78 return quantizeToQuant8<float>(context->getInputBuffer<float>(kInputTensor),
79 context->getOutputBuffer<uint8_t>(kOutputTensor),
80 context->getOutputShape(kOutputTensor));
81 } else if (outputType == OperandType::TENSOR_QUANT8_ASYMM_SIGNED) {
82 return quantizeToQuant8Signed<float>(context->getInputBuffer<float>(kInputTensor),
83 context->getOutputBuffer<int8_t>(kOutputTensor),
84 context->getOutputShape(kOutputTensor));
85 }
86 } else if (inputType == OperandType::TENSOR_FLOAT16) {
87 if (outputType == OperandType::TENSOR_QUANT8_ASYMM) {
88 return quantizeToQuant8<_Float16>(context->getInputBuffer<_Float16>(kInputTensor),
89 context->getOutputBuffer<uint8_t>(kOutputTensor),
90 context->getOutputShape(kOutputTensor));
91 } else if (outputType == OperandType::TENSOR_QUANT8_ASYMM_SIGNED) {
92 return quantizeToQuant8Signed<_Float16>(context->getInputBuffer<_Float16>(kInputTensor),
93 context->getOutputBuffer<int8_t>(kOutputTensor),
94 context->getOutputShape(kOutputTensor));
95 }
96 }
97 NN_RET_CHECK_FAIL() << "Unsupported tensor types combination for QUANTIZE op. (input type: "
98 << inputType << " output type: " << context->getOutputType(kOutputTensor)
99 << ")";
100 }
101
102 } // namespace quantize
103
104 NN_REGISTER_OPERATION_DEFAULT_VALIDATION(QUANTIZE, quantize::prepare, quantize::execute,
105 .allowZeroSizedInput = true);
106
107 } // namespace nn
108 } // namespace android
109