OmniSciDB  06b3bd477c
 All Classes Namespaces Files Functions Variables Typedefs Enumerations Enumerator Friends Macros Pages
QueryMemoryDescriptor.cpp
Go to the documentation of this file.
1 /*
2  * Copyright 2018 MapD Technologies, Inc.
3  *
4  * Licensed under the Apache License, Version 2.0 (the "License");
5  * you may not use this file except in compliance with the License.
6  * You may obtain a copy of the License at
7  *
8  * http://www.apache.org/licenses/LICENSE-2.0
9  *
10  * Unless required by applicable law or agreed to in writing, software
11  * distributed under the License is distributed on an "AS IS" BASIS,
12  * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
13  * See the License for the specific language governing permissions and
14  * limitations under the License.
15  */
16 
17 #include "QueryMemoryDescriptor.h"
18 
19 #include "../Execute.h"
20 #include "../ExpressionRewrite.h"
21 #include "../GroupByAndAggregate.h"
22 #include "../StreamingTopN.h"
23 #include "../UsedColumnsVisitor.h"
24 #include "ColSlotContext.h"
25 
27 
29 extern bool g_enable_columnar_output;
30 
31 namespace {
32 
33 bool is_int_and_no_bigger_than(const SQLTypeInfo& ti, const size_t byte_width) {
34  if (!ti.is_integer()) {
35  return false;
36  }
37  return get_bit_width(ti) <= (byte_width * 8);
38 }
39 
40 std::vector<ssize_t> target_expr_group_by_indices(
41  const std::list<std::shared_ptr<Analyzer::Expr>>& groupby_exprs,
42  const std::vector<Analyzer::Expr*>& target_exprs) {
43  std::vector<ssize_t> indices(target_exprs.size(), -1);
44  for (size_t target_idx = 0; target_idx < target_exprs.size(); ++target_idx) {
45  const auto target_expr = target_exprs[target_idx];
46  if (dynamic_cast<const Analyzer::AggExpr*>(target_expr)) {
47  continue;
48  }
49  const auto var_expr = dynamic_cast<const Analyzer::Var*>(target_expr);
50  if (var_expr && var_expr->get_which_row() == Analyzer::Var::kGROUPBY) {
51  indices[target_idx] = var_expr->get_varno() - 1;
52  continue;
53  }
54  }
55  return indices;
56 }
57 
58 std::vector<ssize_t> target_expr_proj_indices(const RelAlgExecutionUnit& ra_exe_unit,
60  if (ra_exe_unit.input_descs.size() > 1 ||
61  !ra_exe_unit.sort_info.order_entries.empty()) {
62  return {};
63  }
64  std::vector<ssize_t> target_indices(ra_exe_unit.target_exprs.size(), -1);
65  UsedColumnsVisitor columns_visitor;
66  std::unordered_set<int> used_columns;
67  for (const auto& simple_qual : ra_exe_unit.simple_quals) {
68  const auto crt_used_columns = columns_visitor.visit(simple_qual.get());
69  used_columns.insert(crt_used_columns.begin(), crt_used_columns.end());
70  }
71  for (const auto& qual : ra_exe_unit.quals) {
72  const auto crt_used_columns = columns_visitor.visit(qual.get());
73  used_columns.insert(crt_used_columns.begin(), crt_used_columns.end());
74  }
75  for (const auto& target : ra_exe_unit.target_exprs) {
76  const auto col_var = dynamic_cast<const Analyzer::ColumnVar*>(target);
77  if (col_var) {
78  const auto cd = get_column_descriptor_maybe(
79  col_var->get_column_id(), col_var->get_table_id(), cat);
80  if (!cd || !cd->isVirtualCol) {
81  continue;
82  }
83  }
84  const auto crt_used_columns = columns_visitor.visit(target);
85  used_columns.insert(crt_used_columns.begin(), crt_used_columns.end());
86  }
87  for (size_t target_idx = 0; target_idx < ra_exe_unit.target_exprs.size();
88  ++target_idx) {
89  const auto target_expr = ra_exe_unit.target_exprs[target_idx];
90  CHECK(target_expr);
91  const auto& ti = target_expr->get_type_info();
92  // TODO: add proper lazy fetch for varlen types in result set
93  if (ti.is_varlen()) {
94  continue;
95  }
96  const auto col_var = dynamic_cast<const Analyzer::ColumnVar*>(target_expr);
97  if (!col_var) {
98  continue;
99  }
100  if (!ti.is_varlen() &&
101  used_columns.find(col_var->get_column_id()) == used_columns.end()) {
102  // setting target index to be zero so that later it can be decoded properly (in lazy
103  // fetch, the zeroth target index indicates the corresponding rowid column for the
104  // projected entry)
105  target_indices[target_idx] = 0;
106  }
107  }
108  return target_indices;
109 }
110 
112  const size_t group_col_width) {
113  if (range.getType() == ExpressionRangeType::Invalid) {
114  return sizeof(int64_t);
115  }
116  switch (range.getType()) {
118  if (group_col_width == sizeof(int64_t) && range.hasNulls()) {
119  return sizeof(int64_t);
120  }
121  return range.getIntMax() < EMPTY_KEY_32 - 1 ? sizeof(int32_t) : sizeof(int64_t);
124  return sizeof(int64_t); // No compaction for floating point yet.
125  default:
126  UNREACHABLE();
127  }
128  return sizeof(int64_t);
129 }
130 
131 // TODO(miyu): make sure following setting of compact width is correct in all cases.
