1 | /* Copyright 2019 The TensorFlow Authors. All Rights Reserved. |
2 | |
3 | Licensed under the Apache License, Version 2.0 (the "License"); |
4 | you may not use this file except in compliance with the License. |
5 | You may obtain a copy of the License at |
6 | |
7 | http://www.apache.org/licenses/LICENSE-2.0 |
8 | |
9 | Unless required by applicable law or agreed to in writing, software |
10 | distributed under the License is distributed on an "AS IS" BASIS, |
11 | WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. |
12 | See the License for the specific language governing permissions and |
13 | limitations under the License. |
14 | ==============================================================================*/ |
15 | #include <stdint.h> |
16 | |
17 | #include "tensorflow/lite/c/c_api_types.h" |
18 | #include "tensorflow/lite/c/common.h" |
19 | #include "tensorflow/lite/kernels/internal/optimized/optimized_ops.h" |
20 | #include "tensorflow/lite/kernels/internal/reference/reference_ops.h" |
21 | #include "tensorflow/lite/kernels/internal/tensor.h" |
22 | #include "tensorflow/lite/kernels/internal/tensor_ctypes.h" |
23 | #include "tensorflow/lite/kernels/kernel_util.h" |
24 | |
25 | namespace tflite { |
26 | namespace ops { |
27 | namespace builtin { |
28 | namespace gather_nd { |
29 | constexpr int kParams = 0; |
30 | constexpr int kIndices = 1; |
31 | constexpr int kOutputTensor = 0; |
32 | |
33 | TfLiteStatus Prepare(TfLiteContext* context, TfLiteNode* node) { |
34 | TF_LITE_ENSURE_EQ(context, NumInputs(node), 2); |
35 | TF_LITE_ENSURE_EQ(context, NumOutputs(node), 1); |
36 | |
37 | const TfLiteTensor* params; |
38 | TF_LITE_ENSURE_OK(context, GetInputSafe(context, node, kParams, ¶ms)); |
39 | const TfLiteTensor* indices; |
40 | TF_LITE_ENSURE_OK(context, GetInputSafe(context, node, kIndices, &indices)); |
41 | TfLiteTensor* output; |
42 | TF_LITE_ENSURE_OK(context, |
43 | GetOutputSafe(context, node, kOutputTensor, &output)); |
44 | |
45 | switch (params->type) { |
46 | case kTfLiteFloat32: |
47 | case kTfLiteUInt8: |
48 | case kTfLiteInt8: |
49 | case kTfLiteInt16: |
50 | case kTfLiteInt64: |
51 | case kTfLiteInt32: |
52 | case kTfLiteString: |
53 | break; |
54 | default: |
55 | TF_LITE_KERNEL_LOG(context, |
56 | "Params of type '%s' are not supported by gather_nd." , |
57 | TfLiteTypeGetName(params->type)); |
58 | return kTfLiteError; |
59 | } |
60 | switch (indices->type) { |
61 | case kTfLiteInt64: |
62 | case kTfLiteInt32: |
63 | break; |
64 | default: |
65 | TF_LITE_KERNEL_LOG(context, |
66 | "Indices of type '%s' are not supported by gather_nd." , |
67 | TfLiteTypeGetName(indices->type)); |
68 | return kTfLiteError; |
69 | } |
70 | |
71 | const int params_rank = NumDimensions(params); |
72 | const int indices_rank = NumDimensions(indices); |
73 | const int indices_nd = SizeOfDimension(indices, indices_rank - 1); |
74 | if (params_rank < 1) { |
75 | TF_LITE_KERNEL_LOG(context, "Params must be at least a vector." ); |
76 | return kTfLiteError; |
77 | } |
78 | if (indices_rank < 1) { |
79 | TF_LITE_KERNEL_LOG(context, "Indices must be at least a vector." ); |
80 | return kTfLiteError; |
81 | } |
82 | if (indices_nd > params_rank) { |
83 | TF_LITE_KERNEL_LOG( |
84 | context, "Index innermost dimension length must be <= params rank." ); |
85 | return kTfLiteError; |
86 | } |
87 | |
88 | // Assign to output the input type. |
89 | output->type = params->type; |
90 | |
91 | // The result shape is |
92 | // indices.shape[:-1] + params.shape[indices.shape[-1]:] |
93 | const int output_rank = indices_rank + params_rank - indices_nd - 1; |
94 | TfLiteIntArray* output_shape = TfLiteIntArrayCreate(output_rank); |
95 | int output_index = 0; |
96 | for (int i = 0; i < indices_rank - 1; ++i) { |
97 | output_shape->data[output_index++] = indices->dims->data[i]; |
98 | } |
99 | for (int i = indices_nd; i < params_rank; ++i) { |
100 | output_shape->data[output_index++] = params->dims->data[i]; |
101 | } |
102 | return context->ResizeTensor(context, output, output_shape); |
103 | } |
104 | |
105 | template <typename ParamsT, typename IndicesT> |
