OmniSciDB  04ee39c94c
QueryExecutionContext Class Reference

#include <QueryExecutionContext.h>

+ Inheritance diagram for QueryExecutionContext:
+ Collaboration diagram for QueryExecutionContext:

Public Member Functions

 QueryExecutionContext (const RelAlgExecutionUnit &ra_exe_unit, const QueryMemoryDescriptor &, 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 > row_set_mem_owner, const bool output_columnar, const bool sort_on_gpu, RenderInfo *)
 
ResultSetPtr getRowSet (const RelAlgExecutionUnit &ra_exe_unit, const QueryMemoryDescriptor &query_mem_desc) const
 
ResultSetPtr groupBufferToResults (const size_t i) const
 
std::vector< int64_t * > launchGpuCode (const RelAlgExecutionUnit &ra_exe_unit, const std::vector< std::pair< void *, void *>> &cu_functions, const bool hoist_literals, const std::vector< int8_t > &literal_buff, std::vector< std::vector< const int8_t *>> col_buffers, const std::vector< std::vector< int64_t >> &num_rows, const std::vector< std::vector< uint64_t >> &frag_row_offsets, const int32_t scan_limit, Data_Namespace::DataMgr *data_mgr, const unsigned block_size_x, const unsigned grid_size_x, const int device_id, int32_t *error_code, const uint32_t num_tables, const std::vector< int64_t > &join_hash_tables, RenderAllocatorMap *render_allocator_map)
 
std::vector< int64_t * > launchCpuCode (const RelAlgExecutionUnit &ra_exe_unit, const std::vector< std::pair< void *, void *>> &fn_ptrs, const bool hoist_literals, const std::vector< int8_t > &literal_buff, std::vector< std::vector< const int8_t *>> col_buffers, const std::vector< std::vector< int64_t >> &num_rows, const std::vector< std::vector< uint64_t >> &frag_row_offsets, const int32_t scan_limit, int32_t *error_code, const uint32_t num_tables, const std::vector< int64_t > &join_hash_tables)
 
int64_t getAggInitValForIndex (const size_t index) const
 

Private Member Functions

ResultSetPtr groupBufferToDeinterleavedResults (const size_t i) const
 

Private Attributes

std::unique_ptr< CudaAllocatorgpu_allocator_
 
const QueryMemoryDescriptor query_mem_desc_
 
const Executorexecutor_
 
const ExecutorDeviceType device_type_
 
const ExecutorDispatchMode dispatch_mode_
 
std::shared_ptr< RowSetMemoryOwnerrow_set_mem_owner_
 
const bool output_columnar_
 
std::unique_ptr< QueryMemoryInitializerquery_buffers_
 
std::unique_ptr< ResultSetestimator_result_set_
 

Friends

class Executor
 
template<typename META_CLASS_TYPE >
class AggregateReductionEgress
 

Detailed Description

Definition at line 35 of file QueryExecutionContext.h.

Constructor & Destructor Documentation

◆ QueryExecutionContext()

QueryExecutionContext::QueryExecutionContext ( const RelAlgExecutionUnit ra_exe_unit,
const QueryMemoryDescriptor query_mem_desc,
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 row_set_mem_owner,
const bool  output_columnar,
const bool  sort_on_gpu,
RenderInfo render_info 
)

Definition at line 29 of file QueryExecutionContext.cpp.

References CHECK, GPU, gpu_allocator_, RenderInfo::isPotentialInSituRender(), num_rows, query_buffers_, RenderInfo::render_allocator_map_ptr, and sort_on_gpu().

43  : query_mem_desc_(query_mem_desc)
44  , executor_(executor)
45  , device_type_(device_type)
46  , dispatch_mode_(dispatch_mode)
47  , row_set_mem_owner_(row_set_mem_owner)
48  , output_columnar_(output_columnar) {
49  CHECK(executor);
50  auto& data_mgr = executor->catalog_->getDataMgr();
51  if (device_type == ExecutorDeviceType::GPU) {
52  gpu_allocator_ = std::make_unique<CudaAllocator>(&data_mgr, device_id);
53  }
54 
55  auto render_allocator_map = render_info && render_info->isPotentialInSituRender()
56  ? render_info->render_allocator_map_ptr.get()
57  : nullptr;
58  query_buffers_ = std::make_unique<QueryMemoryInitializer>(ra_exe_unit,
59  query_mem_desc,
60  device_id,
61  device_type,
62  dispatch_mode,
63  output_columnar,
65  num_rows,
66  col_buffers,
67  frag_offsets,
68  render_allocator_map,
69  render_info,
70  row_set_mem_owner,
71  gpu_allocator_.get(),
72  executor);
73 }
const int8_t const int64_t * num_rows
const ExecutorDispatchMode dispatch_mode_
const ExecutorDeviceType device_type_
std::unique_ptr< QueryMemoryInitializer > query_buffers_
bool isPotentialInSituRender() const
Definition: RenderInfo.cpp:55
std::shared_ptr< RowSetMemoryOwner > row_set_mem_owner_
std::unique_ptr< RenderAllocatorMap > render_allocator_map_ptr
Definition: RenderInfo.h:32
std::unique_ptr< CudaAllocator > gpu_allocator_
#define CHECK(condition)
Definition: Logger.h:187
const QueryMemoryDescriptor query_mem_desc_
void sort_on_gpu(int64_t *val_buff, int32_t *key_buff, const uint64_t entry_count, const bool desc, const uint32_t chosen_bytes, ThrustAllocator &alloc)
+ Here is the call graph for this function:

Member Function Documentation

◆ getAggInitValForIndex()

int64_t QueryExecutionContext::getAggInitValForIndex ( const size_t  index) const

Definition at line 126 of file QueryExecutionContext.cpp.

