aboutsummaryrefslogtreecommitdiffhomepage
path: root/tensorflow/contrib/lite/kernels/reduce.cc
blob: 31c331a8c61ded203af9ff2ae127cb6f985e2932 (plain)
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
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
/* Copyright 2017 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 <string.h>
#include <vector>
#include "tensorflow/contrib/lite/builtin_op_data.h"
#include "tensorflow/contrib/lite/context.h"
#include "tensorflow/contrib/lite/kernels/internal/quantization_util.h"
#include "tensorflow/contrib/lite/kernels/internal/reference/reference_ops.h"
#include "tensorflow/contrib/lite/kernels/internal/tensor.h"
#include "tensorflow/contrib/lite/kernels/kernel_util.h"
#include "tensorflow/contrib/lite/kernels/op_macros.h"

namespace tflite {
namespace ops {
namespace builtin {
namespace reduce {

// This file has reference implementation of reduce_* operators.
enum KernelType {
  kReference,
};

struct OpContext {
  OpContext(TfLiteContext* context, TfLiteNode* node) {
    params = reinterpret_cast<TfLiteReducerParams*>(node->builtin_data);
    input = GetInput(context, node, 0);
    axis = GetInput(context, node, 1);
    output = GetOutput(context, node, 0);
  }
  TfLiteReducerParams* params;
  const TfLiteTensor* input;
  const TfLiteTensor* axis;
  TfLiteTensor* output;
};

void* Init(TfLiteContext* context, const char* buffer, size_t length) {
  // Creates two temp tensors to store index and axis for internal
  // implementation only.
  auto* scratch_tensor_index = new int;
  context->AddTensors(context, 3, scratch_tensor_index);
  return scratch_tensor_index;
}

void Free(TfLiteContext* context, void* buffer) {
  delete reinterpret_cast<int*>(buffer);
}

// Resizes the temp tensor that stores resolved axis.
TfLiteStatus ResizeTempAxis(TfLiteContext* context, OpContext* op_context,
                            TfLiteTensor* resolved_axis) {
  TfLiteIntArray* axis_size = TfLiteIntArrayCreate(1);
  axis_size->data[0] = static_cast<int>(NumElements(op_context->axis));
  return context->ResizeTensor(context, resolved_axis, axis_size);
}

// Resizes the temp tensor that stores temp sum of reduced elements.
TfLiteStatus ResizeTempSum(TfLiteContext* context, OpContext* op_context,
                           TfLiteTensor* temp_sum) {
  TfLiteIntArray* size = TfLiteIntArrayCreate(1);
  size->data[0] = static_cast<int>(NumElements(op_context->output));
  return context->ResizeTensor(context, temp_sum, size);
}

// Resizes output array based on the input size and resolved axis.
TfLiteStatus ResizeOutputTensor(TfLiteContext* context, OpContext* op_context) {
  size_t num_axis = NumElements(op_context->axis);
  const TfLiteIntArray* input_dims = op_context->input->dims;
  int input_num_dims = NumDimensions(op_context->input);
  const int* axis = GetTensorData<int>(op_context->axis);
  if (op_context->params->keep_dims) {
    TfLiteIntArray* output_dims = TfLiteIntArrayCreate(input_num_dims);
    for (int idx = 0; idx < input_num_dims; ++idx) {
      bool is_axis = false;
      for (int axis_idx = 0; axis_idx < num_axis; ++axis_idx) {
        if (axis[axis_idx] == idx || axis[axis_idx] + input_num_dims == idx) {
          is_axis = true;
          break;
        }
      }
      if (is_axis) {
        output_dims->data[idx] = 1;
      } else {
        output_dims->data[idx] = input_dims->data[idx];
      }
    }
    return context->ResizeTensor(context, op_context->output, output_dims);
  } else {
    // Calculates size of reducing axis.
    int num_reduce_axis = num_axis;
    for (int i = 0; i < num_axis; ++i) {
      int current = axis[i];
      if (current < 0) {
        current += input_num_dims;
      }
      TF_LITE_ENSURE(context, current >= 0 && current < input_num_dims);
      for (int j = 0; j < i; ++j) {
        int previous = axis[j];
        if (previous < 0) {
          previous += input_num_dims;
        }
        if (current == previous) {
          --num_reduce_axis;
          break;
        }
      }
    }
    // Determines output dimensions.
    TfLiteIntArray* output_dims =
        TfLiteIntArrayCreate(input_num_dims - num_reduce_axis);
    int num_skip_axis = 0;
    for (int idx = 0; idx < input_num_dims; ++idx) {
      bool is_axis = false;
      for (int axis_idx = 0; axis_idx < num_axis; ++axis_idx) {
        if (axis[axis_idx] == idx || axis[axis_idx] + input_num_dims == idx) {
          ++num_skip_axis;
          is_axis = true;
          break;
        }
      }
      if (!is_axis) {
        output_dims->data[idx - num_skip_axis] = input_dims->data[idx];
      }
    }
    return context->ResizeTensor(context, op_context->output, output_dims);
  }
}

