OmniSciDB  a987f07e93
 All Classes Namespaces Files Functions Variables Typedefs Enumerations Enumerator Friends Macros Groups Pages
QueryMemoryDescriptor.cpp
Go to the documentation of this file.
1 /*
2  * Copyright 2022 HEAVY.AI, 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 
26 #include <boost/algorithm/cxx11/any_of.hpp>
27 
29 extern bool g_enable_columnar_output;
30 extern size_t g_streaming_topn_max;
31 
32 namespace {
33 
34 bool is_int_and_no_bigger_than(const SQLTypeInfo& ti, const size_t byte_width) {
35  if (!ti.is_integer()) {
36  return false;
37  }
38  return get_bit_width(ti) <= (byte_width * 8);
39 }
40 
42  return range.getIntMin() > INT32_MIN && range.getIntMax() < EMPTY_KEY_32 - 1;
43 }
44 
45 std::vector<int64_t> target_expr_group_by_indices(
46  const std::list<std::shared_ptr<Analyzer::Expr>>& groupby_exprs,
47  const std::vector<Analyzer::Expr*>& target_exprs) {
48  std::vector<int64_t> indices(target_exprs.size(), -1);
49  for (size_t target_idx = 0; target_idx < target_exprs.size(); ++target_idx) {
50  const auto target_expr = target_exprs[target_idx];
51  if (dynamic_cast<const Analyzer::AggExpr*>(target_expr)) {
52  continue;
53  }
54  const auto var_expr = dynamic_cast<const Analyzer::Var*>(target_expr);
55  if (var_expr && var_expr->get_which_row() == Analyzer::Var::kGROUPBY) {
56  indices[target_idx] = var_expr->get_varno() - 1;
57  continue;
58  }
59  }
60  return indices;
61 }
62 
63 std::vector<int64_t> target_expr_proj_indices(const RelAlgExecutionUnit& ra_exe_unit,
65  if (ra_exe_unit.input_descs.size() > 1 ||
66  !ra_exe_unit.sort_info.order_entries.empty()) {
67  return {};
68  }
69  std::vector<int64_t> target_indices(ra_exe_unit.target_exprs.size(), -1);
70  UsedColumnsVisitor columns_visitor;
71  std::unordered_set<int> used_columns;
72  for (const auto& simple_qual : ra_exe_unit.simple_quals) {
73  const auto crt_used_columns = columns_visitor.visit(simple_qual.get());
74  used_columns.insert(crt_used_columns.begin(), crt_used_columns.end());
75  }
76  for (const auto& qual : ra_exe_unit.quals) {
77  const auto crt_used_columns = columns_visitor.visit(qual.get());
78  used_columns.insert(crt_used_columns.begin(), crt_used_columns.end());
79  }
80  for (const auto& target : ra_exe_unit.target_exprs) {
81  const auto col_var = dynamic_cast<const Analyzer::ColumnVar*>(target);
82  if (col_var) {
83  const auto cd = get_column_descriptor_maybe(
84  col_var->get_column_id(), col_var->get_table_id(), cat);
85  if (!cd || !cd->isVirtualCol) {
86  continue;
87  }
88  }
89  const auto crt_used_columns = columns_visitor.visit(target);
90  used_columns.insert(crt_used_columns.begin(), crt_used_columns.end());
91  }
92  for (size_t target_idx = 0; target_idx < ra_exe_unit.target_exprs.size();
93  ++target_idx) {
94  const auto target_expr = ra_exe_unit.target_exprs[target_idx];
95  CHECK(target_expr);
96  const auto& ti = target_expr->get_type_info();
97  // TODO: add proper lazy fetch for varlen types in result set
98  if (ti.is_varlen()) {
99  continue;
100  }
101  const auto col_var = dynamic_cast<const Analyzer::ColumnVar*>(target_expr);
102  if (!col_var) {
103  continue;
104  }
105  if (!ti.is_varlen() &&
106  used_columns.find(col_var->get_column_id()) == used_columns.end()) {
107  // setting target index to be zero so that later it can be decoded properly (in lazy
108  // fetch, the zeroth target index indicates the corresponding rowid column for the
109  // projected entry)
110  target_indices[target_idx] = 0;
111  }
112  }
113  return target_indices;
114 }
115 
117  const size_t group_col_width) {
118  if (range.getType() == ExpressionRangeType::Invalid) {
119  return sizeof(int64_t);
120  }
121  switch (range.getType()) {
123  if (group_col_width == sizeof(int64_t) && range.hasNulls()) {
124  return sizeof(int64_t);
125  }
126  return is_valid_int32_range(range) ? sizeof(int32_t) : sizeof(int64_t);
129  return sizeof(int64_t); // No compaction for floating point yet.
130  default:
131  UNREACHABLE();
132  }
133  return sizeof(int64_t);
134 }
135 
136 // TODO(miyu): make sure following setting of compact width is correct in all cases.
