-
Notifications
You must be signed in to change notification settings - Fork 916
/
reader_impl.cu
1809 lines (1619 loc) · 75.1 KB
/
reader_impl.cu
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
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
726
727
728
729
730
731
732
733
734
735
736
737
738
739
740
741
742
743
744
745
746
747
748
749
750
751
752
753
754
755
756
757
758
759
760
761
762
763
764
765
766
767
768
769
770
771
772
773
774
775
776
777
778
779
780
781
782
783
784
785
786
787
788
789
790
791
792
793
794
795
796
797
798
799
800
801
802
803
804
805
806
807
808
809
810
811
812
813
814
815
816
817
818
819
820
821
822
823
824
825
826
827
828
829
830
831
832
833
834
835
836
837
838
839
840
841
842
843
844
845
846
847
848
849
850
851
852
853
854
855
856
857
858
859
860
861
862
863
864
865
866
867
868
869
870
871
872
873
874
875
876
877
878
879
880
881
882
883
884
885
886
887
888
889
890
891
892
893
894
895
896
897
898
899
900
901
902
903
904
905
906
907
908
909
910
911
912
913
914
915
916
917
918
919
920
921
922
923
924
925
926
927
928
929
930
931
932
933
934
935
936
937
938
939
940
941
942
943
944
945
946
947
948
949
950
951
952
953
954
955
956
957
958
959
960
961
962
963
964
965
966
967
968
969
970
971
972
973
974
975
976
977
978
979
980
981
982
983
984
985
986
987
988
989
990
991
992
993
994
995
996
997
998
999
1000
/*
* Copyright (c) 2019-2021, NVIDIA CORPORATION.
*
* 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.
*/
/**
* @file reader_impl.cu
* @brief cuDF-IO Parquet reader class implementation
*/
#include "reader_impl.hpp"
#include <io/comp/gpuinflate.h>
#include <cudf/detail/utilities/vector_factories.hpp>
#include <cudf/table/table.hpp>
#include <cudf/utilities/error.hpp>
#include <cudf/utilities/traits.hpp>
#include <rmm/cuda_stream_view.hpp>
#include <rmm/device_buffer.hpp>
#include <rmm/device_uvector.hpp>
#include <rmm/exec_policy.hpp>
#include <nvcomp/snappy.h>
#include <algorithm>
#include <array>
#include <numeric>
#include <regex>
namespace cudf {
namespace io {
namespace detail {
namespace parquet {
// Import functionality that's independent of legacy code
using namespace cudf::io::parquet;
using namespace cudf::io;
// bit space we are reserving in column_buffer::user_data
constexpr uint32_t PARQUET_COLUMN_BUFFER_SCHEMA_MASK = (0xffffff);
constexpr uint32_t PARQUET_COLUMN_BUFFER_FLAG_LIST_TERMINATED = (1 << 24);
namespace {
parquet::ConvertedType logical_type_to_converted_type(parquet::LogicalType const& logical)
{
if (logical.isset.STRING) {
return parquet::UTF8;
} else if (logical.isset.MAP) {
return parquet::MAP;
} else if (logical.isset.LIST) {
return parquet::LIST;
} else if (logical.isset.ENUM) {
return parquet::ENUM;
} else if (logical.isset.DECIMAL) {
return parquet::DECIMAL; // TODO set decimal values
} else if (logical.isset.DATE) {
return parquet::DATE;
} else if (logical.isset.TIME) {
if (logical.TIME.unit.isset.MILLIS)
return parquet::TIME_MILLIS;
else if (logical.TIME.unit.isset.MICROS)
return parquet::TIME_MICROS;
} else if (logical.isset.TIMESTAMP) {
if (logical.TIMESTAMP.unit.isset.MILLIS)
return parquet::TIMESTAMP_MILLIS;
else if (logical.TIMESTAMP.unit.isset.MICROS)
return parquet::TIMESTAMP_MICROS;
} else if (logical.isset.INTEGER) {
switch (logical.INTEGER.bitWidth) {
case 8: return logical.INTEGER.isSigned ? INT_8 : UINT_8;
case 16: return logical.INTEGER.isSigned ? INT_16 : UINT_16;
case 32: return logical.INTEGER.isSigned ? INT_32 : UINT_32;
case 64: return logical.INTEGER.isSigned ? INT_64 : UINT_64;
default: break;
}
} else if (logical.isset.UNKNOWN) {
return parquet::NA;
} else if (logical.isset.JSON) {
return parquet::JSON;
} else if (logical.isset.BSON) {
return parquet::BSON;
}
return parquet::UNKNOWN;
}
/**
* @brief Function that translates Parquet datatype to cuDF type enum
*/
type_id to_type_id(SchemaElement const& schema,
bool strings_to_categorical,
type_id timestamp_type_id,
bool strict_decimal_types)
{
parquet::Type physical = schema.type;
parquet::ConvertedType converted_type = schema.converted_type;
int32_t decimal_scale = schema.decimal_scale;
// Logical type used for actual data interpretation; the legacy converted type
// is superceded by 'logical' type whenever available.
