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 #include "CpuOperationUtils.h"
18 #include "OperationResolver.h"
19
20 #include "tensorflow/lite/kernels/internal/optimized/legacy_optimized_ops.h"
21 #include "tensorflow/lite/kernels/internal/reference/reference_ops.h"
22
23 #include "Tracing.h"
24
25 namespace android {
26 namespace nn {
27 namespace transpose {
28
29 constexpr char kOperationName[] = "TRANSPOSE";
30
31 constexpr uint32_t kNumInputs = 2;
32 constexpr uint32_t kInputTensor = 0;
33 constexpr uint32_t kPermTensor = 1;
34
35 constexpr uint32_t kNumOutputs = 1;
36 constexpr uint32_t kOutputTensor = 0;
37
38 namespace {
39
40 template <typename T>
transposeGeneric(const T * inputData,const Shape & inputShape,const int32_t * perm,const Shape & permShape,T * outputData,const Shape & outputShape)41 bool transposeGeneric(const T* inputData, const Shape& inputShape, const int32_t* perm,
42 const Shape& permShape, T* outputData, const Shape& outputShape) {
43 NNTRACE_TRANS("transposeGeneric");
44 // Reverse the permuted axes and convert to 4D due to the way Dims are
45 // constructed.
46 const int32_t kOutputDimensionNum = 4;
47
48 // permData can be NO_VALUE representing a regular 2D matrix transpose
49 int32_t permSize = perm == nullptr ? 2 : static_cast<int32_t>(getSizeOfDimension(permShape, 0));
50 int32_t perm_tmp[2] = {1, 0};
51 if (perm == nullptr) {
52 perm = perm_tmp;
53 }
54 int32_t reversed_perm[kOutputDimensionNum];
55 for (int32_t output_k = 0, input_k = permSize - 1; output_k < permSize; ++output_k, --input_k) {
56 reversed_perm[output_k] = permSize - perm[input_k] - 1;
57 }
58 for (int32_t k = permSize; k < kOutputDimensionNum; ++k) {
59 reversed_perm[k] = k;
60 }
61 NNTRACE_COMP_SWITCH("reference_ops::Transpose");
62 tflite::reference_ops::Transpose(inputData, convertShapeToDims(inputShape), outputData,
63 convertShapeToDims(outputShape), reversed_perm);
64 return true;
65 }
66
67 } // namespace
68
validate(const IOperationValidationContext * context)69 bool validate(const IOperationValidationContext* context) {
70 NN_RET_CHECK_EQ(context->getNumInputs(), kNumInputs);
71 NN_RET_CHECK_EQ(context->getNumOutputs(), kNumOutputs);
72
73 const OperandType inputType = context->getInputType(kInputTensor);
74 if (inputType == OperandType::TENSOR_FLOAT32 || inputType == OperandType::TENSOR_QUANT8_ASYMM) {
75 NN_RET_CHECK(validateHalVersion(context, HalVersion::V1_1));
76 } else if (inputType == OperandType::TENSOR_FLOAT16) {
77 NN_RET_CHECK(validateHalVersion(context, HalVersion::V1_2));
78 } else {
79 NN_RET_CHECK_FAIL() << "Unsupported tensor type for operation " << kOperationName;
80 }
81 return validateInputTypes(context, {inputType, OperandType::TENSOR_INT32}) &&
82 validateOutputTypes(context, {inputType});
83 }
84
prepare(IOperationExecutionContext * context)85 bool prepare(IOperationExecutionContext* context) {
86 // Only the permutation tensor can be omitted.
87 NN_RET_CHECK(!context->isOmittedInput(kInputTensor));
88 NN_RET_CHECK(!context->isOmittedOutput(kOutputTensor));
89
90 const Shape& input = context->getInputShape(kInputTensor);
91 uint32_t numInputDims = getNumberOfDimensions(input);
92 Shape output = context->getOutputShape(kOutputTensor);
93 output.type = input.type;
94 output.offset = input.offset;
95 output.scale = input.scale;
96
97 // permData can be NO_VALUE representing a regular 2D matrix transpose
98 if (context->isOmittedInput(kPermTensor)) {
99 NN_RET_CHECK_EQ(numInputDims, 2);
100 output.dimensions = {getSizeOfDimension(input, 1), getSizeOfDimension(input, 0)};
101 } else {
102 const Shape& permShape = context->getInputShape(kPermTensor);
103 const int32_t* permData = context->getInputBuffer<int32_t>(kPermTensor);
104
105 // Transpose op only supports 1D-4D input arrays.
106 NN_RET_CHECK_LE(numInputDims, 4);
107
108 // perm need to be provided as a 1-D int32 tensor.
109 NN_RET_CHECK(permShape.type == OperandType::TENSOR_INT32);
110 NN_RET_CHECK_EQ(getNumberOfDimensions(permShape), 1);
111 NN_RET_CHECK_EQ(numInputDims, getSizeOfDimension(permShape, 0));
112
113 std::vector<uint32_t> outDims(numInputDims);
114 for (int32_t idx = 0; idx < static_cast<int32_t>(numInputDims); ++idx) {
115 NN_RET_CHECK(permData[idx] >= 0 && permData[idx] < static_cast<int32_t>(numInputDims));
116 outDims[idx] = getSizeOfDimension(input, permData[idx]);
117 }
118 output.dimensions = outDims;
119 }
120 return context->setOutputShape(kOutputTensor, output);
121 }
122
execute(IOperationExecutionContext * context)123 bool execute(IOperationExecutionContext* context) {
124 // Bypass execution in the case of zero-sized input.
125 if (getNumberOfElements(context->getOutputShape(kOutputTensor)) == 0) return true;
126
127 switch (context->getInputType(kInputTensor)) {
128 case OperandType::TENSOR_FLOAT32:
129 return transposeGeneric(context->getInputBuffer<float>(kInputTensor),
130 context->getInputShape(kInputTensor),
131 context->getInputBuffer<int32_t>(kPermTensor),
132 context->getInputShape(kPermTensor),
133 context->getOutputBuffer<float>(kOutputTensor),
134 context->getOutputShape(kOutputTensor));
135 case OperandType::TENSOR_FLOAT16:
136 return transposeGeneric(context->getInputBuffer<_Float16>(kInputTensor),
137 context->getInputShape(kInputTensor),
138 context->getInputBuffer<int32_t>(kPermTensor),
139 context->getInputShape(kPermTensor),
140 context->getOutputBuffer<_Float16>(kOutputTensor),
141 context->getOutputShape(kOutputTensor));
142 case OperandType::TENSOR_QUANT8_ASYMM:
143 return transposeGeneric(context->getInputBuffer<uint8_t>(kInputTensor),
144 context->getInputShape(kInputTensor),
145 context->getInputBuffer<int32_t>(kPermTensor),
146 context->getInputShape(kPermTensor),
147 context->getOutputBuffer<uint8_t>(kOutputTensor),
148 context->getOutputShape(kOutputTensor));
149 default:
150 NN_RET_CHECK_FAIL() << "Unsupported tensor type for operation " << kOperationName;
151 }
152 }
153
154 } // namespace transpose
155
156 NN_REGISTER_OPERATION(TRANSPOSE, transpose::kOperationName, transpose::validate, transpose::prepare,
157 transpose::execute, .allowOmittedOperand = true, .allowZeroSizedInput = true);
158
159 } // namespace nn
160 } // namespace android
161