138  const std::vector<InputTableInfo>& query_infos,
139  const Executor* executor) {
140  int8_t compact_width{4};
141  for (const auto& groupby_expr : ra_exe_unit.groupby_exprs) {
142  const auto expr_range = getExpressionRange(groupby_expr.get(), query_infos, executor);
143  compact_width = std::max(compact_width,
145  expr_range, groupby_expr->get_type_info().get_size()));
146  }
147  return compact_width;
148 }
149 
150 bool use_streaming_top_n(const RelAlgExecutionUnit& ra_exe_unit,
151  const bool output_columnar) {
152  if (g_cluster) {
153  return false; // TODO(miyu)
154  }
155 
156  for (const auto target_expr : ra_exe_unit.target_exprs) {
157  if (dynamic_cast<const Analyzer::AggExpr*>(target_expr)) {
158  return false;
159  }
160  if (dynamic_cast<const Analyzer::WindowFunction*>(target_expr)) {
161  return false;
162  }
163  }
164 
165  // TODO: Allow streaming top n for columnar output
166  if (!output_columnar && ra_exe_unit.sort_info.order_entries.size() == 1 &&
167  ra_exe_unit.sort_info.limit &&
169  const auto only_order_entry = ra_exe_unit.sort_info.order_entries.front();
170  CHECK_GT(only_order_entry.tle_no, int(0));
171  CHECK_LE(static_cast<size_t>(only_order_entry.tle_no),
172  ra_exe_unit.target_exprs.size());
173  const auto order_entry_expr = ra_exe_unit.target_exprs[only_order_entry.tle_no - 1];
174  const auto n = ra_exe_unit.sort_info.offset + ra_exe_unit.sort_info.limit;
175  if ((order_entry_expr->get_type_info().is_number() ||
176  order_entry_expr->get_type_info().is_time()) &&
177  n <= g_streaming_topn_max) {
178  return true;
179  }
180  }
181 
182  return false;
183 }
184 
185 template <class T>
186 inline std::vector<int8_t> get_col_byte_widths(const T& col_expr_list) {
187  std::vector<int8_t> col_widths;
188  size_t col_expr_idx = 0;
189  for (const auto& col_expr : col_expr_list) {
190  if (!col_expr) {
191  // row index
192  col_widths.push_back(sizeof(int64_t));
193  } else {
194  bool is_varlen_projection{false};
195  if constexpr (std::is_same<T, std::list<std::shared_ptr<Analyzer::Expr>>>::value) {
197  !(std::dynamic_pointer_cast<const Analyzer::GeoExpr>(col_expr) == nullptr);
198  } else {
200  !(dynamic_cast<const Analyzer::GeoExpr*>(col_expr) == nullptr);
201  }
202 
203  if (is_varlen_projection) {
204  col_widths.push_back(sizeof(int64_t));
205  ++col_expr_idx;
206  continue;
207  }
208  const auto agg_info = get_target_info(col_expr, g_bigint_count);
209  const auto chosen_type = get_compact_type(agg_info);
210  if ((chosen_type.is_string() && chosen_type.get_compression() == kENCODING_NONE) ||
211  chosen_type.is_array()) {
212  col_widths.push_back(sizeof(int64_t));
213  col_widths.push_back(sizeof(int64_t));
214  ++col_expr_idx;
215  continue;
216  }
217  if (chosen_type.is_geometry()) {
218  for (auto i = 0; i < chosen_type.get_physical_coord_cols(); ++i) {
219  col_widths.push_back(sizeof(int64_t));
220  col_widths.push_back(sizeof(int64_t));
221  }
222  ++col_expr_idx;
223  continue;
224  }
225  const auto col_expr_bitwidth = get_bit_width(chosen_type);
226  CHECK_EQ(size_t(0), col_expr_bitwidth % 8);
227  col_widths.push_back(static_cast<int8_t>(col_expr_bitwidth >> 3));
228  // for average, we'll need to keep the count as well
229  if (agg_info.agg_kind == kAVG) {
230  CHECK(agg_info.is_agg);
231  col_widths.push_back(sizeof(int64_t));
232  }
233  }
234  ++col_expr_idx;
235  }
236  return col_widths;
237 }
238 
239 } // namespace
240 
241 std::unique_ptr<QueryMemoryDescriptor> QueryMemoryDescriptor::init(
242  const Executor* executor,
243  const RelAlgExecutionUnit& ra_exe_unit,
244  const std::vector<InputTableInfo>& query_infos,
245  const ColRangeInfo& col_range_info,
246  const KeylessInfo& keyless_info,
247  const bool allow_multifrag,
248  const ExecutorDeviceType device_type,
249  const int8_t crt_min_byte_width,
250  const bool sort_on_gpu_hint,
251  const size_t shard_count,
252  const size_t max_groups_buffer_entry_count,
253  RenderInfo* render_info,
254  const CountDistinctDescriptors count_distinct_descriptors,
255  const bool must_use_baseline_sort,
256  const bool output_columnar_hint,
257  const bool streaming_top_n_hint) {
258  auto group_col_widths = get_col_byte_widths(ra_exe_unit.groupby_exprs);
259  const bool is_group_by{!group_col_widths.empty()};
260 
261  auto col_slot_context = ColSlotContext(ra_exe_unit.target_exprs, {});
262 
263  const auto min_slot_size = QueryMemoryDescriptor::pick_target_compact_width(
264  ra_exe_unit, query_infos, crt_min_byte_width);
265 
266  col_slot_context.setAllSlotsPaddedSize(min_slot_size);
267  col_slot_context.validate();
268 
269  if (!is_group_by) {
270  CHECK(!must_use_baseline_sort);
271 
272  return std::make_unique<QueryMemoryDescriptor>(
273  executor,
274  ra_exe_unit,
275  query_infos,
276  allow_multifrag,
277  false,
278  false,
279  -1,
280  ColRangeInfo{ra_exe_unit.estimator ? QueryDescriptionType::Estimator
282  0,
283  0,
284  0,
285  false},
286  col_slot_context,
287  std::vector<int8_t>{},
288  /*group_col_compact_width=*/0,
289  std::vector<int64_t>{},
290  /*entry_count=*/1,
291  count_distinct_descriptors,
292  false,
293  output_columnar_hint,
294  render_info && render_info->isInSitu(),
295  must_use_baseline_sort,
296  /*use_streaming_top_n=*/false);
297  }
298 
299  size_t entry_count = 1;
300  auto actual_col_range_info = col_range_info;
301  bool interleaved_bins_on_gpu = false;