auto inferred_converted_type = logical_type_to_converted_type(schema.logical_type);
if (inferred_converted_type != parquet::UNKNOWN) converted_type = inferred_converted_type;
if (inferred_converted_type == parquet::DECIMAL && decimal_scale == 0)
decimal_scale = schema.logical_type.DECIMAL.scale;
switch (converted_type) {
case parquet::UINT_8: return type_id::UINT8;
case parquet::INT_8: return type_id::INT8;
case parquet::UINT_16: return type_id::UINT16;
case parquet::INT_16: return type_id::INT16;
case parquet::UINT_32: return type_id::UINT32;
case parquet::UINT_64: return type_id::UINT64;
case parquet::DATE: return type_id::TIMESTAMP_DAYS;
case parquet::TIME_MILLIS:
return (timestamp_type_id != type_id::EMPTY) ? timestamp_type_id
: type_id::DURATION_MILLISECONDS;
case parquet::TIME_MICROS:
return (timestamp_type_id != type_id::EMPTY) ? timestamp_type_id
: type_id::DURATION_MICROSECONDS;
case parquet::TIMESTAMP_MICROS:
return (timestamp_type_id != type_id::EMPTY) ? timestamp_type_id
: type_id::TIMESTAMP_MICROSECONDS;
case parquet::TIMESTAMP_MILLIS:
return (timestamp_type_id != type_id::EMPTY) ? timestamp_type_id
: type_id::TIMESTAMP_MILLISECONDS;
case parquet::DECIMAL:
if (physical == parquet::INT32)
return type_id::DECIMAL32;
else if (physical == parquet::INT64)
return type_id::DECIMAL64;
else if (physical == parquet::FIXED_LEN_BYTE_ARRAY && schema.type_length <= 8) {
return type_id::DECIMAL64;
} else {
CUDF_EXPECTS(strict_decimal_types == false, "Unsupported decimal type read!");
return type_id::FLOAT64;
}
break;
// maps are just List<Struct<>>.
case parquet::MAP:
case parquet::LIST: return type_id::LIST;
case parquet::NA: return type_id::STRING;
// return type_id::EMPTY; //TODO(kn): enable after Null/Empty column support
default: break;
}
// is it simply a struct?
if (schema.is_struct()) { return type_id::STRUCT; }
// Physical storage type supported by Parquet; controls the on-disk storage
// format in combination with the encoding type.
switch (physical) {
case parquet::BOOLEAN: return type_id::BOOL8;
case parquet::INT32: return type_id::INT32;
case parquet::INT64: return type_id::INT64;
case parquet::FLOAT: return type_id::FLOAT32;
case parquet::DOUBLE: return type_id::FLOAT64;
case parquet::BYTE_ARRAY:
case parquet::FIXED_LEN_BYTE_ARRAY:
// Can be mapped to INT32 (32-bit hash) or STRING
return strings_to_categorical ? type_id::INT32 : type_id::STRING;
case parquet::INT96:
return (timestamp_type_id != type_id::EMPTY) ? timestamp_type_id
: type_id::TIMESTAMP_NANOSECONDS;
default: break;
}
return type_id::EMPTY;
}
/**
* @brief Function that translates cuDF time unit to Parquet clock frequency
*/
constexpr int32_t to_clockrate(type_id timestamp_type_id)
{
switch (timestamp_type_id) {
case type_id::DURATION_SECONDS: return 1;
case type_id::DURATION_MILLISECONDS: return 1000;
case type_id::DURATION_MICROSECONDS: return 1000000;
case type_id::DURATION_NANOSECONDS: return 1000000000;
case type_id::TIMESTAMP_SECONDS: return 1;
case type_id::TIMESTAMP_MILLISECONDS: return 1000;
case type_id::TIMESTAMP_MICROSECONDS: return 1000000;
case type_id::TIMESTAMP_NANOSECONDS: return 1000000000;
default: return 0;
}
}
/**
* @brief Function that returns the required the number of bits to store a value
*/
template <typename T = uint8_t>
T required_bits(uint32_t max_level)
{
return static_cast<T>(CompactProtocolReader::NumRequiredBits(max_level));
}
std::tuple<int32_t, int32_t, int8_t> conversion_info(type_id column_type_id,
type_id timestamp_type_id,
parquet::Type physical,
int8_t converted,
int32_t length)
{
int32_t type_width = (physical == parquet::FIXED_LEN_BYTE_ARRAY) ? length : 0;
int32_t clock_rate = 0;
if (column_type_id == type_id::INT8 or column_type_id == type_id::UINT8) {
type_width = 1; // I32 -> I8
} else if (column_type_id == type_id::INT16 or column_type_id == type_id::UINT16) {
type_width = 2; // I32 -> I16
} else if (column_type_id == type_id::INT32) {
type_width = 4; // str -> hash32
} else if (is_chrono(data_type{column_type_id})) {
clock_rate = to_clockrate(timestamp_type_id);
}
int8_t converted_type = converted;
if (converted_type == parquet::DECIMAL && column_type_id != type_id::FLOAT64 &&
column_type_id != type_id::DECIMAL32 && column_type_id != type_id::DECIMAL64) {
converted_type = parquet::UNKNOWN; // Not converting to float64 or decimal
}
return std::make_tuple(type_width, clock_rate, converted_type);
}
} // namespace
std::string name_from_path(const std::vector<std::string>& path_in_schema)
{
// For the case of lists, we will see a schema that looks like:
// a.list.element.list.element
// where each (list.item) pair represents a level of nesting. According to the parquet spec,
// https://github.com/apache/parquet-format/blob/master/LogicalTypes.md
// the initial field must be named "list" and the inner element must be named "element".
