File size: 5,680 Bytes
25ade36
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
/* Copyright 2021 The TensorFlow Authors. All Rights Reserved.

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.
==============================================================================*/
#include "edge-impulse-sdk/tensorflow/lite/kernels/internal/reference/depth_to_space.h"

#include <stdint.h>

#include "edge-impulse-sdk/tensorflow/lite/c/common.h"
#include "edge-impulse-sdk/tensorflow/lite/kernels/internal/types.h"
#include "edge-impulse-sdk/tensorflow/lite/kernels/kernel_util.h"
#include "edge-impulse-sdk/tensorflow/lite/micro/kernels/kernel_util.h"
#include "edge-impulse-sdk/tensorflow/lite/micro/micro_log.h"

namespace tflite {
namespace {

constexpr int kInputTensor = 0;
constexpr int kOutputTensor = 0;

// input/output tensor shape rank associations
constexpr int kBatchRank = 0;
constexpr int kHeightRank = 1;
constexpr int kWidthRank = 2;
constexpr int kDepthRank = 3;

TfLiteStatus CalculateOpData(TfLiteContext* context, TfLiteNode* node) {
  auto* params =
      reinterpret_cast<TfLiteDepthToSpaceParams*>(node->builtin_data);

  TF_LITE_ENSURE_EQ(context, NumInputs(node), 1);
  TF_LITE_ENSURE_EQ(context, NumOutputs(node), 1);

  MicroContext* micro_context = GetMicroContext(context);

  TfLiteTensor* input =
      micro_context->AllocateTempInputTensor(node, kInputTensor);
  TF_LITE_ENSURE(context, input != nullptr);
  TfLiteTensor* output =
      micro_context->AllocateTempOutputTensor(node, kOutputTensor);
  TF_LITE_ENSURE(context, output != nullptr);

  TF_LITE_ENSURE_EQ(context, NumDimensions(input), 4);

  auto data_type = output->type;
  TF_LITE_ENSURE(context,
                 data_type == kTfLiteFloat32 || data_type == kTfLiteInt8);
  TF_LITE_ENSURE_TYPES_EQ(context, input->type, output->type);

  const int block_size = params->block_size;
  TF_LITE_ENSURE(context, block_size > 0);
  const int input_height = input->dims->data[kHeightRank];
  const int input_width = input->dims->data[kWidthRank];
  const int input_channels = input->dims->data[kDepthRank];
  int output_height = input_height * block_size;
  int output_width = input_width * block_size;
  int output_channels = input_channels / block_size / block_size;

  TF_LITE_ENSURE_EQ(context, input_height, output_height / block_size);
  TF_LITE_ENSURE_EQ(context, input_width, output_width / block_size);
  TF_LITE_ENSURE_EQ(context, input_channels,
                    output_channels * block_size * block_size);

  // We must update the output tensor dimensions.
  // The dims storage is expected to be the same area in memory
  // for both TfLiteTensor and TfLiteEvalTensor.  This is important
  // because TfLiteTensor in the MicroInterpreter is a temporary
  // allocation.  For the KernelRunner interpreter, TfLiteEvalTensor
  // is a temporary allocation.  We must therefore relocate the dims
  // from the FlatBuffer to the persistant storage arena.
  TfLiteEvalTensor* output_eval =
      tflite::micro::GetEvalOutput(context, node, kOutputTensor);
  TF_LITE_ENSURE_OK(context, tflite::micro::CreateWritableTensorDimsWithCopy(
                                 context, output, output_eval));
  output->dims->data[kBatchRank] = input->dims->data[kBatchRank];
  output->dims->data[kHeightRank] = output_height;
  output->dims->data[kWidthRank] = output_width;
  output->dims->data[kDepthRank] = output_channels;

  micro_context->DeallocateTempTfLiteTensor(input);
  micro_context->DeallocateTempTfLiteTensor(output);

  return kTfLiteOk;
}

TfLiteStatus Prepare(TfLiteContext* context, TfLiteNode* node) {
  return CalculateOpData(context, node);
}

TfLiteStatus Eval(TfLiteContext* context, TfLiteNode* node) {
  auto* params =
      reinterpret_cast<TfLiteDepthToSpaceParams*>(node->builtin_data);

  const TfLiteEvalTensor* input =
      tflite::micro::GetEvalInput(context, node, kInputTensor);
  TfLiteEvalTensor* output =
      tflite::micro::GetEvalOutput(context, node, kOutputTensor);

  tflite::DepthToSpaceParams op_params;
  op_params.block_size = static_cast<int32_t>(params->block_size);

  switch (input->type) {  // Already know in/out types are same.
    case kTfLiteFloat32:
      reference_ops::DepthToSpace(op_params,
                                  tflite::micro::GetTensorShape(input),
                                  tflite::micro::GetTensorData<float>(input),
                                  tflite::micro::GetTensorShape(output),
                                  tflite::micro::GetTensorData<float>(output));
      break;
    case kTfLiteInt8:
      reference_ops::DepthToSpace(op_params,
                                  tflite::micro::GetTensorShape(input),
                                  tflite::micro::GetTensorData<int8_t>(input),
                                  tflite::micro::GetTensorShape(output),
                                  tflite::micro::GetTensorData<int8_t>(output));
      break;
    default:
      MicroPrintf("DEPTH_TO_SPACE only supports FLOAT32 and INT8, got %s.",
                  TfLiteTypeGetName(output->type));
      return kTfLiteError;
  }

  return kTfLiteOk;
}

}  // namespace

TfLiteRegistration Register_DEPTH_TO_SPACE() {
  return tflite::micro::RegisterOp(nullptr, Prepare, Eval);
}

}  // namespace tflite