1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
|
/*
* Copyright (C) 2018 The Android Open Source Project
*
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
*/
#define LOG_TAG "Operations"
#include <algorithm>
#include <utility>
#include <vector>
#include "OperationResolver.h"
#include "RNN.h"
#include "nnapi/TypeUtils.h"
namespace android {
namespace nn {
namespace unidirectional_sequence_rnn {
constexpr uint32_t kNumInputs = 7;
constexpr uint32_t kInputTensor = 0;
constexpr uint32_t kWeightsTensor = 1;
constexpr uint32_t kRecurrentWeightsTensor = 2;
constexpr uint32_t kBiasTensor = 3;
constexpr uint32_t kHiddenStateTensor = 4;
constexpr uint32_t kActivationParam = 5;
constexpr uint32_t kTimeMajorParam = 6;
constexpr uint32_t kNumOutputs = 1;
constexpr uint32_t kNumOutputsWithState = 2;
constexpr uint32_t kOutputTensor = 0;
constexpr uint32_t kStateOutputTensor = 1;
namespace {
template <typename T>
void transposeFirstTwoDims(const T* input, const Shape& inputShape, T* output) {
const uint32_t firstDimSize = getSizeOfDimension(inputShape, 0);
const uint32_t secondDimSize = getSizeOfDimension(inputShape, 1);
const uint32_t inputSize = getSizeOfDimension(inputShape, 2);
for (int f = 0; f < firstDimSize; ++f) {
for (int s = 0; s < secondDimSize; ++s) {
for (int i = 0; i < inputSize; ++i) {
const uint32_t inputIndex = f * secondDimSize * inputSize + s * inputSize + i;
const uint32_t outputIndex = s * firstDimSize * inputSize + f * inputSize + i;
output[outputIndex] = input[inputIndex];
}
}
}
}
template <typename T>
bool executeTyped(IOperationExecutionContext* context) {
const T* input = context->getInputBuffer<T>(kInputTensor);
Shape inputShape = context->getInputShape(kInputTensor);
const T* weights = context->getInputBuffer<T>(kWeightsTensor);
Shape weightsShape = context->getInputShape(kWeightsTensor);
const T* recurrentWeights = context->getInputBuffer<T>(kRecurrentWeightsTensor);
Shape recurrentWeightsShape = context->getInputShape(kRecurrentWeightsTensor);
const T* bias = context->getInputBuffer<T>(kBiasTensor);
const T* hiddenState = context->getInputBuffer<T>(kHiddenStateTensor);
int32_t activation = context->getInputValue<int32_t>(kActivationParam);
T* output = context->getOutputBuffer<T>(kOutputTensor);
Shape outputShape = context->getOutputShape(kOutputTensor);
int32_t timeMajor = context->getInputValue<int32_t>(kTimeMajorParam);
// If the input tensors are not in time major format, we transpose the first
// two dimensions, and set input and output pointers to temporary vectors
// which are transposed back after the RNN is applied.
std::vector<T> inputTransposed;
std::vector<T> outputTransposed;
if (!timeMajor) {
// Convert input and output to time major format.
inputTransposed.resize(getNumberOfElements(inputShape));
outputTransposed.resize(getNumberOfElements(outputShape));
transposeFirstTwoDims(input, inputShape, inputTransposed.data());
input = inputTransposed.data();
output = outputTransposed.data();
std::swap(inputShape.dimensions[0], inputShape.dimensions[1]);
std::swap(outputShape.dimensions[0], outputShape.dimensions[1]);
}
const uint32_t maxTime = getSizeOfDimension(inputShape, 0);
const uint32_t batchSize = getSizeOfDimension(inputShape, 1);
const uint32_t inputSize = getSizeOfDimension(inputShape, 2);
const uint32_t numUnits = getSizeOfDimension(weightsShape, 0);
// A shape at a fixed step (removed time dimension).
Shape fixedTimeInputShape = inputShape;
fixedTimeInputShape.dimensions.resize(2);
fixedTimeInputShape.dimensions[0] = inputShape.dimensions[1];
fixedTimeInputShape.dimensions[1] = inputShape.dimensions[2];
for (int i = 0; i < maxTime; ++i) {
RNN::RNNStep<T>(input, fixedTimeInputShape, hiddenState, bias, weights, weightsShape,
recurrentWeights, recurrentWeightsShape, activation, output);
input += batchSize * inputSize;
hiddenState = output;
output += batchSize * numUnits;
}
if (!timeMajor) {
transposeFirstTwoDims(outputTransposed.data(), outputShape,
context->getOutputBuffer<T>(kOutputTensor));
}
if (context->getNumOutputs() == kNumOutputsWithState) {
// We checked that the state output is not omitted during preparation.