// If we are dealing with a list, we want to return the topmost name of the group ("a").
//
// For other nested schemas, like structs we just want to return the bottom-most name. For
// example a struct with the schema
// b.employee.id, the column representing "id" should simply be named "id".
//
// In short, this means : return the highest level of the schema that does not have list
// definitions underneath it.
//
std::string s = (path_in_schema.size() > 0) ? path_in_schema[0] : "";
for (size_t i = 1; i < path_in_schema.size(); i++) {
// The Parquet spec requires that the outer schema field is named "list". However it also
// provides a list of backwards compatibility cases that are applicable as well. Currently
// we are only handling the formal spec. This will get cleaned up and improved when we add
// support for structs. The correct thing to do will probably be to examine the type of
// the SchemaElement itself to concretely identify the start of a nested type of any kind rather
// than trying to derive it from the path string.
if (path_in_schema[i] == "list") {
// Again, strictly speaking, the Parquet spec says the inner field should be named
// "element", but there are some backwards compatibility issues that we have seen in the
// wild. For example, Pandas calls the field "item". We will allow any name for now.
i++;
continue;
}
// otherwise, we've got a real nested column. update the name
s = path_in_schema[i];
}
return s;
}
/**
* @brief Class for parsing dataset metadata
*/
struct metadata : public FileMetaData {
explicit metadata(datasource* source)
{
constexpr auto header_len = sizeof(file_header_s);
constexpr auto ender_len = sizeof(file_ender_s);
const auto len = source->size();
const auto header_buffer = source->host_read(0, header_len);
const auto header = reinterpret_cast<const file_header_s*>(header_buffer->data());
const auto ender_buffer = source->host_read(len - ender_len, ender_len);
const auto ender = reinterpret_cast<const file_ender_s*>(ender_buffer->data());
CUDF_EXPECTS(len > header_len + ender_len, "Incorrect data source");
CUDF_EXPECTS(header->magic == parquet_magic && ender->magic == parquet_magic,
"Corrupted header or footer");
CUDF_EXPECTS(ender->footer_len != 0 && ender->footer_len <= (len - header_len - ender_len),
"Incorrect footer length");
const auto buffer = source->host_read(len - ender->footer_len - ender_len, ender->footer_len);
CompactProtocolReader cp(buffer->data(), ender->footer_len);
CUDF_EXPECTS(cp.read(this), "Cannot parse metadata");
CUDF_EXPECTS(cp.InitSchema(this), "Cannot initialize schema");
}
};
class aggregate_metadata {
std::vector<metadata> const per_file_metadata;
std::map<std::string, std::string> const agg_keyval_map;
size_type const num_rows;
size_type const num_row_groups;
/**
* @brief Create a metadata object from each element in the source vector
*/
auto metadatas_from_sources(std::vector<std::unique_ptr<datasource>> const& sources)
{
std::vector<metadata> metadatas;
std::transform(
sources.cbegin(), sources.cend(), std::back_inserter(metadatas), [](auto const& source) {
return metadata(source.get());
});
return metadatas;
}
/**
* @brief Merge the keyvalue maps from each per-file metadata object into a single map.