133  const std::vector<InputTableInfo>& query_infos,
134  const Executor* executor) {
135  int8_t compact_width{4};
136  for (const auto& groupby_expr : ra_exe_unit.groupby_exprs) {
137  const auto expr_range = getExpressionRange(groupby_expr.get(), query_infos, executor);
138  compact_width = std::max(compact_width,
140  expr_range, groupby_expr->get_type_info().get_size()));
141  }
142  return compact_width;
143 }
144 
145 bool use_streaming_top_n(const RelAlgExecutionUnit& ra_exe_unit,
146  const bool output_columnar) {
147  if (g_cluster) {
148  return false; // TODO(miyu)
149  }
150 
151  for (const auto target_expr : ra_exe_unit.target_exprs) {
152  if (dynamic_cast<const Analyzer::AggExpr*>(target_expr)) {
153  return false;
154  }
155  if (dynamic_cast<const Analyzer::WindowFunction*>(target_expr)) {
156  return false;
157  }
158  }
159 
160  // TODO: Allow streaming top n for columnar output
161  if (!output_columnar && ra_exe_unit.sort_info.order_entries.size() == 1 &&
162  ra_exe_unit.sort_info.limit &&
164  const auto only_order_entry = ra_exe_unit.sort_info.order_entries.front();
165  CHECK_GT(only_order_entry.tle_no, int(0));
166  CHECK_LE(static_cast<size_t>(only_order_entry.tle_no),
167  ra_exe_unit.target_exprs.size());
168  const auto order_entry_expr = ra_exe_unit.target_exprs[only_order_entry.tle_no - 1];
169  const auto n = ra_exe_unit.sort_info.offset + ra_exe_unit.sort_info.limit;
170  if ((order_entry_expr->get_type_info().is_number() ||
171  order_entry_expr->get_type_info().is_time()) &&
172  n <= 100000) { // TODO(miyu): relax?
173  return true;
174  }
175  }
176 
177  return false;
178 }
179 
180 } // namespace
181 
182 std::unique_ptr<QueryMemoryDescriptor> QueryMemoryDescriptor::init(
183  const Executor* executor,
184  const RelAlgExecutionUnit& ra_exe_unit,
185  const std::vector<InputTableInfo>& query_infos,
186  const ColRangeInfo& col_range_info,
187  const KeylessInfo& keyless_info,
188  const bool allow_multifrag,
189  const ExecutorDeviceType device_type,
190  const int8_t crt_min_byte_width,
191  const bool sort_on_gpu_hint,
192  const size_t shard_count,
193  const size_t max_groups_buffer_entry_count,
194  RenderInfo* render_info,
195  const CountDistinctDescriptors count_distinct_descriptors,
196  const bool must_use_baseline_sort,
197  const bool output_columnar_hint,
198  const bool streaming_top_n_hint) {
199  auto group_col_widths = get_col_byte_widths(ra_exe_unit.groupby_exprs, {});
200  const bool is_group_by{!group_col_widths.empty()};
201 
202  auto col_slot_context = ColSlotContext(ra_exe_unit.target_exprs, {});
203 
204  const auto min_slot_size = QueryMemoryDescriptor::pick_target_compact_width(
205  ra_exe_unit, query_infos, crt_min_byte_width);
206 
207  col_slot_context.setAllSlotsPaddedSize(min_slot_size);
208  col_slot_context.validate();
209 
210  if (!is_group_by) {
211  CHECK(!must_use_baseline_sort);
212 
213  return std::make_unique<QueryMemoryDescriptor>(
214  executor,
215  ra_exe_unit,
216  query_infos,
217  allow_multifrag,
218  false,
219  false,
220  -1,
221  ColRangeInfo{ra_exe_unit.estimator ? QueryDescriptionType::Estimator
223  0,
224  0,
225  0,
226  false},
227  col_slot_context,
228  std::vector<int8_t>{},
229  /*group_col_compact_width=*/0,
230  std::vector<ssize_t>{},
231  /*entry_count=*/1,
232  count_distinct_descriptors,
233  false,
234  output_columnar_hint,
235  render_info && render_info->isPotentialInSituRender(),
236  must_use_baseline_sort,
237  /*use_streaming_top_n=*/false);
238  }
239 
240  size_t entry_count = 1;
241  auto actual_col_range_info = col_range_info;
242  bool interleaved_bins_on_gpu = false;
243  bool keyless_hash = false;
244  bool streaming_top_n = false;
245  int8_t group_col_compact_width = 0;
246  int32_t idx_target_as_key = -1;
247  auto output_columnar = output_columnar_hint;
248  std::vector<ssize_t> target_groupby_indices;
249 
250  switch (col_range_info.hash_type_) {
252  if (render_info) {
253  render_info->setInSituDataIfUnset(false);
254  }
255  // keyless hash: whether or not group columns are stored at the beginning of the
256  // output buffer
257  keyless_hash =
258  (!sort_on_gpu_hint ||
260  col_range_info.max, col_range_info.min, col_range_info.bucket)) &&
261  !col_range_info.bucket && !must_use_baseline_sort && keyless_info.keyless;
262 
263  // if keyless, then this target index indicates wheter an entry is empty or not
264  // (acts as a key)
265  idx_target_as_key = keyless_info.target_index;
266 
267  if (group_col_widths.size() > 1) {
268  // col range info max contains the expected cardinality of the output
269  entry_count = static_cast<size_t>(actual_col_range_info.max);
270  actual_col_range_info.bucket = 0;
271  } else {
272  // single column perfect hash
273  entry_count = std::max(
274  GroupByAndAggregate::getBucketedCardinality(col_range_info), int64_t(1));
275  const size_t interleaved_max_threshold{512};
276 
277  if (must_use_baseline_sort) {
278  target_groupby_indices = target_expr_group_by_indices(ra_exe_unit.groupby_exprs,
279  ra_exe_unit.target_exprs);
280  col_slot_context =
281  ColSlotContext(ra_exe_unit.target_exprs, target_groupby_indices);
282  }
283 
284  bool has_varlen_sample_agg = false;
285  for (const auto& target_expr : ra_exe_unit.target_exprs) {