References CHECK, and query_buffers_.

Referenced by AggregateReductionEgress< META_TYPE_CLASS >::operator()(), and AggregateReductionEgress< Experimental::MetaTypeClass< Experimental::Geometry > >::operator()().

126  {
128  return query_buffers_->getAggInitValForIndex(index);
129 }
std::unique_ptr< QueryMemoryInitializer > query_buffers_
#define CHECK(condition)
Definition: Logger.h:187
+ Here is the caller graph for this function:

◆ getRowSet()

ResultSetPtr QueryExecutionContext::getRowSet ( const RelAlgExecutionUnit ra_exe_unit,
const QueryMemoryDescriptor query_mem_desc 
) const

Definition at line 131 of file QueryExecutionContext.cpp.

References CHECK, CHECK_EQ, CPU, device_type_, executor_, GPU, groupBufferToResults(), query_buffers_, query_mem_desc_, row_set_mem_owner_, and QueryMemoryDescriptor::threadsShareMemory().

Referenced by Executor::executePlanWithGroupBy().

133  {
134  std::vector<std::pair<ResultSetPtr, std::vector<size_t>>> results_per_sm;
136  const auto group_by_buffers_size = query_buffers_->getNumBuffers();
138  CHECK_EQ(size_t(1), group_by_buffers_size);
139  return groupBufferToResults(0);
140  }
141  size_t step{query_mem_desc_.threadsShareMemory() ? executor_->blockSize() : 1};
142  for (size_t i = 0; i < group_by_buffers_size; i += step) {
143  results_per_sm.emplace_back(groupBufferToResults(i), std::vector<size_t>{});
144  }
146  return executor_->reduceMultiDeviceResults(
147  ra_exe_unit, results_per_sm, row_set_mem_owner_, query_mem_desc);
148 }
#define CHECK_EQ(x, y)
Definition: Logger.h:195
const ExecutorDeviceType device_type_
std::unique_ptr< QueryMemoryInitializer > query_buffers_
std::shared_ptr< RowSetMemoryOwner > row_set_mem_owner_
ResultSetPtr groupBufferToResults(const size_t i) const
#define CHECK(condition)
Definition: Logger.h:187
const QueryMemoryDescriptor query_mem_desc_
+ Here is the call graph for this function:
+ Here is the caller graph for this function:

◆ groupBufferToDeinterleavedResults()

ResultSetPtr QueryExecutionContext::groupBufferToDeinterleavedResults ( const size_t  i) const
private

Definition at line 75 of file QueryExecutionContext.cpp.

References agg_col_count, CHECK, CPU, executor_, ResultSet::fixupQueryMemoryDescriptor(), QueryMemoryDescriptor::getColOffInBytes(), QueryMemoryDescriptor::getColOffInBytesInNextBin(), QueryMemoryDescriptor::getSlotCount(), output_columnar_, query_buffers_, query_mem_desc_, ResultSetStorage::reduceSingleRow(), and row_set_mem_owner_.

Referenced by groupBufferToResults().

