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 <algorithm>
20 #include <cmath>
21
22 #include "IndexedShapeWrapper.h"
23 #include "OperationResolver.h"
24 #include "OperationsUtils.h"
25 #include "Tracing.h"
26
27 namespace android {
28 namespace nn {
29 namespace quantize {
30
31 constexpr uint32_t kNumInputs = 1;
32 constexpr uint32_t kInputTensor = 0;
33
34 constexpr uint32_t kNumOutputs = 1;
35 constexpr uint32_t kOutputTensor = 0;
36
37 namespace {
38
39 template <typename T>
quantizeToQuant8(const T * inputData,uint8_t * outputData,const Shape & outputShape)40 bool quantizeToQuant8(const T* inputData, uint8_t* outputData, const Shape& outputShape) {
41 NNTRACE_COMP("quantizeToQuant8");
42 uint32_t size = getNumberOfElements(outputShape);
43 for (uint32_t i = 0; i < size; ++i) {
44 outputData[i] = static_cast<uint8_t>(std::max<float>(
45 0.0f, std::min<float>(255.0f, outputShape.offset + std::round(inputData[i] /
46 outputShape.scale))));
47 }
48 return true;
49 }
50
51 template <typename T>
quantizeToQuant8Signed(const T * inputData,int8_t * outputData,const Shape & outputShape)52 bool quantizeToQuant8Signed(const T* inputData, int8_t* outputData, const Shape& outputShape) {
53 NNTRACE_COMP("quantizeToQuant8Signed");
54 uint32_t size = getNumberOfElements(outputShape);
55 for (uint32_t i = 0; i < size; ++i) {
56 outputData[i] = static_cast<int8_t>(std::max<float>(
57 -128.0f,
58 std::min<float>(127.0f, outputShape.offset +
59 std::round(inputData[i] / outputShape.scale))));
60 }
61 return true;
62 }
63
64 } // namespace
65
validate(const IOperationValidationContext * context)66 Result<Version> validate(const IOperationValidationContext* context) {
67 NN_RET_CHECK_EQ(context->getNumInputs(), kNumInputs);
68 NN_RET_CHECK_EQ(context->getNumOutputs(), kNumOutputs);
69
70 const OperandType inputType = context->getInputType(kInputTensor);
71 const OperandType outputType = context->getOutputType(kOutputTensor);
72
73 NN_RET_CHECK(inputType == OperandType::TENSOR_FLOAT16 ||
74 inputType == OperandType::TENSOR_FLOAT32)
75 << "Unsupported input operand type for QUANTIZE op: " << inputType;
76 NN_RET_CHECK(outputType == OperandType::TENSOR_QUANT8_ASYMM ||
77 outputType == OperandType::TENSOR_QUANT8_ASYMM_SIGNED)
78 << "Unsupported output operand type for QUANTIZE op: " << outputType;
79 if (outputType == OperandType::TENSOR_QUANT8_ASYMM_SIGNED) {
80 return Version::ANDROID_R;
81 } else {
82 return Version::ANDROID_Q;
83 }
84 }
85
prepare(IOperationExecutionContext * context)86 bool prepare(IOperationExecutionContext* context) {
87 const Shape& input = context->getInputShape(kInputTensor);
88 Shape output = context->getOutputShape(kOutputTensor);
89 output.dimensions = input.dimensions;
90 return context->setOutputShape(kOutputTensor, output);
91 }
92
execute(IOperationExecutionContext * context)93 bool execute(IOperationExecutionContext* context) {
94 // Bypass execution in the case of zero-sized input.
95 if (getNumberOfElements(context->getOutputShape(kOutputTensor)) == 0) return true;
96
97 const OperandType inputType = context->getInputType(kInputTensor);
98 const OperandType outputType = context->getOutputType(kOutputTensor);
99 if (inputType == OperandType::TENSOR_FLOAT32) {
100 if (outputType == OperandType::TENSOR_QUANT8_ASYMM) {
101 return quantizeToQuant8<float>(context->getInputBuffer<float>(kInputTensor),
102 context->getOutputBuffer<uint8_t>(kOutputTensor),
103 context->getOutputShape(kOutputTensor));
104 } else if (outputType == OperandType::TENSOR_QUANT8_ASYMM_SIGNED) {
105 return quantizeToQuant8Signed<float>(context->getInputBuffer<float>(kInputTensor),
106 context->getOutputBuffer<int8_t>(kOutputTensor),
107 context->getOutputShape(kOutputTensor));
108 }
109 } else if (inputType == OperandType::TENSOR_FLOAT16) {
110 if (outputType == OperandType::TENSOR_QUANT8_ASYMM) {
111 return quantizeToQuant8<_Float16>(context->getInputBuffer<_Float16>(kInputTensor),
112 context->getOutputBuffer<uint8_t>(kOutputTensor),
113 context->getOutputShape(kOutputTensor));
114 } else if (outputType == OperandType::TENSOR_QUANT8_ASYMM_SIGNED) {
115 return quantizeToQuant8Signed<_Float16>(context->getInputBuffer<_Float16>(kInputTensor),
116 context->getOutputBuffer<int8_t>(kOutputTensor),
117 context->getOutputShape(kOutputTensor));
118 }
119 }
120 NN_RET_CHECK_FAIL() << "Unsupported tensor types combination for QUANTIZE op. (input type: "
121 << inputType << " output type: " << context->getOutputType(kOutputTensor)
122 << ")";
123 }
124
125 } // namespace quantize
126
127 NN_REGISTER_OPERATION(QUANTIZE, "QUANTIZE", quantize::validate, quantize::prepare,
128 quantize::execute, .allowZeroSizedInput = true);
129
130 } // namespace nn
131 } // namespace android
132