302  bool keyless_hash = false;
303  bool streaming_top_n = false;
304  int8_t group_col_compact_width = 0;
305  int32_t idx_target_as_key = -1;
306  auto output_columnar = output_columnar_hint;
307  std::vector<int64_t> target_groupby_indices;
308 
309  switch (col_range_info.hash_type_) {
311  if (render_info) {
312  // TODO(croot): this can be removed now thanks to the more centralized
313  // NonInsituQueryClassifier code, but keeping it just in case
314  render_info->setNonInSitu();
315  }
316  // keyless hash: whether or not group columns are stored at the beginning of the
317  // output buffer
318  keyless_hash =
319  (!sort_on_gpu_hint ||
321  col_range_info.max, col_range_info.min, col_range_info.bucket)) &&
322  !col_range_info.bucket && !must_use_baseline_sort && keyless_info.keyless;
323 
324  // if keyless, then this target index indicates wheter an entry is empty or not
325  // (acts as a key)
326  idx_target_as_key = keyless_info.target_index;
327 
328  if (group_col_widths.size() > 1) {
329  // col range info max contains the expected cardinality of the output
330  entry_count = static_cast<size_t>(actual_col_range_info.max);
331  actual_col_range_info.bucket = 0;
332  } else {
333  // single column perfect hash
334  entry_count = std::max(
335  GroupByAndAggregate::getBucketedCardinality(col_range_info), int64_t(1));
336  const size_t interleaved_max_threshold{512};
337 
338  if (must_use_baseline_sort) {
339  target_groupby_indices = target_expr_group_by_indices(ra_exe_unit.groupby_exprs,
340  ra_exe_unit.target_exprs);
341  col_slot_context =
342  ColSlotContext(ra_exe_unit.target_exprs, target_groupby_indices);
343  }
344 
345  bool has_varlen_sample_agg = false;
346  for (const auto& target_expr : ra_exe_unit.target_exprs) {
347  if (target_expr->get_contains_agg()) {
348  const auto agg_expr = dynamic_cast<Analyzer::AggExpr*>(target_expr);
349  CHECK(agg_expr);
350  if (agg_expr->get_aggtype() == kSAMPLE &&
351  agg_expr->get_type_info().is_varlen()) {
352  has_varlen_sample_agg = true;
353  break;
354  }
355  }
356  }
357 
358  interleaved_bins_on_gpu = keyless_hash && !has_varlen_sample_agg &&
359  (entry_count <= interleaved_max_threshold) &&
360  (device_type == ExecutorDeviceType::GPU) &&
362  count_distinct_descriptors) &&
363  !output_columnar;
364  }
365  break;
366  }
368  if (render_info) {
369  // TODO(croot): this can be removed now thanks to the more centralized
370  // NonInsituQueryClassifier code, but keeping it just in case
371  render_info->setNonInSitu();
372  }
373  entry_count = shard_count
374  ? (max_groups_buffer_entry_count + shard_count - 1) / shard_count
375  : max_groups_buffer_entry_count;
376  target_groupby_indices = target_expr_group_by_indices(ra_exe_unit.groupby_exprs,
377  ra_exe_unit.target_exprs);
378  col_slot_context = ColSlotContext(ra_exe_unit.target_exprs, target_groupby_indices);
379 
380  group_col_compact_width =
381  output_columnar ? 8
382  : pick_baseline_key_width(ra_exe_unit, query_infos, executor);
383 
384  actual_col_range_info =
386  break;
387  }
389  CHECK(!must_use_baseline_sort);
390 
391  if (streaming_top_n_hint && use_streaming_top_n(ra_exe_unit, output_columnar)) {
392  streaming_top_n = true;
393  entry_count = ra_exe_unit.sort_info.offset + ra_exe_unit.sort_info.limit;
394  } else {
395  if (ra_exe_unit.use_bump_allocator) {
396  output_columnar = false;
397  entry_count = 0;
398  } else {
399  entry_count = ra_exe_unit.scan_limit
400  ? static_cast<size_t>(ra_exe_unit.scan_limit)
401  : max_groups_buffer_entry_count;
402  }
403  }
404 
405  const auto catalog = executor->getCatalog();
406  CHECK(catalog);
407  target_groupby_indices = executor->plan_state_->allow_lazy_fetch_
408  ? target_expr_proj_indices(ra_exe_unit, *catalog)
409  : std::vector<int64_t>{};
410 
411  col_slot_context = ColSlotContext(ra_exe_unit.target_exprs, target_groupby_indices);
412  break;
413  }
414  default:
415  UNREACHABLE() << "Unknown query type";
416  }
417 
418  return std::make_unique<QueryMemoryDescriptor>(executor,
419  ra_exe_unit,
420  query_infos,
421  allow_multifrag,
422  keyless_hash,
423  interleaved_bins_on_gpu,
424  idx_target_as_key,
425  actual_col_range_info,
426  col_slot_context,
427  group_col_widths,
428  group_col_compact_width,
429  target_groupby_indices,
430  entry_count,
431  count_distinct_descriptors,
432  sort_on_gpu_hint,
433  output_columnar,
434  render_info && render_info->isInSitu(),
435  must_use_baseline_sort,
436  streaming_top_n);
437 }
438 
439 namespace {
440 template <SQLAgg... agg_types>
441 bool any_of(std::vector<Analyzer::Expr*> const& target_exprs) {
442  return boost::algorithm::any_of(target_exprs, [=](Analyzer::Expr const* expr) {
443  auto const* const agg = dynamic_cast<Analyzer::AggExpr const*>(expr);
444  return agg && (... || (agg_types == agg->get_aggtype()));
445  });
446 }
447 } // namespace
448 
450  const Executor* executor,
451  const RelAlgExecutionUnit& ra_exe_unit,
452  const std::vector<InputTableInfo>& query_infos,
453  const bool allow_multifrag,
454  const bool keyless_hash,
455  const bool interleaved_bins_on_gpu,
456  const int32_t idx_target_as_key,
457  const ColRangeInfo& col_range_info,
458  const ColSlotContext& col_slot_context,
459  const std::vector<int8_t>& group_col_widths,
460  const int8_t group_col_compact_width,
461  const std::vector<int64_t>& target_groupby_indices,
462  const size_t entry_count,