286  if (target_expr->get_contains_agg()) {
287  const auto agg_expr = dynamic_cast<Analyzer::AggExpr*>(target_expr);
288  CHECK(agg_expr);
289  if (agg_expr->get_aggtype() == kSAMPLE &&
290  agg_expr->get_type_info().is_varlen()) {
291  has_varlen_sample_agg = true;
292  break;
293  }
294  }
295  }
296 
297  interleaved_bins_on_gpu = keyless_hash && !has_varlen_sample_agg &&
298  (entry_count <= interleaved_max_threshold) &&
299  (device_type == ExecutorDeviceType::GPU) &&
301  count_distinct_descriptors) &&
302  !output_columnar;
303  }
304  break;
305  }
307  if (render_info) {
308  render_info->setInSituDataIfUnset(false);
309  }
310  entry_count = shard_count
311  ? (max_groups_buffer_entry_count + shard_count - 1) / shard_count
312  : max_groups_buffer_entry_count;
313  target_groupby_indices = target_expr_group_by_indices(ra_exe_unit.groupby_exprs,
314  ra_exe_unit.target_exprs);
315  col_slot_context = ColSlotContext(ra_exe_unit.target_exprs, target_groupby_indices);
316 
317  group_col_compact_width =
318  output_columnar ? 8
319  : pick_baseline_key_width(ra_exe_unit, query_infos, executor);
320 
321  actual_col_range_info =
323  break;
324  }
326  CHECK(!must_use_baseline_sort);
327 
328  if (streaming_top_n_hint && use_streaming_top_n(ra_exe_unit, output_columnar)) {
329  streaming_top_n = true;
330  entry_count = ra_exe_unit.sort_info.offset + ra_exe_unit.sort_info.limit;
331  } else {
332  if (ra_exe_unit.use_bump_allocator) {
333  output_columnar = false;
334  entry_count = 0;
335  } else {
336  entry_count = ra_exe_unit.scan_limit
337  ? static_cast<size_t>(ra_exe_unit.scan_limit)
338  : max_groups_buffer_entry_count;
339  }
340  }
341 
342  const auto catalog = executor->getCatalog();
343  CHECK(catalog);
344  target_groupby_indices = executor->plan_state_->allow_lazy_fetch_
345  ? target_expr_proj_indices(ra_exe_unit, *catalog)
346  : std::vector<ssize_t>{};
347 
348  col_slot_context = ColSlotContext(ra_exe_unit.target_exprs, target_groupby_indices);
349  break;
350  }
351  default:
352  UNREACHABLE() << "Unknown query type";
353  }
354 
355  return std::make_unique<QueryMemoryDescriptor>(
356  executor,
357  ra_exe_unit,
358  query_infos,
359  allow_multifrag,
360  keyless_hash,
361  interleaved_bins_on_gpu,
362  idx_target_as_key,
363  actual_col_range_info,
364  col_slot_context,
365  group_col_widths,
366  group_col_compact_width,
367  target_groupby_indices,
368  entry_count,
369  count_distinct_descriptors,
370  sort_on_gpu_hint,
371  output_columnar,
372  render_info && render_info->isPotentialInSituRender(),
373  must_use_baseline_sort,
374  streaming_top_n);
375 }
376 
378  const Executor* executor,
379  const RelAlgExecutionUnit& ra_exe_unit,
380  const std::vector<InputTableInfo>& query_infos,
381  const bool allow_multifrag,
382  const bool keyless_hash,
383  const bool interleaved_bins_on_gpu,
384  const int32_t idx_target_as_key,
385  const ColRangeInfo& col_range_info,
386  const ColSlotContext& col_slot_context,
387  const std::vector<int8_t>& group_col_widths,
388  const int8_t group_col_compact_width,
389  const std::vector<ssize_t>& target_groupby_indices,
390  const size_t entry_count,
391  const CountDistinctDescriptors count_distinct_descriptors,
392  const bool sort_on_gpu_hint,
393  const bool output_columnar_hint,
394  const bool render_output,
395  const bool must_use_baseline_sort,
396  const bool use_streaming_top_n)
397  : executor_(executor)
398  , allow_multifrag_(allow_multifrag)
399  , query_desc_type_(col_range_info.hash_type_)
400  , keyless_hash_(keyless_hash)
401  , interleaved_bins_on_gpu_(interleaved_bins_on_gpu)
402  , idx_target_as_key_(idx_target_as_key)
403  , group_col_widths_(group_col_widths)
404  , group_col_compact_width_(group_col_compact_width)
405  , target_groupby_indices_(target_groupby_indices)
406  , entry_count_(entry_count)
407  , min_val_(col_range_info.min)
408  , max_val_(col_range_info.max)
409  , bucket_(col_range_info.bucket)
410  , has_nulls_(col_range_info.has_nulls)
411  , count_distinct_descriptors_(count_distinct_descriptors)
412  , output_columnar_(false)
413  , render_output_(render_output)
414  , must_use_baseline_sort_(must_use_baseline_sort)
415  , is_table_function_(false)
416  , use_streaming_top_n_(use_streaming_top_n)
417  , force_4byte_float_(false)
418  , col_slot_context_(col_slot_context) {
421 
422  sort_on_gpu_ = sort_on_gpu_hint && canOutputColumnar() && !keyless_hash_;
423 
424  if (sort_on_gpu_) {
425  CHECK(!ra_exe_unit.use_bump_allocator);
426  output_columnar_ = true;
427  } else {
428  switch (query_desc_type_) {
430  output_columnar_ = output_columnar_hint;
431  break;
436  break;
438  output_columnar_ = output_columnar_hint;
439  break;
444  break;
445  default:
446  output_columnar_ = false;
447  break;
448  }
449  }
450 
452  // TODO(adb): Ensure fixed size buffer allocations are correct with all logical column
453  // sizes
454  CHECK(!ra_exe_unit.use_bump_allocator);
457  }
458 
459 #ifdef HAVE_CUDA
460  // Check Streaming Top N heap usage, bail if > max slab size, CUDA ONLY
461  if (use_streaming_top_n_ && executor->catalog_->getDataMgr().gpusPresent()) {
462  const auto thread_count = executor->blockSize() * executor->gridSize();
