-
Notifications
You must be signed in to change notification settings - Fork 74.6k
/
Copy pathsparsity_format_converter.cc
392 lines (348 loc) · 14 KB
/
sparsity_format_converter.cc
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
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
/* Copyright 2020 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 "tensorflow/lite/kernels/internal/utils/sparsity_format_converter.h"
#include <cstdint>
#include <utility>
#include <vector>
namespace tflite {
namespace internal {
namespace sparsity {
namespace {
uint64_t GetFlattenedIndex(const std::vector<int>& indices,
const std::vector<int>& shape) {
uint64_t index = 0;
int sub_elements = 1;
for (int i = shape.size() - 1; i >= 0; i--) {
index += indices[i] * sub_elements;
sub_elements *= shape[i];
}
return index;
}
std::vector<int> TfLiteIntArrayToVector(const TfLiteIntArray* int_array) {
std::vector<int> values;
if (!int_array) {
return values;
}
values.resize(int_array->size);
for (size_t i = 0; i < int_array->size; i++) {
values[i] = int_array->data[i];
}
return values;
}
} // namespace
template <typename T>
FormatConverter<T>::FormatConverter(
const std::vector<int>& shape, const std::vector<int>& traversal_order,
const std::vector<TfLiteDimensionType>& format,
const std::vector<int>& block_size, const std::vector<int>& block_map)
: dense_shape_(shape),
traversal_order_(traversal_order),
block_size_(block_size),
block_map_(block_map) {
dense_size_ = 1;
int block_dim = 0;
blocked_shape_.resize(shape.size());
format_.resize(shape.size() + block_map.size());
for (int i = 0; i < shape.size(); i++) {
format_[i] = format[traversal_order[i]];
dense_size_ *= shape[i];
if (block_dim < block_map.size() && block_map[block_dim] == i) {
blocked_shape_[i] = shape[i] / block_size[block_dim];
block_dim++;
} else {
blocked_shape_[i] = shape[i];
}
}
// Only dense blocks are supported.
for (int i = 0; i < block_map.size(); i++) {
format_[i + shape.size()] = kTfLiteDimDense;
}
}
template <typename T>
TfLiteStatus FormatConverter<T>::DenseToSparse(const T* src_data) {
int num_original_dims = dense_shape_.size();
int num_block_dims = block_map_.size();
int num_expanded_dims = num_original_dims + num_block_dims;
std::vector<int> expanded_shape(num_expanded_dims);
for (int i = 0; i < num_expanded_dims; i++) {
if (i < num_original_dims) {
expanded_shape[i] = blocked_shape_[i];
} else {
expanded_shape[i] = block_size_[i - num_original_dims];
}
}
std::vector<int> shape_offset(num_original_dims);
shape_offset[shape_offset.size() - 1] = 1;
for (int i = num_original_dims - 1; i > 0; --i) {
shape_offset[i - 1] = shape_offset[i] * dense_shape_[i];
}
std::vector<int> expanded_shape_offset(num_expanded_dims);
for (int i = 0; i < num_original_dims; ++i) {
expanded_shape_offset[i] = shape_offset[i];
}
for (int i = 0; i < num_block_dims; ++i) {
int mapped_dim = block_map_[i];
expanded_shape_offset[num_original_dims + i] = shape_offset[mapped_dim];
expanded_shape_offset[mapped_dim] *= block_size_[i];
}
std::vector<int> dst_ordered_offset(num_expanded_dims);
for (int i = 0; i < num_expanded_dims; ++i) {
dst_ordered_offset[i] = expanded_shape_offset[traversal_order_[i]];
}
std::vector<bool> dst_dim_has_nonzeroes(num_expanded_dims);
std::fill(dst_dim_has_nonzeroes.begin(), dst_dim_has_nonzeroes.end(), false);
std::vector<int> inner_compressed_dim(num_expanded_dims);
int most_recent_compressed_dim = -1;
std::vector<int> num_segments_of_next_compressed_dim(num_expanded_dims);
int segment_count = 1;
for (int i = num_expanded_dims - 1; i >= 0; --i) {
inner_compressed_dim[i] = most_recent_compressed_dim;
if (format_[i] == kTfLiteDimSparseCSR) {
most_recent_compressed_dim = i;
num_segments_of_next_compressed_dim[i] = segment_count;
segment_count = 1;
} else {
num_segments_of_next_compressed_dim[i] = -1;
segment_count *= expanded_shape[traversal_order_[i]];
}
}
dim_metadata_.resize(num_expanded_dims * 2);
std::vector<int> dst_sparse_dims;
dst_sparse_dims.reserve(num_expanded_dims);
for (int i = 0; i < num_expanded_dims; ++i) {
dim_metadata_[i * 2].clear();
dim_metadata_[i * 2 + 1].clear();
if (format_[i] == kTfLiteDimDense) {
// If dimension is dense, just store the shape.
dim_metadata_[i * 2].push_back(expanded_shape[traversal_order_[i]]);
} else {
dim_metadata_[i * 2].push_back(0); // Segment array always begins with 0.
dst_sparse_dims.push_back(i); // Add dimension to the sparse list.