463  const CountDistinctDescriptors count_distinct_descriptors,
464  const bool sort_on_gpu_hint,
465  const bool output_columnar_hint,
466  const bool render_output,
467  const bool must_use_baseline_sort,
468  const bool use_streaming_top_n)
469  : executor_(executor)
470  , allow_multifrag_(allow_multifrag)
471  , query_desc_type_(col_range_info.hash_type_)
472  , keyless_hash_(keyless_hash)
473  , interleaved_bins_on_gpu_(interleaved_bins_on_gpu)
474  , idx_target_as_key_(idx_target_as_key)
475  , group_col_widths_(group_col_widths)
476  , group_col_compact_width_(group_col_compact_width)
477  , target_groupby_indices_(target_groupby_indices)
478  , entry_count_(entry_count)
479  , min_val_(col_range_info.min)
480  , max_val_(col_range_info.max)
481  , bucket_(col_range_info.bucket)
482  , has_nulls_(col_range_info.has_nulls)
483  , count_distinct_descriptors_(count_distinct_descriptors)
484  , output_columnar_(false)
485  , render_output_(render_output)
486  , must_use_baseline_sort_(must_use_baseline_sort)
487  , is_table_function_(false)
488  , use_streaming_top_n_(use_streaming_top_n)
489  , force_4byte_float_(false)
490  , col_slot_context_(col_slot_context) {
494 
495  sort_on_gpu_ = sort_on_gpu_hint && canOutputColumnar() && !keyless_hash_;
496  if (sort_on_gpu_) {
497  CHECK(!ra_exe_unit.use_bump_allocator);
498  output_columnar_ = true;
499  } else {
500  switch (query_desc_type_) {
502  output_columnar_ = output_columnar_hint;
503  break;
505  output_columnar_ = output_columnar_hint &&
508  !any_of<kAPPROX_QUANTILE, kMODE>(ra_exe_unit.target_exprs);
509  break;
511  output_columnar_ = output_columnar_hint;
512  break;
514  output_columnar_ = output_columnar_hint &&
517  !any_of<kAPPROX_QUANTILE, kMODE>(ra_exe_unit.target_exprs);
518  break;
519  default:
520  output_columnar_ = false;
521  break;
522  }
523  }
524 
526  // TODO(adb): Ensure fixed size buffer allocations are correct with all logical column
527  // sizes
528  CHECK(!ra_exe_unit.use_bump_allocator);
531  }
532 
533 #ifdef HAVE_CUDA
534  // Check Streaming Top N heap usage, bail if > max slab size, CUDA ONLY
535  if (use_streaming_top_n_ && executor->getDataMgr()->gpusPresent()) {
536  const auto thread_count = executor->blockSize() * executor->gridSize();
537  const auto total_buff_size =
539  if (total_buff_size > executor_->maxGpuSlabSize()) {
540  throw StreamingTopNOOM(total_buff_size);
541  }
542  }
543 #endif
544 }
545 
547  : executor_(nullptr)
548  , allow_multifrag_(false)
549  , query_desc_type_(QueryDescriptionType::Projection)
550  , keyless_hash_(false)
551  , interleaved_bins_on_gpu_(false)
552  , idx_target_as_key_(0)
553  , group_col_compact_width_(0)
554  , entry_count_(0)
555  , min_val_(0)
556  , max_val_(0)
557  , bucket_(0)
558  , has_nulls_(false)
559  , sort_on_gpu_(false)
560  , output_columnar_(false)
561  , render_output_(false)
562  , must_use_baseline_sort_(false)
563  , is_table_function_(false)
564  , use_streaming_top_n_(false)
565  , force_4byte_float_(false) {}
566 
568  const size_t entry_count,
569  const QueryDescriptionType query_desc_type,
570  const bool is_table_function)
571  : executor_(executor)
572  , allow_multifrag_(false)
573  , query_desc_type_(query_desc_type)
574  , keyless_hash_(false)
575  , interleaved_bins_on_gpu_(false)
576  , idx_target_as_key_(0)
577  , group_col_compact_width_(0)
578  , entry_count_(entry_count)
579  , min_val_(0)
580  , max_val_(0)
581  , bucket_(0)
582  , has_nulls_(false)
583  , sort_on_gpu_(false)
584  , output_columnar_(false)
585  , render_output_(false)
586  , must_use_baseline_sort_(false)
587  , is_table_function_(is_table_function)
588  , use_streaming_top_n_(false)
589  , force_4byte_float_(false) {}
590 
592  const int64_t min_val,
593  const int64_t max_val,
594  const bool has_nulls,
595  const std::vector<int8_t>& group_col_widths)
596  : executor_(nullptr)
597  , allow_multifrag_(false)
598  , query_desc_type_(query_desc_type)
599  , keyless_hash_(false)
600  , interleaved_bins_on_gpu_(false)
601  , idx_target_as_key_(0)
602  , group_col_widths_(group_col_widths)
603  , group_col_compact_width_(0)
604  , entry_count_(0)
605  , min_val_(min_val)
606  , max_val_(max_val)
607  , bucket_(0)
608  , has_nulls_(false)
609  , sort_on_gpu_(false)
610  , output_columnar_(false)
611  , render_output_(false)
612  , must_use_baseline_sort_(false)
613  , is_table_function_(false)
614  , use_streaming_top_n_(false)
615  , force_4byte_float_(false) {}
616 
618  // Note that this method does not check ptr reference members (e.g. executor_) or
619  // entry_count_
620  if (query_desc_type_ != other.query_desc_type_) {
621  return false;
622  }
623  if (keyless_hash_ != other.keyless_hash_) {
624  return false;
625  }
627  return false;
628  }
629  if (idx_target_as_key_ != other.idx_target_as_key_) {
630  return false;
631  }
632  if (force_4byte_float_ != other.force_4byte_float_) {
633  return false;
634  }
635  if (group_col_widths_ != other.group_col_widths_) {
636  return false;
637  }
639  return false;
640  }
642  return false;
643  }
644  if (min_val_ != other.min_val_) {
645  return false;
646  }
647  if (max_val_ != other.max_val_) {
648  return false;
649  }
650  if (bucket_ != other.bucket_) {
651  return false;
652  }
653  if (has_nulls_ != other.has_nulls_) {
654  return false;
655  }
657  return false;
658  } else {
659  // Count distinct descriptors can legitimately differ in device only.