*/
auto merge_keyval_metadata()
{
std::map<std::string, std::string> merged;
// merge key/value maps TODO: warn/throw if there are mismatches?
for (auto const& pfm : per_file_metadata) {
for (auto const& kv : pfm.key_value_metadata) {
merged[kv.key] = kv.value;
}
}
return merged;
}
/**
* @brief Sums up the number of rows of each source
*/
size_type calc_num_rows() const
{
return std::accumulate(
per_file_metadata.begin(), per_file_metadata.end(), 0, [](auto& sum, auto& pfm) {
return sum + pfm.num_rows;
});
}
/**
* @brief Sums up the number of row groups of each source
*/
size_type calc_num_row_groups() const
{
return std::accumulate(
per_file_metadata.begin(), per_file_metadata.end(), 0, [](auto& sum, auto& pfm) {
return sum + pfm.row_groups.size();
});
}
public:
aggregate_metadata(std::vector<std::unique_ptr<datasource>> const& sources)
: per_file_metadata(metadatas_from_sources(sources)),
agg_keyval_map(merge_keyval_metadata()),
num_rows(calc_num_rows()),
num_row_groups(calc_num_row_groups())
{
// Verify that the input files have matching numbers of columns
size_type num_cols = -1;
for (auto const& pfm : per_file_metadata) {
if (pfm.row_groups.size() != 0) {
if (num_cols == -1)
num_cols = pfm.row_groups[0].columns.size();
else
CUDF_EXPECTS(num_cols == static_cast<size_type>(pfm.row_groups[0].columns.size()),
"All sources must have the same number of columns");
}
}
// Verify that the input files have matching schemas
for (auto const& pfm : per_file_metadata) {
CUDF_EXPECTS(per_file_metadata[0].schema == pfm.schema,
"All sources must have the same schemas");
}
}
auto const& get_row_group(size_type row_group_index, size_type src_idx) const
{
CUDF_EXPECTS(src_idx >= 0 && src_idx < static_cast<size_type>(per_file_metadata.size()),
"invalid source index");
return per_file_metadata[src_idx].row_groups[row_group_index];
}
auto const& get_column_metadata(size_type row_group_index,
size_type src_idx,
int schema_idx) const
{
auto col = std::find_if(
per_file_metadata[src_idx].row_groups[row_group_index].columns.begin(),
per_file_metadata[src_idx].row_groups[row_group_index].columns.end(),
[schema_idx](ColumnChunk const& col) { return col.schema_idx == schema_idx ? true : false; });
CUDF_EXPECTS(col != std::end(per_file_metadata[src_idx].row_groups[row_group_index].columns),
"Found no metadata for schema index");
return col->meta_data;
}
auto get_num_rows() const { return num_rows; }
auto get_num_row_groups() const { return num_row_groups; }
auto const& get_schema(int schema_idx) const { return per_file_metadata[0].schema[schema_idx]; }
auto const& get_key_value_metadata() const { return agg_keyval_map; }
/**
* @brief Gets the concrete nesting depth of output cudf columns
*
* @param schema_index Schema index of the input column
*
* @return comma-separated index column names in quotes
*/
inline int get_output_nesting_depth(int schema_index) const
{
auto& pfm = per_file_metadata[0];
int depth = 0;
// walk upwards, skipping repeated fields
while (schema_index > 0) {
if (!pfm.schema[schema_index].is_stub()) { depth++; }
schema_index = pfm.schema[schema_index].parent_idx;
}
return depth;
}
/**
* @brief Extracts the pandas "index_columns" section
*
* PANDAS adds its own metadata to the key_value section when writing out the
* dataframe to a file to aid in exact reconstruction. The JSON-formatted
* metadata contains the index column(s) and PANDA-specific datatypes.
*
* @return comma-separated index column names in quotes
*/
std::string get_pandas_index() const
{
auto it = agg_keyval_map.find("pandas");
if (it != agg_keyval_map.end()) {
// Captures a list of quoted strings found inside square brackets after `"index_columns":`
// Inside quotes supports newlines, brackets, escaped quotes, etc.
// One-liner regex:
// "index_columns"\s*:\s*\[\s*((?:"(?:|(?:.*?(?![^\\]")).?)[^\\]?",?\s*)*)\]
// Documented below.