463  const auto total_buff_size =
465  if (total_buff_size > executor_->maxGpuSlabSize()) {
466  throw StreamingTopNOOM(total_buff_size);
467  }
468  }
469 #endif
470 }
471 
473  : executor_(nullptr)
474  , allow_multifrag_(false)
475  , query_desc_type_(QueryDescriptionType::Projection)
476  , keyless_hash_(false)
477  , interleaved_bins_on_gpu_(false)
478  , idx_target_as_key_(0)
479  , group_col_compact_width_(0)
480  , entry_count_(0)
481  , min_val_(0)
482  , max_val_(0)
483  , bucket_(0)
484  , has_nulls_(false)
485  , sort_on_gpu_(false)
486  , output_columnar_(false)
487  , render_output_(false)
488  , must_use_baseline_sort_(false)
489  , is_table_function_(false)
490  , use_streaming_top_n_(false)
491  , force_4byte_float_(false) {}
492 
494  const size_t entry_count,
495  const QueryDescriptionType query_desc_type,
496  const bool is_table_function)
497  : executor_(executor)
498  , allow_multifrag_(false)
499  , query_desc_type_(query_desc_type)
500  , keyless_hash_(false)
501  , interleaved_bins_on_gpu_(false)
502  , idx_target_as_key_(0)
503  , group_col_compact_width_(0)
504  , entry_count_(entry_count)
505  , min_val_(0)
506  , max_val_(0)
507  , bucket_(0)
508  , has_nulls_(false)
509  , sort_on_gpu_(false)
510  , output_columnar_(false)
511  , render_output_(false)
512  , must_use_baseline_sort_(false)
513  , is_table_function_(is_table_function)
514  , use_streaming_top_n_(false)
515  , force_4byte_float_(false) {}
516 
518  const int64_t min_val,
519  const int64_t max_val,
520  const bool has_nulls,
521  const std::vector<int8_t>& group_col_widths)
522  : executor_(nullptr)
523  , allow_multifrag_(false)
524  , query_desc_type_(query_desc_type)
525  , keyless_hash_(false)
526  , interleaved_bins_on_gpu_(false)
527  , idx_target_as_key_(0)
528  , group_col_widths_(group_col_widths)
529  , group_col_compact_width_(0)
530  , entry_count_(0)
531  , min_val_(min_val)
532  , max_val_(max_val)
533  , bucket_(0)
534  , has_nulls_(false)
535  , sort_on_gpu_(false)
536  , output_columnar_(false)
537  , render_output_(false)
538  , must_use_baseline_sort_(false)
539  , is_table_function_(false)
540  , use_streaming_top_n_(false)
541  , force_4byte_float_(false) {}
542 
544  // Note that this method does not check ptr reference members (e.g. executor_) or
545  // entry_count_
546  if (query_desc_type_ != other.query_desc_type_) {
547  return false;
548  }
549  if (keyless_hash_ != other.keyless_hash_) {
550  return false;
551  }
553  return false;
554  }
555  if (idx_target_as_key_ != other.idx_target_as_key_) {
556  return false;
557  }
558  if (force_4byte_float_ != other.force_4byte_float_) {
559  return false;
560  }
561  if (group_col_widths_ != other.group_col_widths_) {
562  return false;
563  }
565  return false;
566  }
568  return false;
569  }
570  if (min_val_ != other.min_val_) {
571  return false;
572  }
573  if (max_val_ != other.max_val_) {
574  return false;
575  }
576  if (bucket_ != other.bucket_) {
577  return false;
578  }
579  if (has_nulls_ != other.has_nulls_) {
580  return false;
581  }
583  return false;
584  } else {
585  // Count distinct descriptors can legitimately differ in device only.
586  for (size_t i = 0; i < count_distinct_descriptors_.size(); ++i) {
587  auto ref_count_distinct_desc = other.count_distinct_descriptors_[i];
588  auto count_distinct_desc = count_distinct_descriptors_[i];
589  count_distinct_desc.device_type = ref_count_distinct_desc.device_type;
590  if (ref_count_distinct_desc != count_distinct_desc) {
591  return false;
592  }
593  }
594  }
595  if (sort_on_gpu_ != other.sort_on_gpu_) {
596  return false;
597  }
598  if (output_columnar_ != other.output_columnar_) {
599  return false;
600  }
601  if (col_slot_context_ != other.col_slot_context_) {
602  return false;
603  }
604  return true;
605 }
606 
607 std::unique_ptr<QueryExecutionContext> QueryMemoryDescriptor::getQueryExecutionContext(
608  const RelAlgExecutionUnit& ra_exe_unit,
609  const Executor* executor,
610  const ExecutorDeviceType device_type,
611  const ExecutorDispatchMode dispatch_mode,
612  const int device_id,
613  const int64_t num_rows,
614  const std::vector<std::vector<const int8_t*>>& col_buffers,
615  const std::vector<std::vector<uint64_t>>& frag_offsets,
616  std::shared_ptr<RowSetMemoryOwner> row_set_mem_owner,
617  const bool output_columnar,
618  const bool sort_on_gpu,
619  RenderInfo* render_info) const {
620  auto timer = DEBUG_TIMER(__func__);
621  if (frag_offsets.empty()) {
622  return nullptr;
623  }
624  return std::unique_ptr<QueryExecutionContext>(
625  new QueryExecutionContext(ra_exe_unit,
626  *this,
627  executor,
628  device_type,
629  dispatch_mode,
630  device_id,
631  num_rows,
632  col_buffers,
633  frag_offsets,
634  row_set_mem_owner,
635  output_columnar,
636  sort_on_gpu,
637  render_info));
638 }
639 
641  const RelAlgExecutionUnit& ra_exe_unit,
642  const std::vector<InputTableInfo>& query_infos,
643  const int8_t crt_min_byte_width) {
644  if (g_bigint_count) {
645  return sizeof(int64_t);
646  }
647  int8_t compact_width{0};
648  auto col_it = ra_exe_unit.input_col_descs.begin();
649  int unnest_array_col_id{std::numeric_limits<int>::min()};
650  for (const auto& groupby_expr : ra_exe_unit.groupby_exprs) {