660  for (size_t i = 0; i < count_distinct_descriptors_.size(); ++i) {
661  auto ref_count_distinct_desc = other.count_distinct_descriptors_[i];
662  auto count_distinct_desc = count_distinct_descriptors_[i];
663  count_distinct_desc.device_type = ref_count_distinct_desc.device_type;
664  if (ref_count_distinct_desc != count_distinct_desc) {
665  return false;
666  }
667  }
668  }
669  if (sort_on_gpu_ != other.sort_on_gpu_) {
670  return false;
671  }
672  if (output_columnar_ != other.output_columnar_) {
673  return false;
674  }
675  if (col_slot_context_ != other.col_slot_context_) {
676  return false;
677  }
678  return true;
679 }
680 
681 std::unique_ptr<QueryExecutionContext> QueryMemoryDescriptor::getQueryExecutionContext(
682  const RelAlgExecutionUnit& ra_exe_unit,
683  const Executor* executor,
684  const ExecutorDeviceType device_type,
685  const ExecutorDispatchMode dispatch_mode,
686  const int device_id,
687  const int outer_table_id,
688  const int64_t num_rows,
689  const std::vector<std::vector<const int8_t*>>& col_buffers,
690  const std::vector<std::vector<uint64_t>>& frag_offsets,
691  std::shared_ptr<RowSetMemoryOwner> row_set_mem_owner,
692  const bool output_columnar,
693  const bool sort_on_gpu,
694  const size_t thread_idx,
695  RenderInfo* render_info) const {
696  auto timer = DEBUG_TIMER(__func__);
697  if (frag_offsets.empty()) {
698  return nullptr;
699  }
700  return std::unique_ptr<QueryExecutionContext>(
701  new QueryExecutionContext(ra_exe_unit,
702  *this,
703  executor,
704  device_type,
705  dispatch_mode,
706  device_id,
707  outer_table_id,
708  num_rows,
709  col_buffers,
710  frag_offsets,
711  row_set_mem_owner,
712  output_columnar,
713  sort_on_gpu,
714  thread_idx,
715  render_info));
716 }
717 
719  const RelAlgExecutionUnit& ra_exe_unit,
720  const std::vector<InputTableInfo>& query_infos,
721  const int8_t crt_min_byte_width) {
722  if (g_bigint_count) {
723  return sizeof(int64_t);
724  }
725  int8_t compact_width{0};
726  auto col_it = ra_exe_unit.input_col_descs.begin();
727  auto const end = ra_exe_unit.input_col_descs.end();
728  int unnest_array_col_id{std::numeric_limits<int>::min()};
729  for (const auto& groupby_expr : ra_exe_unit.groupby_exprs) {
730  const auto uoper = dynamic_cast<Analyzer::UOper*>(groupby_expr.get());
731  if (uoper && uoper->get_optype() == kUNNEST) {
732  const auto& arg_ti = uoper->get_operand()->get_type_info();
733  CHECK(arg_ti.is_array());
734  const auto& elem_ti = arg_ti.get_elem_type();
735  if (elem_ti.is_string() && elem_ti.get_compression() == kENCODING_DICT) {
736  unnest_array_col_id = (*col_it)->getColId();
737  } else {
738  compact_width = crt_min_byte_width;
739  break;
740  }
741  }
742  if (col_it != end) {
743  ++col_it;
744  }
745  }
746  if (!compact_width &&
747  (ra_exe_unit.groupby_exprs.size() != 1 || !ra_exe_unit.groupby_exprs.front())) {
748  compact_width = crt_min_byte_width;
749  }
750  if (!compact_width) {
751  col_it = ra_exe_unit.input_col_descs.begin();
752  std::advance(col_it, ra_exe_unit.groupby_exprs.size());
753  for (const auto target : ra_exe_unit.target_exprs) {
754  const auto& ti = target->get_type_info();
755  const auto agg = dynamic_cast<const Analyzer::AggExpr*>(target);
756  if (agg && agg->get_arg()) {
757  compact_width = crt_min_byte_width;
758  break;
759  }
760 
761  if (agg) {
762  CHECK_EQ(kCOUNT, agg->get_aggtype());
763  CHECK(!agg->get_is_distinct());
764  if (col_it != end) {
765  ++col_it;
766  }
767  continue;
768  }
769 
770  if (is_int_and_no_bigger_than(ti, 4) ||
771  (ti.is_string() && ti.get_compression() == kENCODING_DICT)) {
772  if (col_it != end) {
773  ++col_it;
774  }
775  continue;
776  }
777 
778  const auto uoper = dynamic_cast<Analyzer::UOper*>(target);
779  if (uoper && uoper->get_optype() == kUNNEST &&
780  (*col_it)->getColId() == unnest_array_col_id) {
781  const auto arg_ti = uoper->get_operand()->get_type_info();
782  CHECK(arg_ti.is_array());
783  const auto& elem_ti = arg_ti.get_elem_type();
784  if (elem_ti.is_string() && elem_ti.get_compression() == kENCODING_DICT) {
785  if (col_it != end) {
786  ++col_it;
787  }
788  continue;
789  }
790  }
791 
792  compact_width = crt_min_byte_width;
793  break;
794  }
795  }
796  if (!compact_width) {
797  size_t total_tuples{0};
798  for (const auto& qi : query_infos) {
799  total_tuples += qi.info.getNumTuples();
800  }
801  return total_tuples <= static_cast<size_t>(std::numeric_limits<uint32_t>::max()) ||
802  unnest_array_col_id != std::numeric_limits<int>::min()
803  ? 4
804  : crt_min_byte_width;
805  } else {
806  // TODO(miyu): relax this condition to allow more cases just w/o padding
807  for (auto wid : get_col_byte_widths(ra_exe_unit.target_exprs)) {
808  compact_width = std::max(compact_width, wid);
809  }
810  return compact_width;
811  }
812 }
813 
816 }
817 
820  size_t total_bytes{0};
821  if (keyless_hash_) {
822  // ignore, there's no group column in the output buffer
824  } else {
825  total_bytes += group_col_widths_.size() * getEffectiveKeyWidth();
826  total_bytes = align_to_int64(total_bytes);
827  }
828  total_bytes += getColsSize();
829  return align_to_int64(total_bytes);
830 }
831 
833  return (interleaved_bins_on_gpu_ ? executor_->warpSize() : 1);
834 }
835 
838 }
839 
848 }
849 
855  const size_t num_entries_per_column) const {
856  return col_slot_context_.getTotalBytesOfColumnarBuffers(num_entries_per_column);
857 }
858 
869  const size_t projection_count) const {
870  constexpr size_t row_index_width = sizeof(int64_t);
871  return getTotalBytesOfColumnarBuffers(projection_count) +
872  row_index_width * projection_count;
873 }
874 
875 size_t QueryMemoryDescriptor::getColOnlyOffInBytes(const size_t col_idx) const {
876  return col_slot_context_.getColOnlyOffInBytes(col_idx);
877 }
878 
879 /*
880  * Returns the memory offset in bytes for a specific agg column in the output
881  * memory buffer. Depending on the query type, there may be some extra portion
882  * of memory prepended at the beginning of the buffer. A brief description of
883  * the memory layout is as follows:
884  * 1. projections: index column (64bit) + all target columns
885  * 2. group by: all group columns (64-bit each) + all agg columns
886  * 2a. if keyless, there is no prepending group column stored at the beginning
887  */
888 size_t QueryMemoryDescriptor::getColOffInBytes(const size_t col_idx) const {
889  const auto warp_count = getWarpCount();
890  if (output_columnar_) {
891  CHECK_EQ(size_t(1), warp_count);
892  size_t offset{0};
893  if (!keyless_hash_) {
895  }
896  if (is_table_function_) {
897  for (size_t index = 0; index < col_idx; ++index) {
898  int8_t column_width = getPaddedSlotWidthBytes(index);
899  if (column_width > 0) {
900  offset += align_to_int64(column_width * entry_count_);
901  } else {
902  int64_t flatbuffer_size = getFlatBufferSize(index);
903  CHECK_GT(flatbuffer_size, 0);
904  offset += align_to_int64(flatbuffer_size);
905  }
906  }
907  } else {
908  for (size_t index = 0; index < col_idx; ++index) {
910  }
911  }
912  return offset;
913  }
914 
915  size_t offset{0};
916  if (keyless_hash_) {
917  // ignore, there's no group column in the output buffer
919  } else {
920  offset += group_col_widths_.size() * getEffectiveKeyWidth();
921  offset = align_to_int64(offset);
922  }
923  offset += getColOnlyOffInBytes(col_idx);
924  return offset;
925 }
926 
927 /*
928  * Returns the memory offset for a particular group column in the prepended group
929  * columns portion of the memory.