76  {
78  const auto& result_set = query_buffers_->getResultSet(i);
79  auto deinterleaved_query_mem_desc =
81  deinterleaved_query_mem_desc.setHasInterleavedBinsOnGpu(false);
82  deinterleaved_query_mem_desc.useConsistentSlotWidthSize(8);
83 
84  auto deinterleaved_result_set =
85  std::make_shared<ResultSet>(result_set->getTargetInfos(),
86  std::vector<ColumnLazyFetchInfo>{},
87  std::vector<std::vector<const int8_t*>>{},
88  std::vector<std::vector<int64_t>>{},
89  std::vector<int64_t>{},
91  -1,
92  deinterleaved_query_mem_desc,
94  executor_);
95  auto deinterleaved_storage =
96  deinterleaved_result_set->allocateStorage(executor_->plan_state_->init_agg_vals_);
97  auto deinterleaved_buffer =
98  reinterpret_cast<int64_t*>(deinterleaved_storage->getUnderlyingBuffer());
99  const auto rows_ptr = result_set->getStorage()->getUnderlyingBuffer();
100  size_t deinterleaved_buffer_idx = 0;
102  for (size_t bin_base_off = query_mem_desc_.getColOffInBytes(0), bin_idx = 0;
103  bin_idx < result_set->entryCount();
104  ++bin_idx, bin_base_off += query_mem_desc_.getColOffInBytesInNextBin(0)) {
105  std::vector<int64_t> agg_vals(agg_col_count, 0);
106  memcpy(&agg_vals[0],
107  &executor_->plan_state_->init_agg_vals_[0],
108  agg_col_count * sizeof(agg_vals[0]));
109  ResultSetStorage::reduceSingleRow(rows_ptr + bin_base_off,
110  executor_->warpSize(),
111  false,
112  true,
113  agg_vals,
115  result_set->getTargetInfos(),
116  executor_->plan_state_->init_agg_vals_);
117  for (size_t agg_idx = 0; agg_idx < agg_col_count;
118  ++agg_idx, ++deinterleaved_buffer_idx) {
119  deinterleaved_buffer[deinterleaved_buffer_idx] = agg_vals[agg_idx];
120  }
121  }
122  query_buffers_->resetResultSet(i);
123  return deinterleaved_result_set;
124 }
const int64_t const uint32_t const uint32_t const uint32_t agg_col_count
std::unique_ptr< QueryMemoryInitializer > query_buffers_
std::shared_ptr< RowSetMemoryOwner > row_set_mem_owner_
size_t getColOffInBytesInNextBin(const size_t col_idx) const
static QueryMemoryDescriptor fixupQueryMemoryDescriptor(const QueryMemoryDescriptor &)
Definition: ResultSet.cpp:452
#define CHECK(condition)
Definition: Logger.h:187
size_t getColOffInBytes(const size_t col_idx) const
static bool reduceSingleRow(const int8_t *row_ptr, const int8_t warp_count, const bool is_columnar, const bool replace_bitmap_ptr_with_bitmap_sz, std::vector< int64_t > &agg_vals, const QueryMemoryDescriptor &query_mem_desc, const std::vector< TargetInfo > &targets, const std::vector< int64_t > &agg_init_vals)
const QueryMemoryDescriptor query_mem_desc_
+ Here is the call graph for this function:
+ Here is the caller graph for this function:

◆ groupBufferToResults()

ResultSetPtr QueryExecutionContext::groupBufferToResults ( const size_t  i) const

Definition at line 150 of file QueryExecutionContext.cpp.

References device_type_, groupBufferToDeinterleavedResults(), QueryMemoryDescriptor::interleavedBins(), query_buffers_, and query_mem_desc_.

Referenced by getRowSet().

150  {
153  }
154  return query_buffers_->getResultSetOwned(i);
155 }
const ExecutorDeviceType device_type_
std::unique_ptr< QueryMemoryInitializer > query_buffers_
bool interleavedBins(const ExecutorDeviceType) const
ResultSetPtr groupBufferToDeinterleavedResults(const size_t i) const
const QueryMemoryDescriptor query_mem_desc_
+ Here is the call graph for this function:
+ Here is the caller graph for this function:

◆ launchCpuCode()

std::vector< int64_t * > QueryExecutionContext::launchCpuCode ( const RelAlgExecutionUnit ra_exe_unit,
const std::vector< std::pair< void *, void *>> &  fn_ptrs,
const bool  hoist_literals,
const std::vector< int8_t > &  literal_buff,
std::vector< std::vector< const int8_t *>>  col_buffers,
const std::vector< std::vector< int64_t >> &  num_rows,
const std::vector< std::vector< uint64_t >> &  frag_row_offsets,
const int32_t  scan_limit,
int32_t *  error_code,
const uint32_t  num_tables,
const std::vector< int64_t > &  join_hash_tables 
)

Definition at line 541 of file QueryExecutionContext.cpp.

References align_to_int64(), CHECK, CHECK_EQ, checkCudaErrors(), compact_init_vals(), copy_to_gpu(), CPU, QueryMemoryDescriptor::didOutputColumnar(), error_code, RelAlgExecutionUnit::estimator, estimator_result_set_, executor_, g_dynamic_watchdog_time_limit, QueryMemoryDescriptor::getColsSize(), QueryMemoryDescriptor::getQueryDescriptionType(), gpu_allocator_, logger::INFO, INJECT_TIMER, QueryMemoryDescriptor::isGroupBy(), join_hash_tables, LOG, max_matched, num_rows, output_columnar_, Projection, query_buffers_, query_mem_desc_, to_string(), total_matched, use_streaming_top_n(), and VLOG.

Referenced by Executor::executePlanWithGroupBy(), and Executor::executePlanWithoutGroupBy().