}
}
// This algorithm assumes that the block size is small enough for all the
// elements to fit in cache, so the strided accesses from different traversal
// order and the write-first-erase-later strategy shouldn't be too slow
int dst_dim_idx = num_expanded_dims;
std::vector<int> coordinate(num_expanded_dims, 0);
int dense_tensor_idx = 0;
while (dst_dim_idx >= 0) {
if (dst_dim_idx == num_expanded_dims) {
// We have a complete coordinate. Add the element to the value array if it
// is not zero, or if the last dimension is dense.
if (!IsZero(src_data[dense_tensor_idx])) {
data_.push_back(src_data[dense_tensor_idx]);
// Mark all sparse dimensions that their current indices have nonzeroes.
for (auto dst_dim : dst_sparse_dims) {
if (!dst_dim_has_nonzeroes[dst_dim]) {
// Only add the index to the indices array if the current nonzero
// is the first nonzero of the block.
dim_metadata_[2 * dst_dim + 1].push_back(coordinate[dst_dim]);
dst_dim_has_nonzeroes[dst_dim] = true;
}
}
} else if (format_[num_expanded_dims - 1] == kTfLiteDimDense) {
data_.push_back(src_data[dense_tensor_idx]);
}
--dst_dim_idx;
} else {
int original_dim_idx = traversal_order_[dst_dim_idx];
int dim_size = expanded_shape[original_dim_idx];
if (dst_dim_has_nonzeroes[dst_dim_idx]) {
// If the previous block has nonzeroes, reset the flag to false since
// we have just moved to a new block.
dst_dim_has_nonzeroes[dst_dim_idx] = false;
} else if (format_[dst_dim_idx] == kTfLiteDimSparseCSR) {
// This block is empty. Delete unnecessary values if compressed.
int next_compressed_dim = inner_compressed_dim[dst_dim_idx];
int erase_offset = dim_metadata_[2 * dst_dim_idx + 1].size() *
num_segments_of_next_compressed_dim[dst_dim_idx];
if (next_compressed_dim >= 0) {
auto& segments = dim_metadata_[2 * inner_compressed_dim[dst_dim_idx]];
segments.erase(segments.begin() + 1 + erase_offset, segments.end());
} else {
data_.erase(data_.begin() + erase_offset, data_.end());
}
}
if (++coordinate[dst_dim_idx] < dim_size) {
// The current dst_dim_idx is valid (not out of bound).
dense_tensor_idx += dst_ordered_offset[dst_dim_idx];
++dst_dim_idx;
} else {
// dst_dim_idx has reached its dim size. Update segment array and go
// back to incrementing the previous dimension (dst_dim_idx - 1).
if (format_[dst_dim_idx] == kTfLiteDimSparseCSR) {
dim_metadata_[2 * dst_dim_idx].push_back(
dim_metadata_[2 * dst_dim_idx + 1].size());
}
coordinate[dst_dim_idx] = -1;
dense_tensor_idx -= dst_ordered_offset[dst_dim_idx] * dim_size;
--dst_dim_idx;
}
}
}
return kTfLiteOk;
}
template <typename T>
FormatConverter<T>::FormatConverter(
const std::vector<int>& shape, const std::vector<int>& traversal_order,
const std::vector<TfLiteDimensionType>& format,
const std::vector<int>& dense_size,
const std::vector<std::vector<int>>& segments,
const std::vector<std::vector<int>>& indices,
const std::vector<int>& block_map) {
InitSparseToDenseConverter(shape, traversal_order, format, dense_size,
segments, indices, block_map);
}
template <typename T>
FormatConverter<T>::FormatConverter(const std::vector<int>& shape,
const TfLiteSparsity& sparsity) {
auto traversal_order = TfLiteIntArrayToVector(sparsity.traversal_order);
auto block_map = TfLiteIntArrayToVector(sparsity.block_map);
std::vector<TfLiteDimensionType> format(sparsity.dim_metadata_size);
std::vector<int> dense_size(sparsity.dim_metadata_size);
std::vector<std::vector<int>> segments(sparsity.dim_metadata_size);
std::vector<std::vector<int>> indices(sparsity.dim_metadata_size);
for (int i = 0; i < sparsity.dim_metadata_size; i++) {
format[i] = sparsity.dim_metadata[i].format;
dense_size[i] = sparsity.dim_metadata[i].dense_size;
segments[i] =
TfLiteIntArrayToVector(sparsity.dim_metadata[i].array_segments);
indices[i] = TfLiteIntArrayToVector(sparsity.dim_metadata[i].array_indices);
}
InitSparseToDenseConverter(shape, std::move(traversal_order),
std::move(format), std::move(dense_size),
std::move(segments), std::move(indices),
std::move(block_map));
}