651  const auto uoper = dynamic_cast<Analyzer::UOper*>(groupby_expr.get());
652  if (uoper && uoper->get_optype() == kUNNEST) {
653  const auto& arg_ti = uoper->get_operand()->get_type_info();
654  CHECK(arg_ti.is_array());
655  const auto& elem_ti = arg_ti.get_elem_type();
656  if (elem_ti.is_string() && elem_ti.get_compression() == kENCODING_DICT) {
657  unnest_array_col_id = (*col_it)->getColId();
658  } else {
659  compact_width = crt_min_byte_width;
660  break;
661  }
662  }
663  ++col_it;
664  }
665  if (!compact_width &&
666  (ra_exe_unit.groupby_exprs.size() != 1 || !ra_exe_unit.groupby_exprs.front())) {
667  compact_width = crt_min_byte_width;
668  }
669  if (!compact_width) {
670  col_it = ra_exe_unit.input_col_descs.begin();
671  std::advance(col_it, ra_exe_unit.groupby_exprs.size());
672  for (const auto target : ra_exe_unit.target_exprs) {
673  const auto& ti = target->get_type_info();
674  const auto agg = dynamic_cast<const Analyzer::AggExpr*>(target);
675  if (agg && agg->get_arg()) {
676  compact_width = crt_min_byte_width;
677  break;
678  }
679 
680  if (agg) {
681  CHECK_EQ(kCOUNT, agg->get_aggtype());
682  CHECK(!agg->get_is_distinct());
683  ++col_it;
684  continue;
685  }
686 
687  if (is_int_and_no_bigger_than(ti, 4) ||
688  (ti.is_string() && ti.get_compression() == kENCODING_DICT)) {
689  ++col_it;
690  continue;
691  }
692 
693  const auto uoper = dynamic_cast<Analyzer::UOper*>(target);
694  if (uoper && uoper->get_optype() == kUNNEST &&
695  (*col_it)->getColId() == unnest_array_col_id) {
696  const auto arg_ti = uoper->get_operand()->get_type_info();
697  CHECK(arg_ti.is_array());
698  const auto& elem_ti = arg_ti.get_elem_type();
699  if (elem_ti.is_string() && elem_ti.get_compression() == kENCODING_DICT) {
700  ++col_it;
701  continue;
702  }
703  }
704 
705  compact_width = crt_min_byte_width;
706  break;
707  }
708  }
709  if (!compact_width) {
710  size_t total_tuples{0};
711  for (const auto& qi : query_infos) {
712  total_tuples += qi.info.getNumTuples();
713  }
714  return total_tuples <= static_cast<size_t>(std::numeric_limits<uint32_t>::max()) ||
715  unnest_array_col_id != std::numeric_limits<int>::min()
716  ? 4
717  : crt_min_byte_width;
718  } else {
719  // TODO(miyu): relax this condition to allow more cases just w/o padding
720  for (auto wid : get_col_byte_widths(ra_exe_unit.target_exprs, {})) {
721  compact_width = std::max(compact_width, wid);
722  }
723  return compact_width;
724  }
725 }
726 
729 }
730 
733  size_t total_bytes{0};
734  if (keyless_hash_) {
735  // ignore, there's no group column in the output buffer
737  } else {
738  total_bytes += group_col_widths_.size() * getEffectiveKeyWidth();
739  total_bytes = align_to_int64(total_bytes);
740  }
741  total_bytes += getColsSize();
742  return align_to_int64(total_bytes);
743 }
744 
746  return (interleaved_bins_on_gpu_ ? executor_->warpSize() : 1);
747 }
748 
751 }
752 
761 }
762 
768  const size_t num_entries_per_column) const {
769  return col_slot_context_.getTotalBytesOfColumnarBuffers(num_entries_per_column);
770 }
771 
782  const size_t projection_count) const {
783  constexpr size_t row_index_width = sizeof(int64_t);
784  return getTotalBytesOfColumnarBuffers(projection_count) +
785  row_index_width * projection_count;
786 }
787 
788 size_t QueryMemoryDescriptor::getColOnlyOffInBytes(const size_t col_idx) const {
789  return col_slot_context_.getColOnlyOffInBytes(col_idx);
790 }
791 
792 /*
793  * Returns the memory offset in bytes for a specific agg column in the output
794  * memory buffer. Depending on the query type, there may be some extra portion
795  * of memory prepended at the beginning of the buffer. A brief description of
796  * the memory layout is as follows:
797  * 1. projections: index column (64bit) + all target columns
798  * 2. group by: all group columns (64-bit each) + all agg columns
799  * 2a. if keyless, there is no prepending group column stored at the beginning
800  */
801 size_t QueryMemoryDescriptor::getColOffInBytes(const size_t col_idx) const {
802  const auto warp_count = getWarpCount();
803  if (output_columnar_) {
804  CHECK_EQ(size_t(1), warp_count);
805  size_t offset{0};
806  if (!keyless_hash_) {
808  }
809  for (size_t index = 0; index < col_idx; ++index) {
811  }
812  return offset;
813  }
814 
815  size_t offset{0};
816  if (keyless_hash_) {
817  // ignore, there's no group column in the output buffer
819  } else {
820  offset += group_col_widths_.size() * getEffectiveKeyWidth();
821  offset = align_to_int64(offset);
822  }
823  offset += getColOnlyOffInBytes(col_idx);
824  return offset;
825 }
826 
827 /*
828  * Returns the memory offset for a particular group column in the prepended group
829  * columns portion of the memory.
830  */
832  const size_t group_idx) const {
834  CHECK(group_idx < getGroupbyColCount());
835  size_t offset{0};
836  for (size_t col_idx = 0; col_idx < group_idx; col_idx++) {
837  // TODO(Saman): relax that int64_bit part immediately
838  offset += align_to_int64(
839  std::max(groupColWidth(col_idx), static_cast<int8_t>(sizeof(int64_t))) *
840  getEntryCount());
841  }
842  return offset;
843 }
844 
845 /*
846  * Returns total amount of memory prepended at the beginning of the output memory
847  * buffer.