106 | TfLiteStatus GatherNd(const TfLiteTensor* params, const TfLiteTensor* indices, |
107 | TfLiteTensor* output) { |
108 | return reference_ops::GatherNd( |
109 | GetTensorShape(params), GetTensorData<ParamsT>(params), |
110 | GetTensorShape(indices), GetTensorData<IndicesT>(indices), |
111 | GetTensorShape(output), GetTensorData<ParamsT>(output)); |
112 | } |
113 | |
114 | template <typename IndicesT> |
115 | TfLiteStatus GatherNdString(const TfLiteTensor* params, |
116 | const TfLiteTensor* indices, TfLiteTensor* output) { |
117 | return reference_ops::GatherNdString( |
118 | GetTensorShape(params), params, GetTensorShape(indices), |
119 | GetTensorData<IndicesT>(indices), GetTensorShape(output), output); |
120 | } |
121 | |
122 | template <typename IndicesT> |
123 | TfLiteStatus EvalGatherNd(TfLiteContext* context, const TfLiteTensor* params, |
124 | const TfLiteTensor* indices, TfLiteTensor* output) { |
125 | bool indices_has_only_positive_elements = true; |
126 | const auto* indices_values = GetTensorData<IndicesT>(indices); |
127 | const size_t num_indices = indices->bytes / sizeof(IndicesT); |
128 | for (size_t i = 0; i < num_indices; i++) { |
129 | if (indices_values[i] < 0) { |
130 | indices_has_only_positive_elements = false; |
131 | break; |
132 | } |
133 | } |
134 | TF_LITE_ENSURE(context, indices_has_only_positive_elements); |
135 | |
136 | TfLiteStatus status = kTfLiteError; |
137 | switch (params->type) { |
138 | case kTfLiteFloat32: |
139 | status = GatherNd<float, IndicesT>(params, indices, output); |
140 | break; |
141 | case kTfLiteUInt8: |
142 | status = GatherNd<uint8_t, IndicesT>(params, indices, output); |
143 | break; |
144 | case kTfLiteInt8: |
145 | status = GatherNd<int8_t, IndicesT>(params, indices, output); |
146 | break; |
147 | case kTfLiteInt16: |
148 | status = GatherNd<int16_t, IndicesT>(params, indices, output); |
149 | break; |
150 | case kTfLiteInt32: |
151 | status = GatherNd<int32_t, IndicesT>(params, indices, output); |
152 | break; |
153 | case kTfLiteInt64: |
154 | status = GatherNd<int64_t, IndicesT>(params, indices, output); |
155 | break; |
156 | case kTfLiteString: |
157 | status = GatherNdString<IndicesT>(params, indices, output); |
158 | break; |
159 | default: |
160 | TF_LITE_KERNEL_LOG(context, |
161 | "Params type '%s' are not supported by gather_nd." , |
162 | TfLiteTypeGetName(params->type)); |
163 | return kTfLiteError; |
164 | } |
165 | if (status != kTfLiteOk) { |
166 | TF_LITE_KERNEL_LOG(context, "gather_nd index out of bounds" ); |
167 | } |
168 | return status; |
169 | } |
170 | |
171 | TfLiteStatus Eval(TfLiteContext* context, TfLiteNode* node) { |
172 | const TfLiteTensor* params; |
173 | TF_LITE_ENSURE_OK(context, GetInputSafe(context, node, kParams, ¶ms)); |
174 | const TfLiteTensor* indices; |
175 | TF_LITE_ENSURE_OK(context, GetInputSafe(context, node, kIndices, &indices)); |
176 | TfLiteTensor* output; |
177 | TF_LITE_ENSURE_OK(context, |
178 | GetOutputSafe(context, node, kOutputTensor, &output)); |
179 | |
180 | // Prevent division by 0 in the helper. |
181 | // In TF, GatherND supports empty `params` only when `indices` is also empty. |
182 | TF_LITE_ENSURE(context, |
183 | (NumElements(params) == 0 && NumElements(indices) == 0) || |
184 | NumElements(params) > 0); |
185 | |
186 | switch (indices->type) { |
187 | case kTfLiteInt32: |
188 | return EvalGatherNd<int32_t>(context, params, indices, output); |
189 | case kTfLiteInt64: |
190 | return EvalGatherNd<int64_t>(context, params, indices, output); |
191 | default: |
192 | TF_LITE_KERNEL_LOG(context, |
193 | "Indices of type '%s' are not supported by gather_nd." , |
194 | TfLiteTypeGetName(indices->type)); |
195 | return kTfLiteError; |
196 | } |
197 | } |
198 | } // namespace gather_nd |
199 | |
200 | TfLiteRegistration* Register_GATHER_ND() { |
201 | static TfLiteRegistration r = {/*init*/ nullptr, /*free*/ nullptr, |
202 | gather_nd::Prepare, gather_nd::Eval}; |
203 | return &r; |
204 | } |
205 | } // namespace builtin |
206 | } // namespace ops |
207 | } // namespace tflite |
208 | |