930  */
932  const size_t group_idx) const {
934  CHECK(group_idx < getGroupbyColCount());
935  size_t offset{0};
936  for (size_t col_idx = 0; col_idx < group_idx; col_idx++) {
937  // TODO(Saman): relax that int64_bit part immediately
938  offset += align_to_int64(
939  std::max(groupColWidth(col_idx), static_cast<int8_t>(sizeof(int64_t))) *
940  getEntryCount());
941  }
942  return offset;
943 }
944 
945 /*
946  * Returns total amount of memory prepended at the beginning of the output memory
947  * buffer.
948  */
951  size_t buffer_size{0};
952  for (size_t group_idx = 0; group_idx < getGroupbyColCount(); group_idx++) {
953  buffer_size += align_to_int64(
954  std::max(groupColWidth(group_idx), static_cast<int8_t>(sizeof(int64_t))) *
955  getEntryCount());
956  }
957  return buffer_size;
958 }
959 
960 size_t QueryMemoryDescriptor::getColOffInBytesInNextBin(const size_t col_idx) const {
961  auto warp_count = getWarpCount();
962  if (output_columnar_) {
963  CHECK_EQ(size_t(1), group_col_widths_.size());
964  CHECK_EQ(size_t(1), warp_count);
965  return getPaddedSlotWidthBytes(col_idx);
966  }
967 
968  return warp_count * getRowSize();
969 }
970 
971 size_t QueryMemoryDescriptor::getNextColOffInBytes(const int8_t* col_ptr,
972  const size_t bin,
973  const size_t col_idx) const {
975  size_t offset{0};
976  auto warp_count = getWarpCount();
977  const auto chosen_bytes = getPaddedSlotWidthBytes(col_idx);
978  const auto total_slot_count = getSlotCount();
979  if (col_idx + 1 == total_slot_count) {
980  if (output_columnar_) {
981  return (entry_count_ - bin) * chosen_bytes;
982  } else {
983  return static_cast<size_t>(align_to_int64(col_ptr + chosen_bytes) - col_ptr);
984  }
985  }
986 
987  const auto next_chosen_bytes = getPaddedSlotWidthBytes(col_idx + 1);
988  if (output_columnar_) {
989  CHECK_EQ(size_t(1), group_col_widths_.size());
990  CHECK_EQ(size_t(1), warp_count);
991 
992  offset = align_to_int64(entry_count_ * chosen_bytes);
993 
994  offset += bin * (next_chosen_bytes - chosen_bytes);
995  return offset;
996  }
997 
998  if (next_chosen_bytes == sizeof(int64_t)) {
999  return static_cast<size_t>(align_to_int64(col_ptr + chosen_bytes) - col_ptr);
1000  } else {
1001  return chosen_bytes;
1002  }
1003 }
1004 
1006  const size_t col_idx) const {
1007  const auto chosen_bytes = getPaddedSlotWidthBytes(col_idx);
1008  const auto total_slot_count = getSlotCount();
1009  if (col_idx + 1 == total_slot_count) {
1010  return static_cast<size_t>(align_to_int64(col_ptr + chosen_bytes) - col_ptr);
1011  }
1012 
1013  const auto next_chosen_bytes = getPaddedSlotWidthBytes(col_idx + 1);
1014 
1015  if (next_chosen_bytes == sizeof(int64_t)) {
1016  return static_cast<size_t>(align_to_int64(col_ptr + chosen_bytes) - col_ptr);
1017  } else {
1018  return chosen_bytes;
1019  }
1020 }
1021 
1023  const RelAlgExecutionUnit& ra_exe_unit,
1024  const unsigned thread_count,
1025  const ExecutorDeviceType device_type) const {
1026  if (use_streaming_top_n_) {
1027  const size_t n = ra_exe_unit.sort_info.offset + ra_exe_unit.sort_info.limit;
1028  return streaming_top_n::get_heap_size(getRowSize(), n, thread_count);
1029  }
1030  return getBufferSizeBytes(device_type, entry_count_);
1031 }
1032 
1046  const size_t entry_count) const {
1047  if (keyless_hash_ && !output_columnar_) {
1048  CHECK_GE(group_col_widths_.size(), size_t(1));
1049  auto row_bytes = align_to_int64(getColsSize());
1050 
1051  return (interleavedBins(device_type) ? executor_->warpSize() : 1) * entry_count *
1052  row_bytes;
1053  }
1054 
1055  constexpr size_t row_index_width = sizeof(int64_t);
1056  size_t total_bytes{0};
1057  if (output_columnar_) {
1058  switch (query_desc_type_) {
1060  total_bytes = row_index_width * entry_count + getTotalBytesOfColumnarBuffers();
1061  break;
1063  total_bytes = getTotalBytesOfColumnarBuffers();
1064  break;
1065  default:
1066  total_bytes = sizeof(int64_t) * group_col_widths_.size() * entry_count +
1068  break;
1069  }
1070  } else {
1071  total_bytes = getRowSize() * entry_count;
1072  }
1073  return total_bytes;
1074 }
1075 
1077  const ExecutorDeviceType device_type) const {
1078  return getBufferSizeBytes(device_type, entry_count_);
1079 }
1080 
1082  output_columnar_ = val;
1085  }
1086 }
1087 
1088 /*
1089  * Indicates the query types that are currently allowed to use the logical
1090  * sized columns instead of padded sized ones.