848  */
851  size_t buffer_size{0};
852  for (size_t group_idx = 0; group_idx < getGroupbyColCount(); group_idx++) {
853  buffer_size += align_to_int64(
854  std::max(groupColWidth(group_idx), static_cast<int8_t>(sizeof(int64_t))) *
855  getEntryCount());
856  }
857  return buffer_size;
858 }
859 
860 size_t QueryMemoryDescriptor::getColOffInBytesInNextBin(const size_t col_idx) const {
861  auto warp_count = getWarpCount();
862  if (output_columnar_) {
863  CHECK_EQ(size_t(1), group_col_widths_.size());
864  CHECK_EQ(size_t(1), warp_count);
865  return getPaddedSlotWidthBytes(col_idx);
866  }
867 
868  return warp_count * getRowSize();
869 }
870 
871 size_t QueryMemoryDescriptor::getNextColOffInBytes(const int8_t* col_ptr,
872  const size_t bin,
873  const size_t col_idx) const {
875  size_t offset{0};
876  auto warp_count = getWarpCount();
877  const auto chosen_bytes = getPaddedSlotWidthBytes(col_idx);
878  const auto total_slot_count = getSlotCount();
879  if (col_idx + 1 == total_slot_count) {
880  if (output_columnar_) {
881  return (entry_count_ - bin) * chosen_bytes;
882  } else {
883  return static_cast<size_t>(align_to_int64(col_ptr + chosen_bytes) - col_ptr);
884  }
885  }
886 
887  const auto next_chosen_bytes = getPaddedSlotWidthBytes(col_idx + 1);
888  if (output_columnar_) {
889  CHECK_EQ(size_t(1), group_col_widths_.size());
890  CHECK_EQ(size_t(1), warp_count);
891 
892  offset = align_to_int64(entry_count_ * chosen_bytes);
893 
894  offset += bin * (next_chosen_bytes - chosen_bytes);
895  return offset;
896  }
897 
898  if (next_chosen_bytes == sizeof(int64_t)) {
899  return static_cast<size_t>(align_to_int64(col_ptr + chosen_bytes) - col_ptr);
900  } else {
901  return chosen_bytes;
902  }
903 }
904 
906  const RelAlgExecutionUnit& ra_exe_unit,
907  const unsigned thread_count,
908  const ExecutorDeviceType device_type) const {
909  if (use_streaming_top_n_) {
910  const size_t n = ra_exe_unit.sort_info.offset + ra_exe_unit.sort_info.limit;
911  return streaming_top_n::get_heap_size(getRowSize(), n, thread_count);
912  }
913  return getBufferSizeBytes(device_type, entry_count_);
914 }
915 
928  const size_t entry_count) const {
930  CHECK_GE(group_col_widths_.size(), size_t(1));
931  auto row_bytes = align_to_int64(getColsSize());
932 
933  return (interleavedBins(device_type) ? executor_->warpSize() : 1) * entry_count *
934  row_bytes;
935  }
936 
937  constexpr size_t row_index_width = sizeof(int64_t);
938  size_t total_bytes{0};
939  if (output_columnar_) {
941  ? row_index_width * entry_count
942  : sizeof(int64_t) * group_col_widths_.size() * entry_count) +
944  } else {
945  total_bytes = getRowSize() * entry_count;
946  }
947 
948  return total_bytes;
949 }
950 
952  const ExecutorDeviceType device_type) const {
953  return getBufferSizeBytes(device_type, entry_count_);
954 }
955 
957  output_columnar_ = val;
960  }
961 }
962 
963 /*
964  * Indicates the query types that are currently allowed to use the logical
965  * sized columns instead of padded sized ones.
966  */
968  // In distributed mode, result sets are serialized using rowwise iterators, so we use
969  // consistent slot widths for now
970  return output_columnar_ && !g_cluster &&
972 }
973 
975  size_t total_slot_count = col_slot_context_.getSlotCount();
976 
977  if (target_groupby_indices_.empty()) {
978  return total_slot_count;
979  }
980  return total_slot_count - std::count_if(target_groupby_indices_.begin(),
982  [](const ssize_t i) { return i >= 0; });
983 }
984 
987  getGroupbyColCount() == 1);
988 }
989 
992 }
993 
995  if (g_cluster || is_table_function_) {
996  return true;
997  }
999  return true;
1000  }
1001  if (executor_->isCPUOnly() || render_output_ ||
1005  getGroupbyColCount() > 1)) {
1006  return true;
1007  }
1010 }
1011 
1013  return device_type == ExecutorDeviceType::GPU && !render_output_ &&
1015 }
1016 
1018  return interleaved_bins_on_gpu_ && device_type == ExecutorDeviceType::GPU;
1019 }
1020 
1021 // TODO(Saman): an implementation detail, so move this out of QMD
1023  const ExecutorDeviceType device_type) const {
1024  if (device_type != ExecutorDeviceType::GPU) {
1025  return false;
1026  } else {
1027  auto cuda_mgr = executor_->getCatalog()->getDataMgr().getCudaMgr();
1028  CHECK(cuda_mgr);
1029  return cuda_mgr->isArchVoltaForAll();
1030  }
1031 }
1032 
1034  return col_slot_context_.getColCount();
1035 }
1036 
1039 }
1040 
1041 const int8_t QueryMemoryDescriptor::getPaddedSlotWidthBytes(const size_t slot_idx) const {
1042  return col_slot_context_.getSlotInfo(slot_idx).padded_size;
1043 }
1044 
1046  const size_t slot_idx) const {
1047  return col_slot_context_.getSlotInfo(slot_idx).logical_size;
1048 }
1049 
1051  const size_t col_idx) const {
1052  const auto& col_slots = col_slot_context_.getSlotsForCol(col_idx);
1053  CHECK_EQ(col_slots.size(), size_t(1));
1054  return col_slots.front();
1055 }
1056 
1057 void QueryMemoryDescriptor::useConsistentSlotWidthSize(const int8_t slot_width_size) {
1058  col_slot_context_.setAllSlotsSize(slot_width_size);
1059 }
1060 
1062  // Note: Actual row size may include padding (see ResultSetBufferAccessors.h)
1064 }
1065 
1067  const int8_t actual_min_byte_width) const {
1068  return col_slot_context_.getMinPaddedByteSize(actual_min_byte_width);
1069 }
1070 
1072  const std::vector<std::tuple<int8_t, int8_t>>& slots_for_col) {
1073  col_slot_context_.addColumn(slots_for_col);
1074 }
1075 
1078 }
1079 
1082 }
1083 
1088 }
1089 
1091  switch (query_desc_type_) {
1093  return "Perfect Hash";
1095  return "Baseline Hash";
1097  return "Projection";
1099  return "Non-grouped Aggregate";
1101  return "Estimator";
1102  default:
1103  UNREACHABLE();
1104  }
1105  return "";
1106 }
1107 
1108 std::string QueryMemoryDescriptor::toString() const {
1109  auto str = reductionKey();
1110  str += "\tAllow Multifrag: " + bool_to_string(allow_multifrag_) + "\n";
1111  str += "\tInterleaved Bins on GPU: " + bool_to_string(interleaved_bins_on_gpu_) + "\n";
1112  str += "\tBlocks Share Memory: " + bool_to_string(blocksShareMemory()) + "\n";
1113  str += "\tThreads Share Memory: " + bool_to_string(threadsShareMemory()) + "\n";
1114  str += "\tUses Fast Group Values: " + bool_to_string(usesGetGroupValueFast()) + "\n";
1115  str += "\tLazy Init Groups (GPU): " +