552  {
553  INJECT_TIMER(lauchCpuCode);
554 
556  const auto& init_agg_vals = query_buffers_->init_agg_vals_;
557 
558  std::vector<const int8_t**> multifrag_col_buffers;
559  for (auto& col_buffer : col_buffers) {
560  multifrag_col_buffers.push_back(&col_buffer[0]);
561  }
562  const int8_t*** multifrag_cols_ptr{
563  multifrag_col_buffers.empty() ? nullptr : &multifrag_col_buffers[0]};
564  const uint64_t num_fragments =
565  multifrag_cols_ptr ? static_cast<uint64_t>(col_buffers.size()) : uint64_t(0);
566  const auto num_out_frags = multifrag_cols_ptr ? num_fragments : uint64_t(0);
567 
568  const bool is_group_by{query_mem_desc_.isGroupBy()};
569  std::vector<int64_t*> out_vec;
570  if (ra_exe_unit.estimator) {
571  estimator_result_set_.reset(
572  new ResultSet(ra_exe_unit.estimator, ExecutorDeviceType::CPU, 0, nullptr));
573  out_vec.push_back(
574  reinterpret_cast<int64_t*>(estimator_result_set_->getHostEstimatorBuffer()));
575  } else {
576  if (!is_group_by) {
577  for (size_t i = 0; i < init_agg_vals.size(); ++i) {
578  auto buff = new int64_t[num_out_frags];
579  out_vec.push_back(static_cast<int64_t*>(buff));
580  }
581  }
582  }
583 
584  CHECK_EQ(num_rows.size(), col_buffers.size());
585  std::vector<int64_t> flatened_num_rows;
586  for (auto& nums : num_rows) {
587  flatened_num_rows.insert(flatened_num_rows.end(), nums.begin(), nums.end());
588  }
589  std::vector<uint64_t> flatened_frag_offsets;
590  for (auto& offsets : frag_offsets) {
591  flatened_frag_offsets.insert(
592  flatened_frag_offsets.end(), offsets.begin(), offsets.end());
593  }
594  int64_t rowid_lookup_num_rows{*error_code ? *error_code + 1 : 0};
595  auto num_rows_ptr =
596  rowid_lookup_num_rows ? &rowid_lookup_num_rows : &flatened_num_rows[0];
597  int32_t total_matched_init{0};
598 
599  std::vector<int64_t> cmpt_val_buff;
600  if (is_group_by) {
601  cmpt_val_buff =
603  init_agg_vals,
605  }
606 
607  const int64_t* join_hash_tables_ptr =
608  join_hash_tables.size() == 1
609  ? reinterpret_cast<int64_t*>(join_hash_tables[0])
610  : (join_hash_tables.size() > 1 ? &join_hash_tables[0] : nullptr);
611  if (hoist_literals) {
612  using agg_query = void (*)(const int8_t***, // col_buffers
613  const uint64_t*, // num_fragments
614  const int8_t*, // literals
615  const int64_t*, // num_rows
616  const uint64_t*, // frag_row_offsets
617  const int32_t*, // max_matched
618  int32_t*, // total_matched
619  const int64_t*, // init_agg_value
620  int64_t**, // out
621  int32_t*, // error_code
622  const uint32_t*, // num_tables
623  const int64_t*); // join_hash_tables_ptr
624  if (is_group_by) {
625  reinterpret_cast<agg_query>(fn_ptrs[0].first)(
626  multifrag_cols_ptr,
627  &num_fragments,
628  &literal_buff[0],
629  num_rows_ptr,
630  &flatened_frag_offsets[0],
631  &scan_limit,
632  &total_matched_init,
633  &cmpt_val_buff[0],
634  query_buffers_->getGroupByBuffersPtr(),
635  error_code,
636  &num_tables,
637  join_hash_tables_ptr);
638  } else {
639  reinterpret_cast<agg_query>(fn_ptrs[0].first)(multifrag_cols_ptr,
640  &num_fragments,
641  &literal_buff[0],
642  num_rows_ptr,
643  &flatened_frag_offsets[0],
644  &scan_limit,
645  &total_matched_init,
646  &init_agg_vals[0],
647  &out_vec[0],
648  error_code,
649  &num_tables,
650  join_hash_tables_ptr);
651  }
652  } else {
653  using agg_query = void (*)(const int8_t***, // col_buffers
654  const uint64_t*, // num_fragments
655  const int64_t*, // num_rows
656  const uint64_t*, // frag_row_offsets
657  const int32_t*, // max_matched
658  int32_t*, // total_matched
659  const int64_t*, // init_agg_value
660  int64_t**, // out
661  int32_t*, // error_code
662  const uint32_t*, // num_tables
663  const int64_t*); // join_hash_tables_ptr
664  if (is_group_by) {
665  reinterpret_cast<agg_query>(fn_ptrs[0].first)(
666  multifrag_cols_ptr,
667  &num_fragments,
668  num_rows_ptr,
669  &flatened_frag_offsets[0],
670  &scan_limit,
671  &total_matched_init,
672  &cmpt_val_buff[0],
673  query_buffers_->getGroupByBuffersPtr(),
674  error_code,
675  &num_tables,
676  join_hash_tables_ptr);
677  } else {
678  reinterpret_cast<agg_query>(fn_ptrs[0].first)(multifrag_cols_ptr,
679  &num_fragments,
680  num_rows_ptr,
681  &flatened_frag_offsets[0],
682  &scan_limit,
683  &total_matched_init,
684  &init_agg_vals[0],
685  &out_vec[0],
686  error_code,
687  &num_tables,
688  join_hash_tables_ptr);
689  }
690  }
691 
692  if (ra_exe_unit.estimator) {
693  return {};
694  }
695 
696  if (rowid_lookup_num_rows && *error_code < 0) {
697  *error_code = 0;
698  }
699 
701  query_buffers_->applyStreamingTopNOffsetCpu(query_mem_desc_, ra_exe_unit);
702  }
703 
706  query_buffers_->compactProjectionBuffersCpu(query_mem_desc_, total_matched_init);
707  }
708 
709  return out_vec;
710 }
#define CHECK_EQ(x, y)
Definition: Logger.h:195
bool use_streaming_top_n(const RelAlgExecutionUnit &ra_exe_unit, const bool output_columnar)
const int8_t const int64_t const uint64_t const int32_t const int64_t int64_t uint32_t const int64_t * join_hash_tables
const int8_t const int64_t * num_rows
std::unique_ptr< QueryMemoryInitializer > query_buffers_
#define INJECT_TIMER(DESC)
Definition: measure.h:91
std::vector< int64_t > compact_init_vals(const size_t cmpt_size, const std::vector< int64_t > &init_vec, const QueryMemoryDescriptor &query_mem_desc)
const int8_t const int64_t const uint64_t const int32_t const int64_t int64_t uint32_t const int64_t int32_t * error_code
const std::shared_ptr< Analyzer::Estimator > estimator
#define CHECK(condition)
Definition: Logger.h:187
std::unique_ptr< ResultSet > estimator_result_set_
QueryDescriptionType getQueryDescriptionType() const
const QueryMemoryDescriptor query_mem_desc_
FORCE_INLINE HOST DEVICE T align_to_int64(T addr)
+ Here is the call graph for this function:
+ Here is the caller graph for this function:

◆ launchGpuCode()

std::vector< int64_t * > QueryExecutionContext::launchGpuCode ( const RelAlgExecutionUnit ra_exe_unit,
const std::vector< std::pair< void *, void *>> &  cu_functions,
const bool  hoist_literals,
const std::vector< int8_t > &  literal_buff,
std::vector< std::vector< const int8_t *>>  col_buffers,
const std::vector< std::vector< int64_t >> &  num_rows,
const std::vector< std::vector< uint64_t >> &  frag_row_offsets,
const int32_t  scan_limit,
Data_Namespace::DataMgr data_mgr,
const unsigned  block_size_x,
const unsigned  grid_size_x,
const int  device_id,
int32_t *  error_code,
const uint32_t  num_tables,
const std::vector< int64_t > &  join_hash_tables,
RenderAllocatorMap render_allocator_map 
)

Definition at line 178 of file QueryExecutionContext.cpp.

References agg_col_count, CHECK, CHECK_EQ, checkCudaErrors(), copy_from_gpu(), copy_to_gpu(), QueryMemoryDescriptor::didOutputColumnar(), dispatch_mode_, RelAlgExecutionUnit::estimator, estimator_result_set_, executor_, g_enable_dynamic_watchdog, get_num_allocated_rows_from_gpu(), QueryMemoryDescriptor::getEntryCount(), QueryMemoryDescriptor::getQueryDescriptionType(), RenderAllocatorMap::getRenderAllocator(), GPU, gpu_allocator_, QueryMemoryDescriptor::hasKeylessHash(), INJECT_TIMER, inplace_sort_gpu(), QueryMemoryDescriptor::isGroupBy(), max_matched, num_rows, SortInfo::order_entries, output_columnar_, Projection, query_buffers_, query_mem_desc_, QueryMemoryDescriptor::sharedMemBytes(), RelAlgExecutionUnit::sort_info, QueryMemoryDescriptor::sortOnGpu(), to_string(), RelAlgExecutionUnit::use_bump_allocator, use_speculative_top_n(), use_streaming_top_n(), and VLOG.

Referenced by Executor::executePlanWithGroupBy(), and Executor::executePlanWithoutGroupBy().