template <typename T>
void FormatConverter<T>::InitSparseToDenseConverter(
std::vector<int> shape, std::vector<int> traversal_order,
std::vector<TfLiteDimensionType> format, std::vector<int> dense_size,
std::vector<std::vector<int>> segments,
std::vector<std::vector<int>> indices, std::vector<int> block_map) {
dense_shape_ = std::move(shape);
traversal_order_ = std::move(traversal_order);
block_map_ = std::move(block_map);
format_ = std::move(format);
dense_size_ = 1;
for (int i = 0; i < dense_shape_.size(); i++) {
dense_size_ *= dense_shape_[i];
}
dim_metadata_.resize(2 * format_.size());
for (int i = 0; i < format_.size(); i++) {
if (format_[i] == kTfLiteDimDense) {
dim_metadata_[2 * i] = {dense_size[i]};
} else {
dim_metadata_[2 * i] = std::move(segments[i]);
dim_metadata_[2 * i + 1] = std::move(indices[i]);
}
}
int original_rank = dense_shape_.size();
int block_dim = 0;
blocked_shape_.resize(original_rank);
block_size_.resize(block_map_.size());
for (int i = 0; i < original_rank; i++) {
if (block_dim < block_map_.size() && block_map_[block_dim] == i) {
int orig_dim = traversal_order_[original_rank + block_dim];
block_size_[block_dim] = dense_size[orig_dim];
blocked_shape_[i] = dense_shape_[i] / dense_size[orig_dim];
block_dim++;
} else {
blocked_shape_[i] = dense_shape_[i];
}
}
}
template <typename T>
void FormatConverter<T>::Populate(const T* src_data, std::vector<int> indices,
int level, int prev_idx, int* src_data_ptr,
T* dest_data) {
if (level == indices.size()) {
int orig_rank = dense_shape_.size();
std::vector<int> orig_idx;
orig_idx.resize(orig_rank);
int i = 0;
for (; i < orig_idx.size(); i++) {
int orig_dim = traversal_order_[i];
orig_idx[orig_dim] = indices[i];
}
for (; i < indices.size(); i++) {
const int block_idx = traversal_order_[i] - orig_rank;
const int orig_dim = block_map_[block_idx];
orig_idx[orig_dim] =
orig_idx[orig_dim] * block_size_[block_idx] + indices[i];
}
dest_data[GetFlattenedIndex(orig_idx, dense_shape_)] =
src_data[*src_data_ptr];
*src_data_ptr = *src_data_ptr + 1;
return;
}
const int metadata_idx = 2 * level;
const int shape_of_level = dim_metadata_[metadata_idx][0];
if (format_[level] == kTfLiteDimDense) {
for (int i = 0; i < shape_of_level; i++) {
indices[level] = i;
Populate(src_data, indices, level + 1, prev_idx * shape_of_level + i,
src_data_ptr, dest_data);
}
} else {
const auto& array_segments = dim_metadata_[metadata_idx];
const auto& array_indices = dim_metadata_[metadata_idx + 1];
for (int i = array_segments[prev_idx]; i < array_segments[prev_idx + 1];
i++) {
indices[level] = array_indices[i];
Populate(src_data, indices, level + 1, i, src_data_ptr, dest_data);
}
}
}
template <typename T>
TfLiteStatus FormatConverter<T>::SparseToDense(const T* src_data) {
data_.resize(dense_size_);
std::fill(data_.begin(), data_.end(), T(0));
int total_rank = traversal_order_.size();
int src_data_ptr = 0;
std::vector<int> indices(total_rank);
Populate(src_data, indices, 0, 0, &src_data_ptr, data_.data());
return kTfLiteOk;
}
template <typename T>
TfLiteStatus FormatConverter<T>::SparseToDense(const T* src_data,
const size_t dest_size,
T* dest_data,
TfLiteContext* context) {
if (dest_size != dense_size_) {
TF_LITE_MAYBE_KERNEL_LOG(
context, "unexpected buffer size for densified data, expected %lld.\n",
dense_size_);
return kTfLiteError;
}
// For types like Eigen::half, we cannot do a simple memset() with 0 values.
for (auto i = 0; i < dest_size; i++) {
dest_data[i] = T(0);
}
const int total_rank = traversal_order_.size();
int src_data_ptr = 0;
std::vector<int> indices(total_rank);
Populate(src_data, indices, 0, 0, &src_data_ptr, dest_data);
return kTfLiteOk;
}
template <typename T>
bool FormatConverter<T>::IsZero(const T val) {
return (val == static_cast<T>(0));
}
template class FormatConverter<int32_t>;
template class FormatConverter<int8_t>;
template class FormatConverter<float>;
template class FormatConverter<Eigen::half>;
} // namespace sparsity
} // namespace internal
} // namespace tflite