1091  */
1093  // In distributed mode, result sets are serialized using rowwise iterators, so we use
1094  // consistent slot widths for now
1095  return output_columnar_ && !g_cluster &&
1097  query_desc_type_ == QueryDescriptionType::TableFunction);
1098 }
1099 
1101  size_t total_slot_count = col_slot_context_.getSlotCount();
1102 
1103  if (target_groupby_indices_.empty()) {
1104  return total_slot_count;
1105  }
1106  return total_slot_count - std::count_if(target_groupby_indices_.begin(),
1108  [](const int64_t i) { return i >= 0; });
1109 }
1110 
1113  getGroupbyColCount() == 1);
1114 }
1115 
1118 }
1119 
1121  if (g_cluster) {
1122  return true;
1123  }
1125  return true;
1126  }
1127  if (executor_->isCPUOnly() || render_output_ ||
1132  getGroupbyColCount() > 1)) {
1133  return true;
1134  }
1137 }
1138 
1140  return device_type == ExecutorDeviceType::GPU && !render_output_ &&
1142 }
1143 
1145  return interleaved_bins_on_gpu_ && device_type == ExecutorDeviceType::GPU;
1146 }
1147 
1148 // TODO(Saman): an implementation detail, so move this out of QMD
1150  const ExecutorDeviceType device_type) const {
1151  if (device_type == ExecutorDeviceType::GPU) {
1152  return executor_->cudaMgr()->isArchVoltaOrGreaterForAll();
1153  }
1154  return false;
1155 }
1156 
1158  return col_slot_context_.getColCount();
1159 }
1160 
1163 }
1164 
1165 const int8_t QueryMemoryDescriptor::getPaddedSlotWidthBytes(const size_t slot_idx) const {
1166  return col_slot_context_.getSlotInfo(slot_idx).padded_size;
1167 }
1168 
1170  const int8_t bytes) {
1171  col_slot_context_.setPaddedSlotWidthBytes(slot_idx, bytes);
1172 }
1173 
1175  const size_t slot_idx) const {
1176  return col_slot_context_.getSlotInfo(slot_idx).logical_size;
1177 }
1178 
1180  const size_t col_idx) const {
1181  const auto& col_slots = col_slot_context_.getSlotsForCol(col_idx);
1182  CHECK_EQ(col_slots.size(), size_t(1));
1183  return col_slots.front();
1184 }
1185 
1186 void QueryMemoryDescriptor::useConsistentSlotWidthSize(const int8_t slot_width_size) {
1187  col_slot_context_.setAllSlotsSize(slot_width_size);
1188 }
1189 
1191  // Note: Actual row size may include padding (see ResultSetBufferAccessors.h)
1193 }
1194 
1196  const int8_t actual_min_byte_width) const {
1197  return col_slot_context_.getMinPaddedByteSize(actual_min_byte_width);
1198 }
1199 
1201  const std::vector<std::tuple<int8_t, int8_t>>& slots_for_col) {
1202  col_slot_context_.addColumn(slots_for_col);
1203 }
1204 
1205 void QueryMemoryDescriptor::addColSlotInfoFlatBuffer(const int64_t flatbuffer_size) {
1206  col_slot_context_.addColumnFlatBuffer(flatbuffer_size);
1207 }
1208 
1211 }
1212 
1215 }
1216 
1221 }
1222 
1224  switch (query_desc_type_) {
1226  return "Perfect Hash";
1228  return "Baseline Hash";
1230  return "Projection";
1232  return "Table Function";
1234  return "Non-grouped Aggregate";
1236  return "Estimator";
1237  default:
1238  UNREACHABLE();
1239  }
1240  return "";
1241 }
1242 
1243 std::string QueryMemoryDescriptor::toString() const {
1244  auto str = reductionKey();
1245  str += "\tAllow Multifrag: " + ::toString(allow_multifrag_) + "\n";
1246  str += "\tInterleaved Bins on GPU: " + ::toString(interleaved_bins_on_gpu_) + "\n";
1247  str += "\tBlocks Share Memory: " + ::toString(blocksShareMemory()) + "\n";
1248  str += "\tThreads Share Memory: " + ::toString(threadsShareMemory()) + "\n";
1249  str += "\tUses Fast Group Values: " + ::toString(usesGetGroupValueFast()) + "\n";
1250  str +=
1251  "\tLazy Init Groups (GPU): " + ::toString(lazyInitGroups(ExecutorDeviceType::GPU)) +
1252  "\n";
1253  str += "\tEntry Count: " + std::to_string(entry_count_) + "\n";
1254  str += "\tMin Val (perfect hash only): " + std::to_string(min_val_) + "\n";
1255  str += "\tMax Val (perfect hash only): " + std::to_string(max_val_) + "\n";
1256  str += "\tBucket Val (perfect hash only): " + std::to_string(bucket_) + "\n";
1257  str += "\tSort on GPU: " + ::toString(sort_on_gpu_) + "\n";
1258  str += "\tUse Streaming Top N: " + ::toString(use_streaming_top_n_) + "\n";
1259  str += "\tOutput Columnar: " + ::toString(output_columnar_) + "\n";
1260  str += "\tRender Output: " + ::toString(render_output_) + "\n";
1261  str += "\tUse Baseline Sort: " + ::toString(must_use_baseline_sort_) + "\n";
1262  str += "\tIs Table Function: " + ::toString(is_table_function_) + "\n";
1263  return str;
1264 }
1265 
1267  std::string str;
1268  str += "Query Memory Descriptor State\n";
1269  str += "\tQuery Type: " + queryDescTypeToString() + "\n";
1270  str +=
1271  "\tKeyless Hash: " + ::toString(keyless_hash_) +
1272  (keyless_hash_ ? ", target index for key: " + std::to_string(getTargetIdxForKey())
1273  : "") +
1274  "\n";
1275  str += "\tEffective key width: " + std::to_string(getEffectiveKeyWidth()) + "\n";
1276  str += "\tNumber of group columns: " + std::to_string(getGroupbyColCount()) + "\n";
1277  const auto group_indices_size = targetGroupbyIndicesSize();
1278  if (group_indices_size) {
1279  std::vector<std::string> group_indices_strings;
1280  for (size_t target_idx = 0; target_idx < group_indices_size; ++target_idx) {
1281  group_indices_strings.push_back(std::to_string(getTargetGroupbyIndex(target_idx)));
1282  }
1283  str += "\tTarget group by indices: " +
1284  boost::algorithm::join(group_indices_strings, ",") + "\n";
1285  }
1286  str += "\t" + col_slot_context_.toString();
1287  return str;
1288 }
1289 
1290 std::vector<TargetInfo> target_exprs_to_infos(
1291  const std::vector<Analyzer::Expr*>& targets,
1293  std::vector<TargetInfo> target_infos;
1294  for (const auto target_expr : targets) {
1295  auto target = get_target_info(target_expr, g_bigint_count);
1296  if (query_mem_desc.getQueryDescriptionType() ==
1298  set_notnull(target, false);
1299  target.sql_type.set_notnull(false);
1300  }
1301  target_infos.push_back(target);
1302  }
1303  return target_infos;
1304 }
1305 
1307  int64_t buffer_element_size{0};
1308  for (size_t i = 0; i < col_slot_context_.getSlotCount(); i++) {
1309  try {
1310  const auto slot_element_size = col_slot_context_.varlenOutputElementSize(i);
1311  if (slot_element_size < 0) {
1312  return std::nullopt;
1313  }
1314  buffer_element_size += slot_element_size;
1315  } catch (...) {
1316  continue;
1317  }
1318  }
1319  return buffer_element_size;
1320 }
1321 
1322 size_t QueryMemoryDescriptor::varlenOutputRowSizeToSlot(const size_t slot_idx) const {
1323  int64_t buffer_element_size{0};
1325  for (size_t i = 0; i < slot_idx; i++) {
1326  try {
1327  const auto slot_element_size = col_slot_context_.varlenOutputElementSize(i);
1328  if (slot_element_size < 0) {
1329  continue;
1330  }
1331  buffer_element_size += slot_element_size;
1332  } catch (...) {
1333  continue;
1334  }
1335  }
1336  return buffer_element_size;
1337 }
size_t varlenOutputRowSizeToSlot(const size_t slot_idx) const
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)
void addColSlotInfoFlatBuffer(const int64_t flatbuffer_size)
int64_t getIntMin() const
SQLAgg
Definition: sqldefs.h:73
#define CHECK_EQ(x, y)
Definition: Logger.h:297
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)
std::vector< int64_t > target_expr_proj_indices(const RelAlgExecutionUnit &ra_exe_unit, const Catalog_Namespace::Catalog &cat)
void alignPaddedSlots(const bool sort_on_gpu)
std::string cat(Ts &&...args)
int8_t logical_size
class for a per-database catalog. also includes metadata for the current database and the current use...
