1 | /* Copyright 2017 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 "tensorflow/lite/c/builtin_op_data.h" |
16 | #include "tensorflow/lite/c/common.h" |
17 | #include "tensorflow/lite/kernels/internal/compatibility.h" |
18 | #include "tensorflow/lite/kernels/internal/optimized/optimized_ops.h" |
19 | #include "tensorflow/lite/kernels/internal/reference/integer_ops/l2normalization.h" |
20 | #include "tensorflow/lite/kernels/internal/reference/l2normalization.h" |
21 | #include "tensorflow/lite/kernels/internal/reference/reference_ops.h" |
22 | #include "tensorflow/lite/kernels/internal/tensor.h" |
23 | #include "tensorflow/lite/kernels/internal/tensor_ctypes.h" |
24 | #include "tensorflow/lite/kernels/internal/types.h" |
25 | #include "tensorflow/lite/kernels/kernel_util.h" |
26 | |
27 | namespace tflite { |
28 | namespace ops { |
29 | namespace builtin { |
30 | namespace l2norm { |
31 | |
32 | // This file has two implementation of L2Norm. |
33 | enum KernelType { |
34 | kReference, |
35 | kGenericOptimized, |
36 | }; |
37 | |
38 | constexpr int kInputTensor = 0; |
39 | constexpr int kOutputTensor = 0; |
40 | |
41 | TfLiteStatus Prepare(TfLiteContext* context, TfLiteNode* node) { |
42 | auto* params = reinterpret_cast<TfLiteL2NormParams*>(node->builtin_data); |
43 | |
44 | TF_LITE_ENSURE_EQ(context, NumInputs(node), 1); |
45 | TF_LITE_ENSURE_EQ(context, NumOutputs(node), 1); |
46 | |
47 | const TfLiteTensor* input; |
48 | TF_LITE_ENSURE_OK(context, GetInputSafe(context, node, kInputTensor, &input)); |
49 | TfLiteTensor* output; |
50 | TF_LITE_ENSURE_OK(context, |
51 | GetOutputSafe(context, node, kOutputTensor, &output)); |
52 | |
53 | TF_LITE_ENSURE(context, NumDimensions(input) <= 4); |
54 | |
55 | TF_LITE_ENSURE(context, output->type == kTfLiteFloat32 || |
56 | output->type == kTfLiteUInt8 || |
57 | output->type == kTfLiteInt8); |
58 | TF_LITE_ENSURE_TYPES_EQ(context, input->type, output->type); |
59 | |
60 | if (output->type == kTfLiteUInt8 || output->type == kTfLiteInt8) { |
61 | TF_LITE_ENSURE_EQ(context, output->params.scale, (1. / 128.)); |
62 | if (output->type == kTfLiteUInt8) { |
63 | TF_LITE_ENSURE_EQ(context, output->params.zero_point, 128); |
64 | } |
65 | if (output->type == kTfLiteInt8) { |
66 | TF_LITE_ENSURE_EQ(context, output->params.zero_point, 0); |
67 | } |
68 | } |
69 | |
70 | // TODO(ahentz): For some reason our implementations don't support |
71 | // activations. |
72 | TF_LITE_ENSURE_EQ(context, params->activation, kTfLiteActNone); |
73 | |
74 | TfLiteIntArray* output_size = TfLiteIntArrayCopy(input->dims); |
75 | return context->ResizeTensor(context, output, output_size); |
76 | } |
77 | |
78 | template <KernelType kernel_type> |
79 | TfLiteStatus Eval(TfLiteContext* context, TfLiteNode* node) { |
80 | const TfLiteTensor* input; |
81 | TF_LITE_ENSURE_OK(context, GetInputSafe(context, node, kInputTensor, &input)); |
82 | TfLiteTensor* output; |
83 | TF_LITE_ENSURE_OK(context, |
84 | GetOutputSafe(context, node, kOutputTensor, &output)); |
85 | |
86 | // TODO(b/143912164): instead of hardcode the epsilon here, we should read it |
87 | // from tensorflow, i.e., adding a params. |
88 | // We don't compute epsilon for quantized kernel: |
89 | // |