std::regex index_columns_expr{
R"("index_columns"\s*:\s*\[\s*)" // match preamble, opening square bracket, whitespace
R"(()" // Open first capturing group
R"((?:")" // Open non-capturing group match opening quote
R"((?:|(?:.*?(?![^\\]")).?))" // match empty string or anything between quotes
R"([^\\]?")" // Match closing non-escaped quote
R"(,?\s*)" // Match optional comma and whitespace
R"()*)" // Close non-capturing group and repeat 0 or more times
R"())" // Close first capturing group
R"(\])" // Match closing square brackets
};
std::smatch sm;
if (std::regex_search(it->second, sm, index_columns_expr)) { return sm[1].str(); }
}
return "";
}
/**
* @brief Extracts the column name(s) used for the row indexes in a dataframe
*
* @param names List of column names to load, where index column name(s) will be added
*/
std::vector<std::string> get_pandas_index_names() const
{
std::vector<std::string> names;
auto str = get_pandas_index();
if (str.length() != 0) {
std::regex index_name_expr{R"(\"((?:\\.|[^\"])*)\")"};
std::smatch sm;
while (std::regex_search(str, sm, index_name_expr)) {
if (sm.size() == 2) { // 2 = whole match, first item
if (std::find(names.begin(), names.end(), sm[1].str()) == names.end()) {
std::regex esc_quote{R"(\\")"};
names.emplace_back(std::regex_replace(sm[1].str(), esc_quote, R"(")"));
}
}
str = sm.suffix();
}
}
return names;
}
struct row_group_info {
size_type const index;
size_t const start_row; // TODO source index
size_type const source_index;
row_group_info(size_type index, size_t start_row, size_type source_index)
: index(index), start_row(start_row), source_index(source_index)
{
}
};
/**
* @brief Filters and reduces down to a selection of row groups
*
* @param row_groups Lists of row group to reads, one per source
* @param row_start Starting row of the selection
* @param row_count Total number of rows selected
*
* @return List of row group indexes and its starting row
*/
auto select_row_groups(std::vector<std::vector<size_type>> const& row_groups,
size_type& row_start,
size_type& row_count) const
{
if (!row_groups.empty()) {
std::vector<row_group_info> selection;
CUDF_EXPECTS(row_groups.size() == per_file_metadata.size(),
"Must specify row groups for each source");
row_count = 0;
for (size_t src_idx = 0; src_idx < row_groups.size(); ++src_idx) {
for (auto const& rowgroup_idx : row_groups[src_idx]) {
CUDF_EXPECTS(
rowgroup_idx >= 0 &&
rowgroup_idx < static_cast<size_type>(per_file_metadata[src_idx].row_groups.size()),
"Invalid rowgroup index");
selection.emplace_back(rowgroup_idx, row_count, src_idx);
row_count += get_row_group(rowgroup_idx, src_idx).num_rows;
}
}
return selection;
}
row_start = std::max(row_start, 0);
if (row_count < 0) {
row_count = static_cast<size_type>(
std::min<int64_t>(get_num_rows(), std::numeric_limits<size_type>::max()));
}
row_count = min(row_count, get_num_rows() - row_start);
CUDF_EXPECTS(row_count >= 0, "Invalid row count");
CUDF_EXPECTS(row_start <= get_num_rows(), "Invalid row start");
std::vector<row_group_info> selection;
size_type count = 0;
for (size_t src_idx = 0; src_idx < per_file_metadata.size(); ++src_idx) {
for (size_t rg_idx = 0; rg_idx < per_file_metadata[src_idx].row_groups.size(); ++rg_idx) {
auto const chunk_start_row = count;
count += get_row_group(rg_idx, src_idx).num_rows;
if (count > row_start || count == 0) {
selection.emplace_back(rg_idx, chunk_start_row, src_idx);
}
if (count >= row_start + row_count) { break; }
}
}
return selection;
}
/**
* @brief Filters and reduces down to a selection of columns
*
* @param use_names List of paths of column names to select
* @param include_index Whether to always include the PANDAS index column(s)
* @param strings_to_categorical Type conversion parameter
* @param timestamp_type_id Type conversion parameter
* @param strict_decimal_types Type conversion parameter
*
* @return input column information, output column information, list of output column schema
* indices
*/
auto select_columns(std::vector<std::string> const& use_names,
bool include_index,
bool strings_to_categorical,
type_id timestamp_type_id,
bool strict_decimal_types) const
{
auto find_schema_child = [&](SchemaElement const& schema_elem, std::string const& name) {
auto const& col_schema_idx = std::find_if(
schema_elem.children_idx.cbegin(),
schema_elem.children_idx.cend(),
[&](size_t col_schema_idx) { return get_schema(col_schema_idx).name == name; });
return (col_schema_idx != schema_elem.children_idx.end()) ? static_cast<int>(*col_schema_idx)
: -1;
};
std::vector<column_buffer> output_columns;
std::vector<input_column_info> input_columns;
std::vector<int> nesting;
// Return true if column path is valid. e.g. if the path is {"struct1", "child1"}, then it is
// valid if "struct1.child1" exists in this file's schema. If "struct1" exists but "child1" is
// not a child of "struct1" then the function will return false for "struct1"
std::function<bool(column_name_info const*, int, std::vector<column_buffer>&)> build_column =
[&](column_name_info const* col_name_info,
int schema_idx,
std::vector<column_buffer>& out_col_array) {
if (schema_idx < 0) { return false; }
auto const& schema_elem = get_schema(schema_idx);
// if schema_elem is a stub then it does not exist in the column_name_info and column_buffer
// hierarchy. So continue on
if (schema_elem.is_stub()) {
// is this legit?