1117  str += "\tEntry Count: " + std::to_string(entry_count_) + "\n";
1118  str += "\tMin Val (perfect hash only): " + std::to_string(min_val_) + "\n";
1119  str += "\tMax Val (perfect hash only): " + std::to_string(max_val_) + "\n";
1120  str += "\tBucket Val (perfect hash only): " + std::to_string(bucket_) + "\n";
1121  str += "\tSort on GPU: " + bool_to_string(sort_on_gpu_) + "\n";
1122  str += "\tUse Streaming Top N: " + bool_to_string(use_streaming_top_n_) + "\n";
1123  str += "\tOutput Columnar: " + bool_to_string(output_columnar_) + "\n";
1124  str += "\tRender Output: " + bool_to_string(render_output_) + "\n";
1125  str += "\tUse Baseline Sort: " + bool_to_string(must_use_baseline_sort_) + "\n";
1126  return str;
1127 }
1128 
1130  std::string str;
1131  str += "Query Memory Descriptor State\n";
1132  str += "\tQuery Type: " + queryDescTypeToString() + "\n";
1133  str +=
1134  "\tKeyless Hash: " + bool_to_string(keyless_hash_) +
1135  (keyless_hash_ ? ", target index for key: " + std::to_string(getTargetIdxForKey())
1136  : "") +
1137  "\n";
1138  str += "\tEffective key width: " + std::to_string(getEffectiveKeyWidth()) + "\n";
1139  str += "\tNumber of group columns: " + std::to_string(getGroupbyColCount()) + "\n";
1140  const auto group_indices_size = targetGroupbyIndicesSize();
1141  if (group_indices_size) {
1142  std::vector<std::string> group_indices_strings;
1143  for (size_t target_idx = 0; target_idx < group_indices_size; ++target_idx) {
1144  group_indices_strings.push_back(std::to_string(getTargetGroupbyIndex(target_idx)));
1145  }
1146  str += "\tTarget group by indices: " +
1147  boost::algorithm::join(group_indices_strings, ",") + "\n";
1148  }
1149  str += "\t" + col_slot_context_.toString();
1150  return str;
1151 }
1152 
1153 std::vector<TargetInfo> target_exprs_to_infos(
1154  const std::vector<Analyzer::Expr*>& targets,
1156  std::vector<TargetInfo> target_infos;
1157  for (const auto target_expr : targets) {
1158  auto target = get_target_info(target_expr, g_bigint_count);
1159  if (query_mem_desc.getQueryDescriptionType() ==
1161  set_notnull(target, false);
1162  target.sql_type.set_notnull(false);
1163  }
1164  target_infos.push_back(target);
1165  }
1166  return target_infos;
1167 }
int8_t getMinPaddedByteSize(const int8_t actual_min_byte_width) const
std::vector< Analyzer::Expr * > target_exprs
static bool many_entries(const int64_t max_val, const int64_t min_val, const int64_t bucket)
#define CHECK_EQ(x, y)
Definition: Logger.h:205
size_t getBufferSizeBytes(const RelAlgExecutionUnit &ra_exe_unit, const unsigned thread_count, const ExecutorDeviceType device_type) const
bool g_enable_smem_group_by
static int64_t getBucketedCardinality(const ColRangeInfo &col_range_info)
void alignPaddedSlots(const bool sort_on_gpu)
std::string cat(Ts &&...args)
int8_t logical_size
const int8_t const int64_t * num_rows
class for a per-database catalog. also includes metadata for the current database and the current use...
Definition: Catalog.h:86
size_t getTotalBytesOfColumnarProjections(const size_t projection_count) const
std::vector< int8_t > get_col_byte_widths(const T &col_expr_list, const std::vector< ssize_t > &col_exprs_to_not_project)
ExecutorDeviceType
std::unique_ptr< QueryExecutionContext > getQueryExecutionContext(const RelAlgExecutionUnit &, const Executor *executor, const ExecutorDeviceType device_type, const ExecutorDispatchMode dispatch_mode, const int device_id, const int64_t num_rows, const std::vector< std::vector< const int8_t * >> &col_buffers, const std::vector< std::vector< uint64_t >> &frag_offsets, std::shared_ptr< RowSetMemoryOwner >, const bool output_columnar, const bool sort_on_gpu, RenderInfo *) const
std::string toString() const
bool isLogicalSizedColumnsAllowed() const
TargetInfo get_target_info(const PointerType target_expr, const bool bigint_count)
Definition: TargetInfo.h:78
int8_t pick_baseline_key_component_width(const ExpressionRange &range, const size_t group_col_width)
const std::list< Analyzer::OrderEntry > order_entries
std::string join(T const &container, std::string const &delim)
std::vector< InputDescriptor > input_descs
static std::unique_ptr< QueryMemoryDescriptor > init(const Executor *executor, const RelAlgExecutionUnit &ra_exe_unit, const std::vector< InputTableInfo > &query_infos, const ColRangeInfo &col_range_info, const KeylessInfo &keyless_info, const bool allow_multifrag, const ExecutorDeviceType device_type, const int8_t crt_min_byte_width, const bool sort_on_gpu_hint, const size_t shard_count, const size_t max_groups_buffer_entry_count, RenderInfo *render_info, const CountDistinctDescriptors count_distinct_descriptors, const bool must_use_baseline_sort, const bool output_columnar_hint, const bool streaming_top_n_hint)
#define UNREACHABLE()
Definition: Logger.h:241
void setOutputColumnar(const bool val)
const SortAlgorithm algorithm
#define CHECK_GE(x, y)
Definition: Logger.h:210
size_t getAllSlotsPaddedSize() const
size_t getAllSlotsAlignedPaddedSize() const
size_t getNextColOffInBytes(const int8_t *col_ptr, const size_t bin, const size_t col_idx) const
size_t getEffectiveKeyWidth() const
bool use_streaming_top_n(const RelAlgExecutionUnit &ra_exe_unit, const bool output_columnar)
const std::list< std::shared_ptr< Analyzer::Expr > > groupby_exprs
T visit(const Analyzer::Expr *expr) const
#define CHECK_GT(x, y)
Definition: Logger.h:209
void setAllSlotsSize(const int8_t slot_width_size)
std::string to_string(char const *&&v)
void useConsistentSlotWidthSize(const int8_t slot_width_size)
const SlotSize & getSlotInfo(const size_t slot_idx) const
size_t getColOnlyOffInBytes(const size_t col_idx) const
ExecutorDispatchMode
size_t getColOnlyOffInBytes(const size_t slot_idx) const
const size_t limit
bool g_enable_columnar_output
Definition: Execute.cpp:90
int8_t groupColWidth(const size_t key_idx) const
size_t get_bit_width(const SQLTypeInfo &ti)
CHECK(cgen_state)
const ColumnDescriptor * get_column_descriptor_maybe(const int col_id, const int table_id, const Catalog_Namespace::Catalog &cat)
Definition: Execute.h:149
std::vector< CountDistinctDescriptor > CountDistinctDescriptors
Definition: CountDistinct.h:35
const SortInfo sort_info
size_t getCompactByteWidth() const
Provides column info and slot info for the output buffer and some metadata helpers.