90 | // epsilon_float = (epsilon_quant - zp) * scale |
91 | // so |
92 | // espsilon_quant = epsilon_float / scale + zp |
93 | // We know epsilon_float is just a very small number to avoid division by |
94 | // zero error, and scale is > 1, so the integer value of epsilon for quant |
95 | // is just dominated by the zero point. |
96 | // Also, GetInvSqrtQuantizedMultiplierExp handles the scenario where the sum |
97 | // of input value squared is zero case well. |
98 | // So we don't even need to do handle the epsilon for quantized kernel case. |
99 | const float epsilon = 1e-6f; |
100 | if (output->type == kTfLiteFloat32) { |
101 | #define TF_LITE_L2NORM(type) \ |
102 | tflite::L2NormalizationParams op_params; \ |
103 | op_params.input_zero_point = 0; \ |
104 | type::L2Normalization(op_params, GetTensorShape(input), \ |
105 | GetTensorData<float>(input), GetTensorShape(output), \ |
106 | GetTensorData<float>(output), epsilon) |
107 | |
108 | if (kernel_type == kReference) { |
109 | TF_LITE_L2NORM(reference_ops); |
110 | } |
111 | if (kernel_type == kGenericOptimized) { |
112 | TF_LITE_L2NORM(optimized_ops); |
113 | } |
114 | #undef TF_LITE_L2NORM |
115 | } else if (output->type == kTfLiteUInt8) { |
116 | #define TF_LITE_L2NORM(type) \ |
117 | tflite::L2NormalizationParams op_params; \ |
118 | op_params.input_zero_point = input->params.zero_point; \ |
119 | type::L2Normalization(op_params, GetTensorShape(input), \ |
120 | GetTensorData<uint8>(input), GetTensorShape(output), \ |
121 | GetTensorData<uint8>(output)) |
122 | |
123 | if (kernel_type == kReference) { |
124 | TF_LITE_L2NORM(reference_ops); |
125 | } |
126 | if (kernel_type == kGenericOptimized) { |
127 | TF_LITE_L2NORM(optimized_ops); |
128 | } |
129 | #undef TF_LITE_L2NORM |
130 | } else if (output->type == kTfLiteInt8) { |
131 | const auto input_shape = GetTensorShape(input); |
132 | const auto output_shape = GetTensorShape(output); |
133 | const int trailing_dim = input_shape.DimensionsCount() - 1; |
134 | const int depth = |
135 | MatchingDim(input_shape, trailing_dim, output_shape, trailing_dim); |
136 | const int outer_size = |
137 | MatchingFlatSizeSkipDim(input_shape, trailing_dim, output_shape); |
138 | reference_integer_ops::L2Normalization(input->params.zero_point, outer_size, |
139 | depth, GetTensorData<int8>(input), |
140 | GetTensorData<int8>(output)); |
141 | } else { |
142 | TF_LITE_KERNEL_LOG(context, "Output type is %s, requires float." , |
143 | TfLiteTypeGetName(output->type)); |
144 | return kTfLiteError; |
145 | } |
146 | |
147 | return kTfLiteOk; |
148 | } |
149 | |
150 | } // namespace l2norm |
151 | |
152 | TfLiteRegistration* Register_L2NORM_REF() { |
153 | static TfLiteRegistration r = {nullptr, nullptr, l2norm::Prepare, |
154 | l2norm::Eval<l2norm::kReference>}; |
155 | return &r; |
156 | } |
157 | |
158 | TfLiteRegistration* Register_L2NORM_GENERIC_OPT() { |
159 | static TfLiteRegistration r = {nullptr, nullptr, l2norm::Prepare, |
160 | l2norm::Eval<l2norm::kGenericOptimized>}; |
161 | return &r; |
162 | } |
163 | |
164 | TfLiteRegistration* Register_L2_NORMALIZATION() { |
165 | return Register_L2NORM_GENERIC_OPT(); |
166 | } |
167 | |
168 | } // namespace builtin |
169 | } // namespace ops |
170 | } // namespace tflite |
171 | |