CUDF_EXPECTS(schema_elem.num_children == 1, "Unexpected number of children for stub");
auto child_col_name_info = (col_name_info) ? &col_name_info->children[0] : nullptr;
return build_column(child_col_name_info, schema_elem.children_idx[0], out_col_array);
}
// if we're at the root, this is a new output column
auto const col_type =
to_type_id(schema_elem, strings_to_categorical, timestamp_type_id, strict_decimal_types);
auto const dtype = col_type == type_id::DECIMAL32 || col_type == type_id::DECIMAL64
? data_type{col_type, numeric::scale_type{-schema_elem.decimal_scale}}
: data_type{col_type};
column_buffer output_col(dtype, schema_elem.repetition_type == OPTIONAL);
// store the index of this element if inserted in out_col_array
nesting.push_back(static_cast<int>(out_col_array.size()));
output_col.name = schema_elem.name;
// build each child
bool path_is_valid = false;
if (col_name_info == nullptr or col_name_info->children.empty()) {
// add all children of schema_elem.
// At this point, we can no longer pass a col_name_info to build_column
for (int idx = 0; idx < schema_elem.num_children; idx++) {
path_is_valid |=
build_column(nullptr, schema_elem.children_idx[idx], output_col.children);
}
} else {
for (size_t idx = 0; idx < col_name_info->children.size(); idx++) {
path_is_valid |=
build_column(&col_name_info->children[idx],
find_schema_child(schema_elem, col_name_info->children[idx].name),
output_col.children);
}
}
// if I have no children, we're at a leaf and I'm an input column (that is, one with actual
// data stored) so add me to the list.
if (schema_elem.num_children == 0) {
input_column_info& input_col =
input_columns.emplace_back(input_column_info{schema_idx, schema_elem.name});
std::copy(nesting.cbegin(), nesting.cend(), std::back_inserter(input_col.nesting));
path_is_valid = true; // If we're able to reach leaf then path is valid
}
if (path_is_valid) { out_col_array.push_back(std::move(output_col)); }
nesting.pop_back();
return path_is_valid;
};
std::vector<int> output_column_schemas;
//
// there is not necessarily a 1:1 mapping between input columns and output columns.
// For example, parquet does not explicitly store a ColumnChunkDesc for struct columns.
// The "structiness" is simply implied by the schema. For example, this schema:
// required group field_id=1 name {
// required binary field_id=2 firstname (String);
// required binary field_id=3 middlename (String);
// required binary field_id=4 lastname (String);
// }
// will only contain 3 internal columns of data (firstname, middlename, lastname). But of
// course "name" is ultimately the struct column we want to return.
//
// "firstname", "middlename" and "lastname" represent the input columns in the file that we
// process to produce the final cudf "name" column.
//
// A user can ask for a single field out of the struct e.g. firstname.
// In this case they'll pass a fully qualified name to the schema element like
// ["name", "firstname"]
//
auto const& root = get_schema(0);
if (use_names.empty()) {
for (auto const& schema_idx : root.children_idx) {
build_column(nullptr, schema_idx, output_columns);
output_column_schemas.push_back(schema_idx);
}
} else {
struct path_info {
std::string full_path;
int schema_idx;
};
// Convert schema into a vector of every possible path
std::vector<path_info> all_paths;
std::function<void(std::string, int)> add_path = [&](std::string path_till_now,
int schema_idx) {
auto const& schema_elem = get_schema(schema_idx);
std::string curr_path = path_till_now + schema_elem.name;
all_paths.push_back({curr_path, schema_idx});
for (auto const& child_idx : schema_elem.children_idx) {
add_path(curr_path + ".", child_idx);
}
};
for (auto const& child_idx : get_schema(0).children_idx) {
add_path("", child_idx);
}
// Find which of the selected paths are valid and get their schema index
std::vector<path_info> valid_selected_paths;
for (auto const& selected_path : use_names) {
auto found_path =
std::find_if(all_paths.begin(), all_paths.end(), [&](path_info& valid_path) {
return valid_path.full_path == selected_path;
});
if (found_path != all_paths.end()) {
valid_selected_paths.push_back({selected_path, found_path->schema_idx});
}
}
// Now construct paths as vector of strings for further consumption
std::vector<std::vector<std::string>> use_names3;
std::transform(valid_selected_paths.begin(),
valid_selected_paths.end(),
std::back_inserter(use_names3),
[&](path_info const& valid_path) {
auto schema_idx = valid_path.schema_idx;
std::vector<std::string> result_path;
do {
SchemaElement const& elem = get_schema(schema_idx);
result_path.push_back(elem.name);
schema_idx = elem.parent_idx;
} while (schema_idx > 0);
return std::vector<std::string>(result_path.rbegin(), result_path.rend());
});
std::vector<column_name_info> selected_columns;
if (include_index) {
std::vector<std::string> index_names = get_pandas_index_names();
std::transform(index_names.cbegin(),
index_names.cend(),
std::back_inserter(selected_columns),
[](std::string const& name) { return column_name_info(name); });
}
// Merge the vector use_names into a set of hierarchical column_name_info objects
/* This is because if we have columns like this:
* col1
* / \
* s3 f4
* / \
* f5 f6
*
* there may be common paths in use_names like:
* {"col1", "s3", "f5"}, {"col1", "f4"}
* which means we want the output to contain
* col1
* / \
* s3 f4
* /
* f5
*
* rather than
* col1 col1
* | |
* s3 f4
* |
* f5
*/
for (auto const& path : use_names3) {
auto array_to_find_in = &selected_columns;
for (size_t depth = 0; depth < path.size(); ++depth) {
// Check if the path exists in our selected_columns and if not, add it.