Definition: Catalog.h:132
size_t getTotalBytesOfColumnarProjections(const size_t projection_count) const
int64_t getTargetGroupbyIndex(const size_t target_idx) const
ExecutorDeviceType
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, const int device_id)
std::string toString() const
bool isLogicalSizedColumnsAllowed() const
void setPaddedSlotWidthBytes(const size_t slot_idx, const int8_t bytes)
std::vector< int8_t > get_col_byte_widths(const T &col_expr_list)
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:333
void setOutputColumnar(const bool val)
const SortAlgorithm algorithm
#define CHECK_GE(x, y)
Definition: Logger.h:302
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)
size_t g_streaming_topn_max
Definition: ResultSet.cpp:51
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:301
void setAllSlotsSize(const int8_t slot_width_size)
TargetInfo get_target_info(const Analyzer::Expr *target_expr, const bool bigint_count)
Definition: TargetInfo.h:88
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 SQLTypeInfo get_compact_type(const TargetInfo &target)
bool is_varlen_projection(const Analyzer::Expr *target_expr, const SQLTypeInfo &ti)
const size_t limit
bool g_enable_columnar_output
Definition: Execute.cpp:99
int8_t groupColWidth(const size_t key_idx) const
size_t get_bit_width(const SQLTypeInfo &ti)
void addColumnFlatBuffer(const int64_t flatbuffer_size)
const ColumnDescriptor * get_column_descriptor_maybe(const int col_id, const int table_id, const Catalog_Namespace::Catalog &cat)
Definition: Execute.h:220
std::vector< CountDistinctDescriptor > CountDistinctDescriptors
Definition: CountDistinct.h:34
size_t getCompactByteWidth() const
Provides column info and slot info for the output buffer and some metadata helpers.
size_t getGroupbyColCount() const
bool is_integer() const
Definition: sqltypes.h:578
std::unique_ptr< QueryExecutionContext > getQueryExecutionContext(const RelAlgExecutionUnit &, const Executor *executor, const ExecutorDeviceType device_type, const ExecutorDispatchMode dispatch_mode, const int device_id, const int outer_table_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, const size_t thread_idx, RenderInfo *) const
bool lazyInitGroups(const ExecutorDeviceType) const
size_t targetGroupbyIndicesSize() const
size_t getPrependedGroupBufferSizeInBytes() const
size_t getTotalBytesOfColumnarBuffers() const
std::vector< int64_t > target_groupby_indices_
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)
bool g_bigint_count
CountDistinctDescriptors count_distinct_descriptors_
bool is_valid_int32_range(const ExpressionRange &range)
void validate() const
int get_varno() const
Definition: Analyzer.h:284
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
int64_t varlenOutputElementSize(const size_t slot_idx) const
const SQLTypeInfo & get_type_info() const
Definition: Analyzer.h:83
QueryDescriptionType getQueryDescriptionType() const
std::vector< int64_t > target_expr_group_by_indices(const std::list< std::shared_ptr< Analyzer::Expr >> &groupby_exprs, const std::vector< Analyzer::Expr * > &target_exprs)
std::optional< size_t > varlenOutputBufferElemSize() const
void addColumn(const std::vector< std::tuple< int8_t, int8_t >> &slots_for_col)
#define CHECK_LT(x, y)
Definition: Logger.h:299
#define CHECK_LE(x, y)
Definition: Logger.h:300
size_t getNextColOffInBytesRowOnly(const int8_t *col_ptr, const size_t col_idx) const
const Expr * get_operand() const
Definition: Analyzer.h:380
QueryDescriptionType query_desc_type_
int8_t padded_size
Definition: sqldefs.h:78
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:79
#define CHECK(condition)
Definition: Logger.h:289
#define DEBUG_TIMER(name)
Definition: Logger.h:407
size_t getColCount() const
std::vector< int8_t > group_col_widths_
#define EMPTY_KEY_32
QueryDescriptionType
Definition: Types.h:29
bool g_cluster
void setPaddedSlotWidthBytes(const size_t slot_idx, const int8_t bytes)
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::string queryDescTypeToString() const
bool any_of(std::vector< Analyzer::Expr * > const &target_exprs)
std::list< std::shared_ptr< const InputColDescriptor > > input_col_descs
constexpr double n
Definition: Utm.h:38
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)
int64_t getFlatBufferSize(const size_t slot_idx) const
const int8_t getSlotIndexForSingleSlotCol(const size_t col_idx) const
const int8_t getLogicalSlotWidthBytes(const size_t slot_idx) const
Definition: sqldefs.h:74
size_t getColOffInBytes(const size_t col_idx) const
size_t getColOffInBytesInNextBin(const size_t col_idx) const
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