194  {
195  INJECT_TIMER(lauchGpuCode);
196 #ifdef HAVE_CUDA
199  const auto& init_agg_vals = query_buffers_->init_agg_vals_;
200 
201  bool is_group_by{query_mem_desc_.isGroupBy()};
202 
203  RenderAllocator* render_allocator = nullptr;
204  if (render_allocator_map) {
205  render_allocator = render_allocator_map->getRenderAllocator(device_id);
206  }
207 
208  auto cu_func = static_cast<CUfunction>(cu_functions[device_id].first);
209  std::vector<int64_t*> out_vec;
210  uint32_t num_fragments = col_buffers.size();
211  std::vector<int32_t> error_codes(grid_size_x * block_size_x);
212 
213  CUevent start0, stop0; // preparation
214  cuEventCreate(&start0, 0);
215  cuEventCreate(&stop0, 0);
216  CUevent start1, stop1; // cuLaunchKernel
217  cuEventCreate(&start1, 0);
218  cuEventCreate(&stop1, 0);
219  CUevent start2, stop2; // finish
220  cuEventCreate(&start2, 0);
221  cuEventCreate(&stop2, 0);
222 
224  cuEventRecord(start0, 0);
225  }
226 
228  initializeDynamicWatchdog(cu_functions[device_id].second, device_id);
229  }
230 
231  auto kernel_params = prepareKernelParams(col_buffers,
232  literal_buff,
233  num_rows,
234  frag_offsets,
235  scan_limit,
236  init_agg_vals,
237  error_codes,
238  num_tables,
240  data_mgr,
241  device_id,
242  hoist_literals,
243  is_group_by);
244 
245  CHECK_EQ(static_cast<size_t>(KERN_PARAM_COUNT), kernel_params.size());
246  CHECK_EQ(CUdeviceptr(0), kernel_params[GROUPBY_BUF]);
247 
248  const unsigned block_size_y = 1;
249  const unsigned block_size_z = 1;
250  const unsigned grid_size_y = 1;
251  const unsigned grid_size_z = 1;
252  const auto total_thread_count = block_size_x * grid_size_x;
253  const auto err_desc = kernel_params[ERROR_CODE];
254 
255  if (is_group_by) {
256  CHECK(!(query_buffers_->getGroupByBuffersSize() == 0) || render_allocator);
257  bool can_sort_on_gpu = query_mem_desc_.sortOnGpu();
258  auto gpu_group_by_buffers =
259  query_buffers_->createAndInitializeGroupByBufferGpu(ra_exe_unit,
261  kernel_params[INIT_AGG_VALS],
262  device_id,
264  block_size_x,
265  grid_size_x,
266  executor_->warpSize(),
267  can_sort_on_gpu,
269  render_allocator);
270  if (ra_exe_unit.use_bump_allocator) {
271  const auto max_matched = static_cast<int32_t>(gpu_group_by_buffers.entry_count);
272  copy_to_gpu(data_mgr,
273  kernel_params[MAX_MATCHED],
274  &max_matched,
275  sizeof(max_matched),
276  device_id);
277  }
278 
279  kernel_params[GROUPBY_BUF] = gpu_group_by_buffers.first;
280  std::vector<void*> param_ptrs;
281  for (auto& param : kernel_params) {
282  param_ptrs.push_back(&param);
283  }
284 
286  cuEventRecord(stop0, 0);
287  cuEventSynchronize(stop0);
288  float milliseconds0 = 0;
289  cuEventElapsedTime(&milliseconds0, start0, stop0);
290  VLOG(1) << "Device " << std::to_string(device_id)
291  << ": launchGpuCode: group-by prepare: " << std::to_string(milliseconds0)
292  << " ms";
293  cuEventRecord(start1, 0);
294  }
295 
296  if (hoist_literals) {
298  cuLaunchKernel(cu_func,
299  grid_size_x,
300  grid_size_y,
301  grid_size_z,
302  block_size_x,
303  block_size_y,
304  block_size_z,
306  nullptr,
307  &param_ptrs[0],
308  nullptr));
309  } else {
310  param_ptrs.erase(param_ptrs.begin() + LITERALS); // TODO(alex): remove
312  cuLaunchKernel(cu_func,
313  grid_size_x,
314  grid_size_y,
315  grid_size_z,
316  block_size_x,
317  block_size_y,
318  block_size_z,
320  nullptr,
321  &param_ptrs[0],
322  nullptr));
323  }
325  executor_->registerActiveModule(cu_functions[device_id].second, device_id);
326  cuEventRecord(stop1, 0);
327  cuEventSynchronize(stop1);
328  executor_->unregisterActiveModule(cu_functions[device_id].second, device_id);
329  float milliseconds1 = 0;
330  cuEventElapsedTime(&milliseconds1, start1, stop1);
331  VLOG(1) << "Device " << std::to_string(device_id)
332  << ": launchGpuCode: group-by cuLaunchKernel: "
333  << std::to_string(milliseconds1) << " ms";
334  cuEventRecord(start2, 0);
335  }
336 
337  gpu_allocator_->copyFromDevice(reinterpret_cast<int8_t*>(error_codes.data()),
338  reinterpret_cast<int8_t*>(err_desc),
339  error_codes.size() * sizeof(error_codes[0]));
340  *error_code = aggregate_error_codes(error_codes);
341  if (*error_code > 0) {
342  return {};
343  }
344 
345  if (!render_allocator) {
347  query_buffers_->applyStreamingTopNOffsetGpu(data_mgr,
349  gpu_group_by_buffers,
350  ra_exe_unit,
351  total_thread_count,
352  device_id);
353  } else {
354  if (use_speculative_top_n(ra_exe_unit, query_mem_desc_)) {
357  gpu_group_by_buffers,
358  data_mgr,
359  device_id);
360  }
364  query_buffers_->compactProjectionBuffersGpu(
366  data_mgr,
367  gpu_group_by_buffers,
369  data_mgr, kernel_params[TOTAL_MATCHED], device_id),
370  device_id);
371  } else {
372  size_t num_allocated_rows{0};
373  if (ra_exe_unit.use_bump_allocator) {
374  num_allocated_rows = get_num_allocated_rows_from_gpu(
375  data_mgr, kernel_params[TOTAL_MATCHED], device_id);
376  // First, check the error code. If we ran out of slots, don't copy data back
377  // into the ResultSet or update ResultSet entry count
378  if (*error_code < 0) {
379  return {};
380  }
381  }
382  query_buffers_->copyGroupByBuffersFromGpu(
383  data_mgr,
385  ra_exe_unit.use_bump_allocator ? num_allocated_rows
387  gpu_group_by_buffers,
388  ra_exe_unit,
389  block_size_x,
390  grid_size_x,
391  device_id,
392  can_sort_on_gpu && query_mem_desc_.hasKeylessHash());
393  if (num_allocated_rows) {
394  CHECK(ra_exe_unit.use_bump_allocator);
395  CHECK(!query_buffers_->result_sets_.empty());
396  query_buffers_->result_sets_.front()->updateStorageEntryCount(
397  num_allocated_rows);
398  }
399  }
400  } else {
401  query_buffers_->copyGroupByBuffersFromGpu(
402  data_mgr,
405  gpu_group_by_buffers,
406  ra_exe_unit,
407  block_size_x,
408  grid_size_x,
409  device_id,
410  can_sort_on_gpu && query_mem_desc_.hasKeylessHash());
411  }
412  }
413  }
414  } else {
415  std::vector<CUdeviceptr> out_vec_dev_buffers;
416  const size_t agg_col_count{ra_exe_unit.estimator ? size_t(1) : init_agg_vals.size()};
417  if (ra_exe_unit.estimator) {