// Initializes temp tensors to store index and resolved axis.
TfLiteStatus InitializeTemporaries(TfLiteContext* context, TfLiteNode* node,
                                   OpContext* op_context) {
  // Creates a temp index to iterate through input data.
  int* scratch_tensor_index = reinterpret_cast<int*>(node->user_data);
  TfLiteIntArrayFree(node->temporaries);
  node->temporaries = TfLiteIntArrayCreate(3);
  node->temporaries->data[0] = *scratch_tensor_index;
  TfLiteTensor* scratch_tensor = GetTemporary(context, node, /*index=*/0);
  scratch_tensor->type = kTfLiteInt32;
  scratch_tensor->allocation_type = kTfLiteArenaRw;
  TfLiteIntArray* index_size = TfLiteIntArrayCreate(1);
  index_size->data[0] = NumDimensions(op_context->input);
  TF_LITE_ENSURE_OK(context,
                    context->ResizeTensor(context, scratch_tensor, index_size));

  // Creates a temp tensor to store resolved axis given input data.
  node->temporaries->data[1] = *scratch_tensor_index + 1;
  TfLiteTensor* resolved_axis = GetTemporary(context, node, /*index=*/1);
  resolved_axis->type = kTfLiteInt32;
  // Creates a temp tensor to store temp sums when calculating mean.
  node->temporaries->data[2] = *scratch_tensor_index + 2;
  TfLiteTensor* temp_sum = GetTemporary(context, node, /*index=*/2);
  switch (op_context->input->type) {
    case kTfLiteFloat32:
      temp_sum->type = kTfLiteFloat32;
      break;
    case kTfLiteInt32:
      temp_sum->type = kTfLiteInt64;
      break;
    case kTfLiteInt64:
      temp_sum->type = kTfLiteInt64;
      break;
    case kTfLiteUInt8:
      temp_sum->type = kTfLiteInt32;
      break;
    default:
      return kTfLiteError;
  }
  return kTfLiteOk;
}

TfLiteStatus PrepareSimple(TfLiteContext* context, TfLiteNode* node) {
  TF_LITE_ENSURE_EQ(context, NumInputs(node), 2);
  TF_LITE_ENSURE_EQ(context, NumOutputs(node), 1);

  OpContext op_context(context, node);
  TF_LITE_ENSURE_OK(context, InitializeTemporaries(context, node, &op_context));

  TfLiteTensor* resolved_axis = GetTemporary(context, node, /*index=*/1);
  // Leaves work to Eval if axis is not constant; else resizes output.
  if (!IsConstantTensor(op_context.axis)) {
    SetTensorToDynamic(op_context.output);
    SetTensorToDynamic(resolved_axis);
    return kTfLiteOk;
  }
  resolved_axis->allocation_type = kTfLiteArenaRw;
  TF_LITE_ENSURE_OK(context,
                    ResizeTempAxis(context, &op_context, resolved_axis));
  TF_LITE_ENSURE_OK(context, ResizeOutputTensor(context, &op_context));
  return kTfLiteOk;
}

TfLiteStatus PrepareMean(TfLiteContext* context, TfLiteNode* node) {
  TF_LITE_ENSURE_OK(context, PrepareSimple(context, node));

  // reduce_mean requires a buffer to store intermediate sum result.
  OpContext op_context(context, node);
  TfLiteTensor* temp_sum = GetTemporary(context, node, /*index=*/2);
  if (!IsConstantTensor(op_context.axis)) {
    SetTensorToDynamic(temp_sum);
    return kTfLiteOk;
  }
  temp_sum->allocation_type = kTfLiteArenaRw;
  return ResizeTempSum(context, &op_context, temp_sum);
}

template <KernelType kernel_type>
TfLiteStatus EvalMean(TfLiteContext* context, TfLiteNode* node) {
  OpContext op_context(context, node);
  int num_axis = static_cast<int>(NumElements(op_context.axis));
  TfLiteTensor* temp_index = GetTemporary(context, node, /*index=*/0);
  TfLiteTensor* resolved_axis = GetTemporary(context, node, /*index=*/1);
  TfLiteTensor* temp_sum = GetTemporary(context, node, /*index=*/2);
  // Resize the output tensor if the output tensor is dynamic.
  if (IsDynamicTensor(op_context.output)) {
    TF_LITE_ENSURE_OK(context,
                      ResizeTempAxis(context, &op_context, resolved_axis));
    TF_LITE_ENSURE_OK(context, ResizeOutputTensor(context, &op_context));
    TF_LITE_ENSURE_OK(context, ResizeTempSum(context, &op_context, temp_sum));
  }