size_t getGroupbyColCount() const
std::vector< ssize_t > target_expr_group_by_indices(const std::list< std::shared_ptr< Analyzer::Expr >> &groupby_exprs, const std::vector< Analyzer::Expr * > &target_exprs)
bool is_integer() const
Definition: sqltypes.h:417
bool lazyInitGroups(const ExecutorDeviceType) const
size_t targetGroupbyIndicesSize() const
size_t getPrependedGroupBufferSizeInBytes() const
size_t getTotalBytesOfColumnarBuffers() const
static int8_t pick_target_compact_width(const RelAlgExecutionUnit &ra_exe_unit, const std::vector< InputTableInfo > &query_infos, const int8_t crt_min_byte_width)
std::string bool_to_string(const bool val)
bool g_bigint_count
CountDistinctDescriptors count_distinct_descriptors_
void validate() const
int get_varno() const
Definition: Analyzer.h:276
const int8_t getPaddedSlotWidthBytes(const size_t slot_idx) const
ExpressionRange getExpressionRange(const Analyzer::BinOper *expr, const std::vector< InputTableInfo > &query_infos, const Executor *, boost::optional< std::list< std::shared_ptr< Analyzer::Expr >>> simple_quals)
bool hasNulls() const
const SQLTypeInfo & get_type_info() const
Definition: Analyzer.h:79
ssize_t getTargetGroupbyIndex(const size_t target_idx) const
QueryDescriptionType getQueryDescriptionType() const
void addColumn(const std::vector< std::tuple< int8_t, int8_t >> &slots_for_col)
#define CHECK_LE(x, y)
Definition: Logger.h:208
const Expr * get_operand() const
Definition: Analyzer.h:372
QueryDescriptionType query_desc_type_
int8_t padded_size
Definition: sqldefs.h:76
int8_t updateActualMinByteWidth(const int8_t actual_min_byte_width) const
size_t getTotalBytesOfColumnarBuffers(const size_t entry_count) const
bool operator==(const QueryMemoryDescriptor &other) const
Descriptor for the result set buffer layout.
bool is_int_and_no_bigger_than(const SQLTypeInfo &ti, const size_t byte_width)
std::list< std::shared_ptr< Analyzer::Expr > > quals
ExpressionRangeType getType() const
size_t get_heap_size(const size_t row_size, const size_t n, const size_t thread_count)
int64_t getIntMax() const
bool isWarpSyncRequired(const ExecutorDeviceType) const
std::string toString() const
size_t getSlotCount() const
void setAllSlotsPaddedSizeToLogicalSize()
bool interleavedBins(const ExecutorDeviceType) const
bool g_enable_watchdog false
Definition: Execute.cpp:74
#define DEBUG_TIMER(name)
Definition: Logger.h:313
size_t getColCount() const
std::vector< int8_t > group_col_widths_
#define EMPTY_KEY_32
QueryDescriptionType
Definition: Types.h:26
bool g_cluster
std::vector< TargetInfo > target_exprs_to_infos(const std::vector< Analyzer::Expr * > &targets, const QueryMemoryDescriptor &query_mem_desc)
const std::vector< size_t > & getSlotsForCol(const size_t col_idx) const
std::vector< ssize_t > target_expr_proj_indices(const RelAlgExecutionUnit &ra_exe_unit, const Catalog_Namespace::Catalog &cat)
std::string queryDescTypeToString() const
void sort_on_gpu(int64_t *val_buff, int32_t *idx_buff, const uint64_t entry_count, const bool desc, const uint32_t chosen_bytes, ThrustAllocator &alloc)
std::list< std::shared_ptr< const InputColDescriptor > > input_col_descs
void addColSlotInfo(const std::vector< std::tuple< int8_t, int8_t >> &slots_for_col)
const size_t offset
static bool countDescriptorsLogicallyEmpty(const CountDistinctDescriptors &count_distinct_descriptors)
void setAllUnsetSlotsPaddedSize(const int8_t padded_size)
const int8_t getSlotIndexForSingleSlotCol(const size_t col_idx) const
size_t getBufferColSlotCount() const
const int8_t getLogicalSlotWidthBytes(const size_t slot_idx) const
size_t getColOffInBytes(const size_t col_idx) const
size_t getColOffInBytesInNextBin(const size_t col_idx) const
std::vector< ssize_t > target_groupby_indices_
FORCE_INLINE HOST DEVICE T align_to_int64(T addr)
int8_t pick_baseline_key_width(const RelAlgExecutionUnit &ra_exe_unit, const std::vector< InputTableInfo > &query_infos, const Executor *executor)
std::string reductionKey() const
std::list< std::shared_ptr< Analyzer::Expr > > simple_quals
void set_notnull(TargetInfo &target, const bool not_null)
int32_t getTargetIdxForKey() const
size_t getPrependedGroupColOffInBytes(const size_t group_idx) const