auto const& name_to_find = path[depth];
auto found_col = std::find_if(
array_to_find_in->begin(),
array_to_find_in->end(),
[&name_to_find](column_name_info const& col) { return col.name == name_to_find; });
if (found_col == array_to_find_in->end()) {
auto& col = array_to_find_in->emplace_back(name_to_find);
array_to_find_in = &col.children;
} else {
// Path exists. go down further.
array_to_find_in = &found_col->children;
}
}
}
for (auto& col : selected_columns) {
auto const& top_level_col_schema_idx = find_schema_child(root, col.name);
bool valid_column = build_column(&col, top_level_col_schema_idx, output_columns);
if (valid_column) output_column_schemas.push_back(top_level_col_schema_idx);
}
}
return std::make_tuple(
std::move(input_columns), std::move(output_columns), std::move(output_column_schemas));
}
};
/**
* @brief Generate depth remappings for repetition and definition levels.
*
* When dealing with columns that contain lists, we must examine incoming
* repetition and definition level pairs to determine what range of output nesting
* is indicated when adding new values. This function generates the mappings of
* the R/D levels to those start/end bounds
*
* @param remap Maps column schema index to the R/D remapping vectors for that column
* @param src_col_schema The column schema to generate the new mapping for
* @param md File metadata information
*/
void generate_depth_remappings(std::map<int, std::pair<std::vector<int>, std::vector<int>>>& remap,
int src_col_schema,
aggregate_metadata const& md)
{
// already generated for this level
if (remap.find(src_col_schema) != remap.end()) { return; }
auto schema = md.get_schema(src_col_schema);
int max_depth = md.get_output_nesting_depth(src_col_schema);
CUDF_EXPECTS(remap.find(src_col_schema) == remap.end(),
"Attempting to remap a schema more than once");
auto inserted =
remap.insert(std::pair<int, std::pair<std::vector<int>, std::vector<int>>>{src_col_schema, {}});
auto& depth_remap = inserted.first->second;
std::vector<int>& rep_depth_remap = (depth_remap.first);
rep_depth_remap.resize(schema.max_repetition_level + 1);
std::vector<int>& def_depth_remap = (depth_remap.second);
def_depth_remap.resize(schema.max_definition_level + 1);
// the key:
// for incoming level values R/D
// add values starting at the shallowest nesting level X has repetition level R
// until you reach the deepest nesting level Y that corresponds to the repetition level R1
// held by the nesting level that has definition level D
//
// Example: a 3 level struct with a list at the bottom
//
// R / D Depth
// level0 0 / 1 0
// level1 0 / 2 1
// level2 0 / 3 2
// list 0 / 3 3
// element 1 / 4 4
//
// incoming R/D : 0, 0 -> add values from depth 0 to 3 (def level 0 always maps to depth 0)
// incoming R/D : 0, 1 -> add values from depth 0 to 3
// incoming R/D : 0, 2 -> add values from depth 0 to 3
// incoming R/D : 1, 4 -> add values from depth 4 to 4
//
// Note : the -validity- of values is simply checked by comparing the incoming D value against the
// D value of the given nesting level (incoming D >= the D for the nesting level == valid,
// otherwise NULL). The tricky part is determining what nesting levels to add values at.
//
// For schemas with no repetition level (no lists), X is always 0 and Y is always max nesting
// depth.