T* stateOutput = context->getOutputBuffer<T>(kStateOutputTensor);
std::copy(hiddenState, hiddenState + batchSize * numUnits, stateOutput);
}
return true;
}
} // namespace
bool validate(const IOperationValidationContext* context) {
NN_RET_CHECK_EQ(context->getNumInputs(), kNumInputs);
const int numOutputs = context->getNumOutputs();
NN_RET_CHECK(numOutputs == kNumOutputs || numOutputs == kNumOutputsWithState);
OperandType inputType = context->getInputType(kInputTensor);
if (inputType != OperandType::TENSOR_FLOAT16 && inputType != OperandType::TENSOR_FLOAT32) {
LOG(ERROR) << "Unsupported input operand type for UNIDIRECTIONAL_SEQUENCE_RNN op: "
<< inputType;
return false;
}
NN_RET_CHECK(validateInputTypes(context, {inputType, inputType, inputType, inputType, inputType,
OperandType::INT32, OperandType::INT32}));
std::vector<OperandType> outputTypes = {inputType};
Version minVersionSupported = Version::ANDROID_Q;
if (numOutputs == kNumOutputsWithState) {
minVersionSupported = Version::ANDROID_R;
outputTypes.push_back(inputType);
}
NN_RET_CHECK(validateOutputTypes(context, outputTypes));
return validateVersion(context, minVersionSupported);
}
bool prepare(IOperationExecutionContext* context) {
Shape input = context->getInputShape(kInputTensor);
Shape weights = context->getInputShape(kWeightsTensor);
Shape recurrentWeights = context->getInputShape(kRecurrentWeightsTensor);
Shape bias = context->getInputShape(kBiasTensor);
Shape hiddenState = context->getInputShape(kHiddenStateTensor);
int32_t timeMajor = context->getInputValue<int32_t>(kTimeMajorParam);
NN_RET_CHECK(timeMajor == 0 || timeMajor == 1);
const uint32_t batchSize =
timeMajor ? getSizeOfDimension(input, 1) : getSizeOfDimension(input, 0);
const uint32_t maxTime =
timeMajor ? getSizeOfDimension(input, 0) : getSizeOfDimension(input, 1);
const uint32_t numUnits = getSizeOfDimension(weights, 0);
const uint32_t inputSize = getSizeOfDimension(input, 2);
NN_RET_CHECK_EQ(getNumberOfDimensions(input), 3);
NN_RET_CHECK_EQ(getNumberOfDimensions(weights), 2);
NN_RET_CHECK_EQ(getNumberOfDimensions(recurrentWeights), 2);
NN_RET_CHECK_EQ(getNumberOfDimensions(bias), 1);
NN_RET_CHECK_EQ(getNumberOfDimensions(hiddenState), 2);
NN_RET_CHECK_EQ(inputSize, getSizeOfDimension(weights, 1));
NN_RET_CHECK_EQ(numUnits, getSizeOfDimension(bias, 0));
NN_RET_CHECK_EQ(numUnits, getSizeOfDimension(recurrentWeights, 0));
NN_RET_CHECK_EQ(numUnits, getSizeOfDimension(recurrentWeights, 1));
NN_RET_CHECK_EQ(batchSize, getSizeOfDimension(hiddenState, 0));
NN_RET_CHECK_EQ(numUnits, getSizeOfDimension(hiddenState, 1));
Shape output = context->getOutputShape(kOutputTensor);
output.dimensions.resize(3);
output.dimensions[0] = timeMajor ? maxTime : batchSize;
output.dimensions[1] = timeMajor ? batchSize : maxTime;
output.dimensions[2] = numUnits;
if (context->getNumOutputs() == kNumOutputsWithState) {
NN_RET_CHECK(!context->isOmittedOutput(kStateOutputTensor));
Shape outputStateShape = context->getInputShape(kHiddenStateTensor);
outputStateShape.dimensions.resize(2);
outputStateShape.dimensions[0] = batchSize;
outputStateShape.dimensions[1] = numUnits;
NN_RET_CHECK(context->setOutputShape(kStateOutputTensor, outputStateShape));
}
return context->setOutputShape(kOutputTensor, output);
}
bool execute(IOperationExecutionContext* context) {
if (context->getInputType(kInputTensor) == OperandType::TENSOR_FLOAT16) {
executeTyped<_Float16>(context);
} else {
executeTyped<float>(context);
}
return true;
}
} // namespace unidirectional_sequence_rnn
NN_REGISTER_OPERATION(UNIDIRECTIONAL_SEQUENCE_RNN, "UNIDIRECTIONAL_SEQUENCE_RNN",
unidirectional_sequence_rnn::validate, unidirectional_sequence_rnn::prepare,
unidirectional_sequence_rnn::execute);
} // namespace nn
} // namespace android
|