#define TF_LITE_MEAN(kernel_type, data_type, temp_data_type)        \
  kernel_type::Mean<>(                                              \
      GetTensorData<data_type>(op_context.input),                   \
      op_context.input->dims->data, op_context.input->dims->size,   \
      GetTensorData<data_type>(op_context.output),                  \
      op_context.output->dims->data, op_context.output->dims->size, \
      GetTensorData<int>(op_context.axis), num_axis,                \
      op_context.params->keep_dims, GetTensorData<int>(temp_index), \
      GetTensorData<int>(resolved_axis),                            \
      GetTensorData<temp_data_type>(temp_sum))

  if (kernel_type == kReference) {
    switch (op_context.input->type) {
      case kTfLiteFloat32:
        TF_LITE_ENSURE(context, TF_LITE_MEAN(reference_ops, float, float));
        break;
      case kTfLiteInt32:
        TF_LITE_ENSURE(context, TF_LITE_MEAN(reference_ops, int, int64_t));
        break;
      case kTfLiteInt64:
        TF_LITE_ENSURE(context, TF_LITE_MEAN(reference_ops, int64_t, int64_t));
        break;
      case kTfLiteUInt8:
        TF_LITE_ENSURE_EQ(context, op_context.input->params.scale,
                          op_context.output->params.scale);
        TF_LITE_ENSURE_EQ(context, op_context.input->params.zero_point,
                          op_context.output->params.zero_point);
        TF_LITE_ENSURE(context, TF_LITE_MEAN(reference_ops, uint8_t, int));
        break;
      default:
        return kTfLiteError;
    }
  }
#undef TF_LITE_MEAN
  return kTfLiteOk;
}

template <KernelType kernel_type>
TfLiteStatus EvalSum(TfLiteContext* context, TfLiteNode* node) {
  OpContext op_context(context, node);
  int num_axis = static_cast<int>(NumElements(op_context.axis));
  TfLiteTensor* temp_index = GetTemporary(context, node, /*index=*/0);
  TfLiteTensor* resolved_axis = GetTemporary(context, node, /*index=*/1);
  // Resize the output tensor if the output tensor is dynamic.
  if (IsDynamicTensor(op_context.output)) {
    TF_LITE_ENSURE_OK(context,
                      ResizeTempAxis(context, &op_context, resolved_axis));
    TF_LITE_ENSURE_OK(context, ResizeOutputTensor(context, &op_context));
  }

#define TF_LITE_SUM(kernel_type, data_type)                         \
  kernel_type::Sum<>(                                               \
      GetTensorData<data_type>(op_context.input),                   \
      op_context.input->dims->data, op_context.input->dims->size,   \
      GetTensorData<data_type>(op_context.output),                  \
      op_context.output->dims->data, op_context.output->dims->size, \
      GetTensorData<int>(op_context.axis), num_axis,                \
      op_context.params->keep_dims, GetTensorData<int>(temp_index), \
      GetTensorData<int>(resolved_axis))

  if (kernel_type == kReference) {
    switch (op_context.input->type) {
      case kTfLiteFloat32:
        TF_LITE_ENSURE(context, TF_LITE_SUM(reference_ops, float));
        break;
      case kTfLiteInt32:
        TF_LITE_ENSURE(context, TF_LITE_SUM(reference_ops, int));
        break;
      case kTfLiteInt64:
        TF_LITE_ENSURE(context, TF_LITE_SUM(reference_ops, int64_t));
        break;
      case kTfLiteUInt8:
        TF_LITE_ENSURE_EQ(context, op_context.input->params.scale,
                          op_context.output->params.scale);
        TF_LITE_ENSURE_EQ(context, op_context.input->params.zero_point,
                          op_context.output->params.zero_point);
        TF_LITE_ENSURE(context, TF_LITE_SUM(reference_ops, uint8_t));
        break;
      default:
        return kTfLiteError;
    }
  }
#undef TF_LITE_SUM
  return kTfLiteOk;
}

}  // namespace reduce

TfLiteRegistration* Register_MEAN_REF() {
  static TfLiteRegistration r = {reduce::Init, reduce::Free,
                                 reduce::PrepareMean,
                                 reduce::EvalMean<reduce::kReference>};
  return &r;
}

TfLiteRegistration* Register_SUM_REF() {
  static TfLiteRegistration r = {reduce::Init, reduce::Free,
                                 reduce::PrepareSimple,
                                 reduce::EvalSum<reduce::kReference>};
  return &r;
}

// TODO(kanlig): add optimized implementation of Mean.
TfLiteRegistration* Register_MEAN() { return Register_MEAN_REF(); }
TfLiteRegistration* Register_SUM() { return Register_SUM_REF(); }

}  // namespace builtin
}  // namespace ops
}  // namespace tflite