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
|
/*
* 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 "OperationResolver.h"
#include "OperationsUtils.h"
#include "Tracing.h"
namespace android {
namespace nn {
namespace channel_shuffle {
constexpr char kOperationName[] = "CHANNEL_SHUFFLE";
constexpr uint32_t kNumInputs = 3;
constexpr uint32_t kInputTensor = 0;
constexpr uint32_t kNumGroups = 1;
constexpr uint32_t kInputAxis = 2;
constexpr uint32_t kNumOutputs = 1;
constexpr uint32_t kOutputTensor = 0;
template <typename T>
inline bool eval(const T* inputData, const Shape& inputShape, int32_t numGroups, int32_t axis,
T* outputData) {
const uint32_t outerSize = getNumberOfElements(inputShape, 0, axis);
const uint32_t axisSize = getSizeOfDimension(inputShape, axis);
const uint32_t innerSize =
getNumberOfElements(inputShape, axis + 1, getNumberOfDimensions(inputShape));
const uint32_t groupSize = axisSize / numGroups;
for (uint32_t outer = 0; outer < outerSize; ++outer) {
for (uint32_t inner = 0; inner < innerSize; ++inner) {
const T* inputBase = inputData + outer * axisSize * innerSize + inner;
T* outputBase = outputData + outer * axisSize * innerSize + inner;
for (uint32_t i = 0; i < groupSize; i++) {
for (uint32_t j = 0; j < static_cast<uint32_t>(numGroups);
j++, outputBase += innerSize) {
*outputBase = inputBase[innerSize * (i + j * groupSize)];
}
}
}
}
return true;
}
bool validate(const IOperationValidationContext* context) {
NN_RET_CHECK_EQ(context->getNumInputs(), kNumInputs);
NN_RET_CHECK_EQ(context->getNumOutputs(), kNumOutputs);
auto inputType = context->getInputType(kInputTensor);
NN_RET_CHECK(inputType == OperandType::TENSOR_FLOAT16 ||
inputType == OperandType::TENSOR_FLOAT32 ||
inputType == OperandType::TENSOR_QUANT8_ASYMM ||
inputType == OperandType::TENSOR_QUANT8_ASYMM_SIGNED)
<< "Unsupported tensor type for operation " << kOperationName;
const Shape& inputShape = context->getInputShape(kInputTensor);
if (hasKnownRank(inputShape)) {
NN_RET_CHECK_LE(getNumberOfDimensions(inputShape), 4);
}
NN_RET_CHECK(validateInputTypes(context, {inputType, OperandType::INT32, OperandType::INT32}));
NN_RET_CHECK(validateOutputTypes(context, {inputType}));
if (inputType == OperandType::TENSOR_QUANT8_ASYMM_SIGNED) {
return validateHalVersion(context, HalVersion::V1_3);
} else {
return validateHalVersion(context, HalVersion::V1_2);
}
}
bool prepare(IOperationExecutionContext* context) {
Shape input = context->getInputShape(kInputTensor);
int32_t numGroups = context->getInputValue<int32_t>(kNumGroups);
int32_t axis = context->getInputValue<int32_t>(kInputAxis);
NN_RET_CHECK(handleNegativeAxis(input, &axis));
NN_RET_CHECK(numGroups > 0);
NN_RET_CHECK(getSizeOfDimension(input, axis) % numGroups == 0);
return context->setOutputShape(kOutputTensor, input);
}
bool execute(IOperationExecutionContext* context) {
int32_t numGroups = context->getInputValue<int32_t>(kNumGroups);
int32_t axis = context->getInputValue<int32_t>(kInputAxis);
NN_RET_CHECK(handleNegativeAxis(context->getInputShape(kInputTensor), &axis));
switch (context->getInputType(kInputTensor)) {
case OperandType::TENSOR_FLOAT16:
return eval(context->getInputBuffer<_Float16>(kInputTensor),
context->getInputShape(kInputTensor), numGroups, axis,
context->getOutputBuffer<_Float16>(kOutputTensor));
case OperandType::TENSOR_FLOAT32:
return eval(context->getInputBuffer<float>(kInputTensor),
context->getInputShape(kInputTensor), numGroups, axis,
context->getOutputBuffer<float>(kOutputTensor));
case OperandType::TENSOR_QUANT8_ASYMM:
return eval(context->getInputBuffer<uint8_t>(kInputTensor),
context->getInputShape(kInputTensor), numGroups, axis,
context->getOutputBuffer<uint8_t>(kOutputTensor));
case OperandType::TENSOR_QUANT8_ASYMM_SIGNED:
return eval(context->getInputBuffer<int8_t>(kInputTensor),
context->getInputShape(kInputTensor), numGroups, axis,
context->getOutputBuffer<int8_t>(kOutputTensor));
default:
NN_RET_CHECK_FAIL() << "Unsupported tensor type for operation " << kOperationName;
}
}
} // namespace channel_shuffle
NN_REGISTER_OPERATION(CHANNEL_SHUFFLE, channel_shuffle::kOperationName, channel_shuffle::validate,
channel_shuffle::prepare, channel_shuffle::execute);
} // namespace nn
} // namespace android
|