//
// compute "X" from above
for (int s_idx = schema.max_repetition_level; s_idx >= 0; s_idx--) {
auto find_shallowest = [&](int r) {
int shallowest = -1;
int cur_depth = max_depth - 1;
int schema_idx = src_col_schema;
while (schema_idx > 0) {
auto cur_schema = md.get_schema(schema_idx);
if (cur_schema.max_repetition_level == r) {
// if this is a repeated field, map it one level deeper
shallowest = cur_schema.is_stub() ? cur_depth + 1 : cur_depth;
}
if (!cur_schema.is_stub()) { cur_depth--; }
schema_idx = cur_schema.parent_idx;
}
return shallowest;
};
rep_depth_remap[s_idx] = find_shallowest(s_idx);
}
// compute "Y" from above
for (int s_idx = schema.max_definition_level; s_idx >= 0; s_idx--) {
auto find_deepest = [&](int d) {
SchemaElement prev_schema;
int schema_idx = src_col_schema;
int r1 = 0;
while (schema_idx > 0) {
SchemaElement cur_schema = md.get_schema(schema_idx);
if (cur_schema.max_definition_level == d) {
// if this is a repeated field, map it one level deeper
r1 = cur_schema.is_stub() ? prev_schema.max_repetition_level
: cur_schema.max_repetition_level;
break;
}
prev_schema = cur_schema;
schema_idx = cur_schema.parent_idx;
}
// we now know R1 from above. return the deepest nesting level that has the
// same repetition level
schema_idx = src_col_schema;
int depth = max_depth - 1;
while (schema_idx > 0) {
SchemaElement cur_schema = md.get_schema(schema_idx);
if (cur_schema.max_repetition_level == r1) {
// if this is a repeated field, map it one level deeper
depth = cur_schema.is_stub() ? depth + 1 : depth;
break;
}
if (!cur_schema.is_stub()) { depth--; }
prev_schema = cur_schema;
schema_idx = cur_schema.parent_idx;
}
return depth;
};
def_depth_remap[s_idx] = find_deepest(s_idx);
}
}
/**
* @copydoc cudf::io::detail::parquet::read_column_chunks
*/
std::future<void> reader::impl::read_column_chunks(
std::vector<std::unique_ptr<datasource::buffer>>& page_data,
hostdevice_vector<gpu::ColumnChunkDesc>& chunks, // TODO const?
size_t begin_chunk,
size_t end_chunk,
const std::vector<size_t>& column_chunk_offsets,
std::vector<size_type> const& chunk_source_map,
rmm::cuda_stream_view stream)
{
// Transfer chunk data, coalescing adjacent chunks
std::vector<std::future<size_t>> read_tasks;
for (size_t chunk = begin_chunk; chunk < end_chunk;) {
const size_t io_offset = column_chunk_offsets[chunk];
size_t io_size = chunks[chunk].compressed_size;
size_t next_chunk = chunk + 1;
const bool is_compressed = (chunks[chunk].codec != parquet::Compression::UNCOMPRESSED);
while (next_chunk < end_chunk) {
const size_t next_offset = column_chunk_offsets[next_chunk];
const bool is_next_compressed =
(chunks[next_chunk].codec != parquet::Compression::UNCOMPRESSED);
if (next_offset != io_offset + io_size || is_next_compressed != is_compressed) {
// Can't merge if not contiguous or mixing compressed and uncompressed
// Not coalescing uncompressed with compressed chunks is so that compressed buffers can be
// freed earlier (immediately after decompression stage) to limit peak memory requirements
break;
}
io_size += chunks[next_chunk].compressed_size;
next_chunk++;
}
if (io_size != 0) {
auto& source = _sources[chunk_source_map[chunk]];
if (source->is_device_read_preferred(io_size)) {
auto buffer = rmm::device_buffer(io_size, stream);
auto fut_read_size = source->device_read_async(
io_offset, io_size, static_cast<uint8_t*>(buffer.data()), stream);
read_tasks.emplace_back(std::move(fut_read_size));
page_data[chunk] = datasource::buffer::create(std::move(buffer));
} else {
auto const buffer = source->host_read(io_offset, io_size);
page_data[chunk] =
datasource::buffer::create(rmm::device_buffer(buffer->data(), buffer->size(), stream));
}
auto d_compdata = page_data[chunk]->data();
do {
chunks[chunk].compressed_data = d_compdata;
d_compdata += chunks[chunk].compressed_size;
} while (++chunk != next_chunk);
} else {
chunk = next_chunk;
}
}
auto sync_fn = [](decltype(read_tasks) read_tasks) {
for (auto& task : read_tasks) {
task.wait();
}
};
return std::async(std::launch::deferred, sync_fn, std::move(read_tasks));
}
/**
* @copydoc cudf::io::detail::parquet::count_page_headers
*/
size_t reader::impl::count_page_headers(hostdevice_vector<gpu::ColumnChunkDesc>& chunks,
rmm::cuda_stream_view stream)
{
size_t total_pages = 0;
chunks.host_to_device(stream);
gpu::DecodePageHeaders(chunks.device_ptr(), chunks.size(), stream);
chunks.device_to_host(stream, true);
for (size_t c = 0; c < chunks.size(); c++) {
total_pages += chunks[c].num_data_pages + chunks[c].num_dict_pages;
}
return total_pages;
}
/**
* @copydoc cudf::io::detail::parquet::decode_page_headers
*/
void reader::impl::decode_page_headers(hostdevice_vector<gpu::ColumnChunkDesc>& chunks,
hostdevice_vector<gpu::PageInfo>& pages,
rmm::cuda_stream_view stream)
{
// IMPORTANT : if you change how pages are stored within a chunk (dist pages, then data pages),
// please update preprocess_nested_columns to reflect this.
for (size_t c = 0, page_count = 0; c < chunks.size(); c++) {