418  estimator_result_set_.reset(new ResultSet(
419  ra_exe_unit.estimator, ExecutorDeviceType::GPU, device_id, data_mgr));
420  out_vec_dev_buffers.push_back(reinterpret_cast<CUdeviceptr>(
421  estimator_result_set_->getDeviceEstimatorBuffer()));
422  } else {
423  for (size_t i = 0; i < agg_col_count; ++i) {
424  CUdeviceptr out_vec_dev_buffer =
425  num_fragments
426  ? reinterpret_cast<CUdeviceptr>(gpu_allocator_->alloc(
427  block_size_x * grid_size_x * sizeof(int64_t) * num_fragments))
428  : 0;
429  out_vec_dev_buffers.push_back(out_vec_dev_buffer);
430  }
431  }
432  auto out_vec_dev_ptr = gpu_allocator_->alloc(agg_col_count * sizeof(CUdeviceptr));
433  gpu_allocator_->copyToDevice(out_vec_dev_ptr,
434  reinterpret_cast<int8_t*>(out_vec_dev_buffers.data()),
435  agg_col_count * sizeof(CUdeviceptr));
436  kernel_params[GROUPBY_BUF] = reinterpret_cast<CUdeviceptr>(out_vec_dev_ptr);
437  std::vector<void*> param_ptrs;
438  for (auto& param : kernel_params) {
439  param_ptrs.push_back(&param);
440  }
441 
443  cuEventRecord(stop0, 0);
444  cuEventSynchronize(stop0);
445  float milliseconds0 = 0;
446  cuEventElapsedTime(&milliseconds0, start0, stop0);
447  VLOG(1) << "Device " << std::to_string(device_id)
448  << ": launchGpuCode: prepare: " << std::to_string(milliseconds0) << " ms";
449  cuEventRecord(start1, 0);
450  }
451 
452  if (hoist_literals) {
453  checkCudaErrors(cuLaunchKernel(cu_func,
454  grid_size_x,
455  grid_size_y,
456  grid_size_z,
457  block_size_x,
458  block_size_y,
459  block_size_z,
460  0,
461  nullptr,
462  &param_ptrs[0],
463  nullptr));
464  } else {
465  param_ptrs.erase(param_ptrs.begin() + LITERALS); // TODO(alex): remove
466  checkCudaErrors(cuLaunchKernel(cu_func,
467  grid_size_x,
468  grid_size_y,
469  grid_size_z,
470  block_size_x,
471  block_size_y,
472  block_size_z,
473  0,
474  nullptr,
475  &param_ptrs[0],
476  nullptr));
477  }
478 
480  executor_->registerActiveModule(cu_functions[device_id].second, device_id);
481  cuEventRecord(stop1, 0);
482  cuEventSynchronize(stop1);
483  executor_->unregisterActiveModule(cu_functions[device_id].second, device_id);
484  float milliseconds1 = 0;
485  cuEventElapsedTime(&milliseconds1, start1, stop1);
486  VLOG(1) << "Device " << std::to_string(device_id)
487  << ": launchGpuCode: cuLaunchKernel: " << std::to_string(milliseconds1)
488  << " ms";
489  cuEventRecord(start2, 0);
490  }
491 
492  copy_from_gpu(data_mgr,
493  &error_codes[0],
494  err_desc,
495  error_codes.size() * sizeof(error_codes[0]),
496  device_id);
497  *error_code = aggregate_error_codes(error_codes);
498  if (*error_code > 0) {
499  return {};
500  }
501  if (ra_exe_unit.estimator) {
503  estimator_result_set_->syncEstimatorBuffer();
504  return {};
505  }
506  for (size_t i = 0; i < agg_col_count; ++i) {
507  int64_t* host_out_vec =
508  new int64_t[block_size_x * grid_size_x * sizeof(int64_t) * num_fragments];
509  copy_from_gpu(data_mgr,
510  host_out_vec,
511  out_vec_dev_buffers[i],
512  block_size_x * grid_size_x * sizeof(int64_t) * num_fragments,
513  device_id);
514  out_vec.push_back(host_out_vec);
515  }
516  }
517  const auto count_distinct_bitmap_mem = query_buffers_->getCountDistinctBitmapPtr();
518  if (count_distinct_bitmap_mem) {
519  copy_from_gpu(data_mgr,
520  query_buffers_->getCountDistinctHostPtr(),
521  count_distinct_bitmap_mem,
522  query_buffers_->getCountDistinctBitmapBytes(),
523  device_id);
524  }
525 
527  cuEventRecord(stop2, 0);
528  cuEventSynchronize(stop2);
529  float milliseconds2 = 0;
530  cuEventElapsedTime(&milliseconds2, start2, stop2);
531  VLOG(1) << "Device " << std::to_string(device_id)
532  << ": launchGpuCode: finish: " << std::to_string(milliseconds2) << " ms";
533  }
534 
535  return out_vec;
536 #else
537  return {};
538 #endif
539 }
#define CHECK_EQ(x, y)
Definition: Logger.h:195
RenderAllocator * getRenderAllocator(size_t device_id)
bool use_streaming_top_n(const RelAlgExecutionUnit &ra_exe_unit, const bool output_columnar)
const int8_t const int64_t const uint64_t const int32_t const int64_t int64_t uint32_t const int64_t * join_hash_tables
const int64_t const uint32_t const uint32_t const uint32_t agg_col_count
const int8_t const int64_t * num_rows
size_t sharedMemBytes(const ExecutorDeviceType) const
void checkCudaErrors(CUresult err)
Definition: sample.cpp:38
const std::list< Analyzer::OrderEntry > order_entries
unsigned long long CUdeviceptr
Definition: nocuda.h:27
bool use_speculative_top_n(const RelAlgExecutionUnit &ra_exe_unit, const QueryMemoryDescriptor &query_mem_desc)
const ExecutorDispatchMode dispatch_mode_
bool g_enable_dynamic_watchdog
Definition: Execute.cpp:70
void inplace_sort_gpu(const std::list< Analyzer::OrderEntry > &order_entries, const QueryMemoryDescriptor &query_mem_desc, const GpuGroupByBuffers &group_by_buffers, Data_Namespace::DataMgr *data_mgr, const int device_id)
std::string to_string(char const *&&v)
std::unique_ptr< QueryMemoryInitializer > query_buffers_
void copy_to_gpu(Data_Namespace::DataMgr *data_mgr, CUdeviceptr dst, const void *src, const size_t num_bytes, const int device_id)
Definition: GpuMemUtils.cpp:31
const SortInfo sort_info
#define INJECT_TIMER(DESC)
Definition: measure.h:91
const int8_t const int64_t const uint64_t const int32_t const int64_t int64_t uint32_t const int64_t int32_t * error_code
void copy_from_gpu(Data_Namespace::DataMgr *data_mgr, void *dst, const CUdeviceptr src, const size_t num_bytes, const int device_id)
void * CUfunction
Definition: nocuda.h:24
const int8_t const int64_t const uint64_t const int32_t * max_matched
size_t get_num_allocated_rows_from_gpu(Data_Namespace::DataMgr *data_mgr, CUdeviceptr projection_size_gpu, const int device_id)
std::unique_ptr< CudaAllocator > gpu_allocator_
#define CHECK(condition)
Definition: Logger.h:187
std::unique_ptr< ResultSet > estimator_result_set_
QueryDescriptionType getQueryDescriptionType() const
const QueryMemoryDescriptor query_mem_desc_
#define VLOG(n)
Definition: Logger.h:277
+ Here is the call graph for this function:
+ Here is the caller graph for this function:

Friends And Related Function Documentation

◆ AggregateReductionEgress

template<typename META_CLASS_TYPE >
friend class AggregateReductionEgress
friend

Definition at line 144 of file QueryExecutionContext.h.

◆ Executor

friend class Executor
friend

Definition at line 140 of file QueryExecutionContext.h.

Member Data Documentation

◆ device_type_

const ExecutorDeviceType QueryExecutionContext::device_type_
private

Definition at line 133 of file QueryExecutionContext.h.

Referenced by getRowSet(), and groupBufferToResults().

◆ dispatch_mode_

const ExecutorDispatchMode QueryExecutionContext::dispatch_mode_
private

Definition at line 134 of file QueryExecutionContext.h.

Referenced by launchGpuCode().

◆ estimator_result_set_

std::unique_ptr<ResultSet> QueryExecutionContext::estimator_result_set_
mutableprivate

◆ executor_

const Executor* QueryExecutionContext::executor_
private

◆ gpu_allocator_

std::unique_ptr<CudaAllocator> QueryExecutionContext::gpu_allocator_
private

Definition at line 128 of file QueryExecutionContext.h.

Referenced by launchCpuCode(), launchGpuCode(), and QueryExecutionContext().

◆ output_columnar_

const bool QueryExecutionContext::output_columnar_
private

◆ query_buffers_

◆ query_mem_desc_

◆ row_set_mem_owner_

std::shared_ptr<RowSetMemoryOwner> QueryExecutionContext::row_set_mem_owner_
private

Definition at line 135 of file QueryExecutionContext.h.

Referenced by getRowSet(), and groupBufferToDeinterleavedResults().


The documentation for this class was generated from the following files: