OmniSciDB  94e8789169
 All Classes Namespaces Files Functions Variables Typedefs Enumerations Enumerator Friends Macros Pages
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 GpuCompilationContext *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, const size_t shared_memory_size, 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 CpuCompilationContext *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_
 
QueryMemoryDescriptor query_mem_desc_
 
const Executorexecutor_
 
const ExecutorDeviceType device_type_
 
const ExecutorDispatchMode dispatch_mode_
 
std::shared_ptr
< RowSetMemoryOwner
row_set_mem_owner_
 
const bool output_columnar_
 
std::unique_ptr
< QueryMemoryInitializer
query_buffers_
 
std::unique_ptr< ResultSetestimator_result_set_
 

Friends

class Executor
 

Detailed Description

Definition at line 38 of file QueryExecutionContext.h.

Constructor & Destructor Documentation

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(), query_buffers_, query_mem_desc, 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,
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 ExecutorDispatchMode dispatch_mode_
const ExecutorDeviceType device_type_
std::unique_ptr< QueryMemoryInitializer > query_buffers_
QueryMemoryDescriptor query_mem_desc_
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_
bool isPotentialInSituRender() const
Definition: RenderInfo.cpp:64
#define CHECK(condition)
Definition: Logger.h:197
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)

+ Here is the call graph for this function:

Member Function Documentation

int64_t QueryExecutionContext::getAggInitValForIndex ( const size_t  index) const

Definition at line 128 of file QueryExecutionContext.cpp.

References CHECK, and query_buffers_.

Referenced by Executor::executePlanWithoutGroupBy().

128  {
130  return query_buffers_->getAggInitValForIndex(index);
131 }
std::unique_ptr< QueryMemoryInitializer > query_buffers_
#define CHECK(condition)
Definition: Logger.h:197

+ Here is the caller graph for this function:

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

Definition at line 133 of file QueryExecutionContext.cpp.

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

Referenced by Executor::executePlanWithGroupBy().

135  {
136  auto timer = DEBUG_TIMER(__func__);
137  std::vector<std::pair<ResultSetPtr, std::vector<size_t>>> results_per_sm;
139  const auto group_by_buffers_size = query_buffers_->getNumBuffers();
141  CHECK_EQ(size_t(1), group_by_buffers_size);
142  return groupBufferToResults(0);
143  }
144  size_t step{query_mem_desc_.threadsShareMemory() ? executor_->blockSize() : 1};
145  for (size_t i = 0; i < group_by_buffers_size; i += step) {
146  results_per_sm.emplace_back(groupBufferToResults(i), std::vector<size_t>{});
147  }
149  return executor_->reduceMultiDeviceResults(
150  ra_exe_unit, results_per_sm, row_set_mem_owner_, query_mem_desc);
151 }
#define CHECK_EQ(x, y)
Definition: Logger.h:205
const ExecutorDeviceType device_type_
std::unique_ptr< QueryMemoryInitializer > query_buffers_
QueryMemoryDescriptor query_mem_desc_
std::shared_ptr< RowSetMemoryOwner > row_set_mem_owner_
ResultSetPtr groupBufferToResults(const size_t i) const
#define CHECK(condition)
Definition: Logger.h:197
#define DEBUG_TIMER(name)
Definition: Logger.h:313

+ Here is the call graph for this function:

+ Here is the caller graph for this function:

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

Definition at line 75 of file QueryExecutionContext.cpp.

References 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_->getCatalog(),
95  executor_->blockSize(),
96  executor_->gridSize());
97  auto deinterleaved_storage =
98  deinterleaved_result_set->allocateStorage(executor_->plan_state_->init_agg_vals_);
99  auto deinterleaved_buffer =
100  reinterpret_cast<int64_t*>(deinterleaved_storage->getUnderlyingBuffer());
101  const auto rows_ptr = result_set->getStorage()->getUnderlyingBuffer();
102  size_t deinterleaved_buffer_idx = 0;
103  const size_t agg_col_count{query_mem_desc_.getSlotCount()};
104  for (size_t bin_base_off = query_mem_desc_.getColOffInBytes(0), bin_idx = 0;
105  bin_idx < result_set->entryCount();
106  ++bin_idx, bin_base_off += query_mem_desc_.getColOffInBytesInNextBin(0)) {
107  std::vector<int64_t> agg_vals(agg_col_count, 0);
108  memcpy(&agg_vals[0],
109  &executor_->plan_state_->init_agg_vals_[0],
110  agg_col_count * sizeof(agg_vals[0]));
111  ResultSetStorage::reduceSingleRow(rows_ptr + bin_base_off,
112  executor_->warpSize(),
113  false,
114  true,
115  agg_vals,
117  result_set->getTargetInfos(),
118  executor_->plan_state_->init_agg_vals_);
119  for (size_t agg_idx = 0; agg_idx < agg_col_count;
120  ++agg_idx, ++deinterleaved_buffer_idx) {
121  deinterleaved_buffer[deinterleaved_buffer_idx] = agg_vals[agg_idx];
122  }
123  }
124  query_buffers_->resetResultSet(i);
125  return deinterleaved_result_set;
126 }
std::unique_ptr< QueryMemoryInitializer > query_buffers_
QueryMemoryDescriptor query_mem_desc_
std::shared_ptr< RowSetMemoryOwner > row_set_mem_owner_
static QueryMemoryDescriptor fixupQueryMemoryDescriptor(const QueryMemoryDescriptor &)
Definition: ResultSet.cpp:483
#define CHECK(condition)
Definition: Logger.h:197
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)
size_t getColOffInBytesInNextBin(const size_t col_idx) const

+ Here is the call graph for this function:

+ Here is the caller graph for this function:

ResultSetPtr QueryExecutionContext::groupBufferToResults ( const size_t  i) const

Definition at line 153 of file QueryExecutionContext.cpp.

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

Referenced by getRowSet().

153  {
156  }
157  return query_buffers_->getResultSetOwned(i);
158 }
const ExecutorDeviceType device_type_
std::unique_ptr< QueryMemoryInitializer > query_buffers_
QueryMemoryDescriptor query_mem_desc_
ResultSetPtr groupBufferToDeinterleavedResults(const size_t i) const
bool interleavedBins(const ExecutorDeviceType) const

+ Here is the call graph for this function:

+ Here is the caller graph for this function:

std::vector< int64_t * > QueryExecutionContext::launchCpuCode ( const RelAlgExecutionUnit ra_exe_unit,
const CpuCompilationContext 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 571 of file QueryExecutionContext.cpp.

References align_to_int64(), CHECK, CHECK_EQ, compact_init_vals(), CPU, DEBUG_TIMER, QueryMemoryDescriptor::didOutputColumnar(), RelAlgExecutionUnit::estimator, estimator_result_set_, CpuCompilationContext::func(), QueryMemoryDescriptor::getColsSize(), QueryMemoryDescriptor::getQueryDescriptionType(), generate_TableFunctionsFactory_init::i, INJECT_TIMER, QueryMemoryDescriptor::isGroupBy(), Projection, query_buffers_, query_mem_desc_, and QueryMemoryDescriptor::useStreamingTopN().

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

582  {
583  auto timer = DEBUG_TIMER(__func__);
584  INJECT_TIMER(lauchCpuCode);
585 
587  const auto& init_agg_vals = query_buffers_->init_agg_vals_;
588 
589  std::vector<const int8_t**> multifrag_col_buffers;
590  for (auto& col_buffer : col_buffers) {
591  multifrag_col_buffers.push_back(&col_buffer[0]);
592  }
593  const int8_t*** multifrag_cols_ptr{
594  multifrag_col_buffers.empty() ? nullptr : &multifrag_col_buffers[0]};
595  const uint64_t num_fragments =
596  multifrag_cols_ptr ? static_cast<uint64_t>(col_buffers.size()) : uint64_t(0);
597  const auto num_out_frags = multifrag_cols_ptr ? num_fragments : uint64_t(0);
598 
599  const bool is_group_by{query_mem_desc_.isGroupBy()};
600  std::vector<int64_t*> out_vec;
601  if (ra_exe_unit.estimator) {
602  estimator_result_set_.reset(
603  new ResultSet(ra_exe_unit.estimator, ExecutorDeviceType::CPU, 0, nullptr));
604  out_vec.push_back(
605  reinterpret_cast<int64_t*>(estimator_result_set_->getHostEstimatorBuffer()));
606  } else {
607  if (!is_group_by) {
608  for (size_t i = 0; i < init_agg_vals.size(); ++i) {
609  auto buff = new int64_t[num_out_frags];
610  out_vec.push_back(static_cast<int64_t*>(buff));
611  }
612  }
613  }
614 
615  CHECK_EQ(num_rows.size(), col_buffers.size());
616  std::vector<int64_t> flatened_num_rows;
617  for (auto& nums : num_rows) {
618  flatened_num_rows.insert(flatened_num_rows.end(), nums.begin(), nums.end());
619  }
620  std::vector<uint64_t> flatened_frag_offsets;
621  for (auto& offsets : frag_offsets) {
622  flatened_frag_offsets.insert(
623  flatened_frag_offsets.end(), offsets.begin(), offsets.end());
624  }
625  int64_t rowid_lookup_num_rows{*error_code ? *error_code + 1 : 0};
626  auto num_rows_ptr =
627  rowid_lookup_num_rows ? &rowid_lookup_num_rows : &flatened_num_rows[0];
628  int32_t total_matched_init{0};
629 
630  std::vector<int64_t> cmpt_val_buff;
631  if (is_group_by) {
632  cmpt_val_buff =
634  init_agg_vals,
636  }
637 
638  CHECK(native_code);
639  const int64_t* join_hash_tables_ptr =
640  join_hash_tables.size() == 1
641  ? reinterpret_cast<int64_t*>(join_hash_tables[0])
642  : (join_hash_tables.size() > 1 ? &join_hash_tables[0] : nullptr);
643  if (hoist_literals) {
644  using agg_query = void (*)(const int8_t***, // col_buffers
645  const uint64_t*, // num_fragments
646  const int8_t*, // literals
647  const int64_t*, // num_rows
648  const uint64_t*, // frag_row_offsets
649  const int32_t*, // max_matched
650  int32_t*, // total_matched
651  const int64_t*, // init_agg_value
652  int64_t**, // out
653  int32_t*, // error_code
654  const uint32_t*, // num_tables
655  const int64_t*); // join_hash_tables_ptr
656  if (is_group_by) {
657  reinterpret_cast<agg_query>(native_code->func())(
658  multifrag_cols_ptr,
659  &num_fragments,
660  &literal_buff[0],
661  num_rows_ptr,
662  &flatened_frag_offsets[0],
663  &scan_limit,
664  &total_matched_init,
665  &cmpt_val_buff[0],
666  query_buffers_->getGroupByBuffersPtr(),
667  error_code,
668  &num_tables,
669  join_hash_tables_ptr);
670  } else {
671  reinterpret_cast<agg_query>(native_code->func())(multifrag_cols_ptr,
672  &num_fragments,
673  &literal_buff[0],
674  num_rows_ptr,
675  &flatened_frag_offsets[0],
676  &scan_limit,
677  &total_matched_init,
678  &init_agg_vals[0],
679  &out_vec[0],
680  error_code,
681  &num_tables,
682  join_hash_tables_ptr);
683  }
684  } else {
685  using agg_query = void (*)(const int8_t***, // col_buffers
686  const uint64_t*, // num_fragments
687  const int64_t*, // num_rows
688  const uint64_t*, // frag_row_offsets
689  const int32_t*, // max_matched
690  int32_t*, // total_matched
691  const int64_t*, // init_agg_value
692  int64_t**, // out
693  int32_t*, // error_code
694  const uint32_t*, // num_tables
695  const int64_t*); // join_hash_tables_ptr
696  if (is_group_by) {
697  reinterpret_cast<agg_query>(native_code->func())(
698  multifrag_cols_ptr,
699  &num_fragments,
700  num_rows_ptr,
701  &flatened_frag_offsets[0],
702  &scan_limit,
703  &total_matched_init,
704  &cmpt_val_buff[0],
705  query_buffers_->getGroupByBuffersPtr(),
706  error_code,
707  &num_tables,
708  join_hash_tables_ptr);
709  } else {
710  reinterpret_cast<agg_query>(native_code->func())(multifrag_cols_ptr,
711  &num_fragments,
712  num_rows_ptr,
713  &flatened_frag_offsets[0],
714  &scan_limit,
715  &total_matched_init,
716  &init_agg_vals[0],
717  &out_vec[0],
718  error_code,
719  &num_tables,
720  join_hash_tables_ptr);
721  }
722  }
723 
724  if (ra_exe_unit.estimator) {
725  return {};
726  }
727 
728  if (rowid_lookup_num_rows && *error_code < 0) {
729  *error_code = 0;
730  }
731 
733  query_buffers_->applyStreamingTopNOffsetCpu(query_mem_desc_, ra_exe_unit);
734  }
735 
738  query_buffers_->compactProjectionBuffersCpu(query_mem_desc_, total_matched_init);
739  }
740 
741  return out_vec;
742 }
#define CHECK_EQ(x, y)
Definition: Logger.h:205
std::unique_ptr< QueryMemoryInitializer > query_buffers_
#define INJECT_TIMER(DESC)
Definition: measure.h:93
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 std::shared_ptr< Analyzer::Estimator > estimator
QueryDescriptionType getQueryDescriptionType() const
QueryMemoryDescriptor query_mem_desc_
#define CHECK(condition)
Definition: Logger.h:197
#define DEBUG_TIMER(name)
Definition: Logger.h:313
std::unique_ptr< ResultSet > estimator_result_set_
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:

std::vector< int64_t * > QueryExecutionContext::launchGpuCode ( const RelAlgExecutionUnit ra_exe_unit,
const GpuCompilationContext 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,
const size_t  shared_memory_size,
int32_t *  error_code,
const uint32_t  num_tables,
const std::vector< int64_t > &  join_hash_tables,
RenderAllocatorMap render_allocator_map 
)

Definition at line 181 of file QueryExecutionContext.cpp.

References CHECK, CHECK_EQ, checkCudaErrors(), copy_from_gpu(), copy_to_gpu(), DEBUG_TIMER, QueryMemoryDescriptor::didOutputColumnar(), dispatch_mode_, RelAlgExecutionUnit::estimator, estimator_result_set_, executor_, g_enable_dynamic_watchdog, g_enable_runtime_query_interrupt, get_num_allocated_rows_from_gpu(), QueryMemoryDescriptor::getEntryCount(), GpuCompilationContext::getNativeCode(), QueryMemoryDescriptor::getQueryDescriptionType(), RenderAllocatorMap::getRenderAllocator(), GPU, gpu_allocator_, QueryMemoryDescriptor::hasKeylessHash(), generate_TableFunctionsFactory_init::i, INJECT_TIMER, inplace_sort_gpu(), QueryMemoryDescriptor::isGroupBy(), SortInfo::order_entries, output_columnar_, Projection, query_buffers_, query_mem_desc_, RelAlgExecutionUnit::sort_info, QueryMemoryDescriptor::sortOnGpu(), to_string(), RelAlgExecutionUnit::use_bump_allocator, use_speculative_top_n(), QueryMemoryDescriptor::useStreamingTopN(), and VLOG.

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

198  {
199  auto timer = DEBUG_TIMER(__func__);
200  INJECT_TIMER(lauchGpuCode);
201 #ifdef HAVE_CUDA
204  const auto& init_agg_vals = query_buffers_->init_agg_vals_;
205 
206  bool is_group_by{query_mem_desc_.isGroupBy()};
207 
208  RenderAllocator* render_allocator = nullptr;
209  if (render_allocator_map) {
210  render_allocator = render_allocator_map->getRenderAllocator(device_id);
211  }
212 
213  CHECK(cu_functions);
214  const auto native_code = cu_functions->getNativeCode(device_id);
215  auto cu_func = static_cast<CUfunction>(native_code.first);
216  std::vector<int64_t*> out_vec;
217  uint32_t num_fragments = col_buffers.size();
218  std::vector<int32_t> error_codes(grid_size_x * block_size_x);
219 
220  CUevent start0, stop0; // preparation
221  cuEventCreate(&start0, 0);
222  cuEventCreate(&stop0, 0);
223  CUevent start1, stop1; // cuLaunchKernel
224  cuEventCreate(&start1, 0);
225  cuEventCreate(&stop1, 0);
226  CUevent start2, stop2; // finish
227  cuEventCreate(&start2, 0);
228  cuEventCreate(&stop2, 0);
229 
231  (g_enable_runtime_query_interrupt && !render_allocator)) {
232  cuEventRecord(start0, 0);
233  }
234 
236  initializeDynamicWatchdog(native_code.second, device_id);
237  }
238 
239  if (g_enable_runtime_query_interrupt && !render_allocator) {
240  initializeRuntimeInterrupter(native_code.second, device_id);
241  }
242 
243  auto kernel_params = prepareKernelParams(col_buffers,
244  literal_buff,
245  num_rows,
246  frag_offsets,
247  scan_limit,
248  init_agg_vals,
249  error_codes,
250  num_tables,
251  join_hash_tables,
252  data_mgr,
253  device_id,
254  hoist_literals,
255  is_group_by);
256 
257  CHECK_EQ(static_cast<size_t>(KERN_PARAM_COUNT), kernel_params.size());
258  CHECK_EQ(CUdeviceptr(0), kernel_params[GROUPBY_BUF]);
259 
260  const unsigned block_size_y = 1;
261  const unsigned block_size_z = 1;
262  const unsigned grid_size_y = 1;
263  const unsigned grid_size_z = 1;
264  const auto total_thread_count = block_size_x * grid_size_x;
265  const auto err_desc = kernel_params[ERROR_CODE];
266 
267  if (is_group_by) {
268  CHECK(!(query_buffers_->getGroupByBuffersSize() == 0) || render_allocator);
269  bool can_sort_on_gpu = query_mem_desc_.sortOnGpu();
270  auto gpu_group_by_buffers =
271  query_buffers_->createAndInitializeGroupByBufferGpu(ra_exe_unit,
273  kernel_params[INIT_AGG_VALS],
274  device_id,
276  block_size_x,
277  grid_size_x,
278  executor_->warpSize(),
279  can_sort_on_gpu,
281  render_allocator);
282  if (ra_exe_unit.use_bump_allocator) {
283  const auto max_matched = static_cast<int32_t>(gpu_group_by_buffers.entry_count);
284  copy_to_gpu(data_mgr,
285  kernel_params[MAX_MATCHED],
286  &max_matched,
287  sizeof(max_matched),
288  device_id);
289  }
290 
291  kernel_params[GROUPBY_BUF] = gpu_group_by_buffers.first;
292  std::vector<void*> param_ptrs;
293  for (auto& param : kernel_params) {
294  param_ptrs.push_back(&param);
295  }
296 
298  (g_enable_runtime_query_interrupt && !render_allocator)) {
299  cuEventRecord(stop0, 0);
300  cuEventSynchronize(stop0);
301  float milliseconds0 = 0;
302  cuEventElapsedTime(&milliseconds0, start0, stop0);
303  VLOG(1) << "Device " << std::to_string(device_id)
304  << ": launchGpuCode: group-by prepare: " << std::to_string(milliseconds0)
305  << " ms";
306  cuEventRecord(start1, 0);
307  }
308 
309  if (hoist_literals) {
310  checkCudaErrors(cuLaunchKernel(cu_func,
311  grid_size_x,
312  grid_size_y,
313  grid_size_z,
314  block_size_x,
315  block_size_y,
316  block_size_z,
317  shared_memory_size,
318  nullptr,
319  &param_ptrs[0],
320  nullptr));
321  } else {
322  param_ptrs.erase(param_ptrs.begin() + LITERALS); // TODO(alex): remove
323  checkCudaErrors(cuLaunchKernel(cu_func,
324  grid_size_x,
325  grid_size_y,
326  grid_size_z,
327  block_size_x,
328  block_size_y,
329  block_size_z,
330  shared_memory_size,
331  nullptr,
332  &param_ptrs[0],
333  nullptr));
334  }
336  (g_enable_runtime_query_interrupt && !render_allocator)) {
337  executor_->registerActiveModule(native_code.second, device_id);
338  cuEventRecord(stop1, 0);
339  cuEventSynchronize(stop1);
340  executor_->unregisterActiveModule(native_code.second, device_id);
341  float milliseconds1 = 0;
342  cuEventElapsedTime(&milliseconds1, start1, stop1);
343  VLOG(1) << "Device " << std::to_string(device_id)
344  << ": launchGpuCode: group-by cuLaunchKernel: "
345  << std::to_string(milliseconds1) << " ms";
346  cuEventRecord(start2, 0);
347  }
348 
349  gpu_allocator_->copyFromDevice(reinterpret_cast<int8_t*>(error_codes.data()),
350  reinterpret_cast<int8_t*>(err_desc),
351  error_codes.size() * sizeof(error_codes[0]));
352  *error_code = aggregate_error_codes(error_codes);
353  if (*error_code > 0) {
354  return {};
355  }
356 
357  if (!render_allocator) {
359  query_buffers_->applyStreamingTopNOffsetGpu(data_mgr,
361  gpu_group_by_buffers,
362  ra_exe_unit,
363  total_thread_count,
364  device_id);
365  } else {
366  if (use_speculative_top_n(ra_exe_unit, query_mem_desc_)) {
367  try {
370  gpu_group_by_buffers,
371  data_mgr,
372  device_id);
373  } catch (const std::bad_alloc&) {
374  throw SpeculativeTopNFailed("Failed during in-place GPU sort.");
375  }
376  }
380  query_buffers_->compactProjectionBuffersGpu(
382  data_mgr,
383  gpu_group_by_buffers,
385  data_mgr, kernel_params[TOTAL_MATCHED], device_id),
386  device_id);
387  } else {
388  size_t num_allocated_rows{0};
389  if (ra_exe_unit.use_bump_allocator) {
390  num_allocated_rows = get_num_allocated_rows_from_gpu(
391  data_mgr, kernel_params[TOTAL_MATCHED], device_id);
392  // First, check the error code. If we ran out of slots, don't copy data back
393  // into the ResultSet or update ResultSet entry count
394  if (*error_code < 0) {
395  return {};
396  }
397  }
398  query_buffers_->copyGroupByBuffersFromGpu(
399  data_mgr,
401  ra_exe_unit.use_bump_allocator ? num_allocated_rows
403  gpu_group_by_buffers,
404  &ra_exe_unit,
405  block_size_x,
406  grid_size_x,
407  device_id,
408  can_sort_on_gpu && query_mem_desc_.hasKeylessHash());
409  if (num_allocated_rows) {
410  CHECK(ra_exe_unit.use_bump_allocator);
411  CHECK(!query_buffers_->result_sets_.empty());
412  query_buffers_->result_sets_.front()->updateStorageEntryCount(
413  num_allocated_rows);
414  }
415  }
416  } else {
417  query_buffers_->copyGroupByBuffersFromGpu(
418  data_mgr,
421  gpu_group_by_buffers,
422  &ra_exe_unit,
423  block_size_x,
424  grid_size_x,
425  device_id,
426  can_sort_on_gpu && query_mem_desc_.hasKeylessHash());
427  }
428  }
429  }
430  } else {
431  std::vector<CUdeviceptr> out_vec_dev_buffers;
432  const size_t agg_col_count{ra_exe_unit.estimator ? size_t(1) : init_agg_vals.size()};
433  // by default, non-grouped aggregate queries generate one result per available thread
434  // in the lifetime of (potentially multi-fragment) kernel execution.
435  // We can reduce these intermediate results internally in the device and hence have
436  // only one result per device, if GPU shared memory optimizations are enabled.
437  const auto num_results_per_agg_col =
438  shared_memory_size ? 1 : block_size_x * grid_size_x * num_fragments;
439  const auto output_buffer_size_per_agg = num_results_per_agg_col * sizeof(int64_t);
440  if (ra_exe_unit.estimator) {
441  estimator_result_set_.reset(new ResultSet(
442  ra_exe_unit.estimator, ExecutorDeviceType::GPU, device_id, data_mgr));
443  out_vec_dev_buffers.push_back(reinterpret_cast<CUdeviceptr>(
444  estimator_result_set_->getDeviceEstimatorBuffer()));
445  } else {
446  for (size_t i = 0; i < agg_col_count; ++i) {
447  CUdeviceptr out_vec_dev_buffer =
448  num_fragments ? reinterpret_cast<CUdeviceptr>(
449  gpu_allocator_->alloc(output_buffer_size_per_agg))
450  : 0;
451  out_vec_dev_buffers.push_back(out_vec_dev_buffer);
452  if (shared_memory_size) {
453  CHECK_EQ(output_buffer_size_per_agg, size_t(8));
454  gpu_allocator_->copyToDevice(reinterpret_cast<int8_t*>(out_vec_dev_buffer),
455  reinterpret_cast<const int8_t*>(&init_agg_vals[i]),
456  output_buffer_size_per_agg);
457  }
458  }
459  }
460  auto out_vec_dev_ptr = gpu_allocator_->alloc(agg_col_count * sizeof(CUdeviceptr));
461  gpu_allocator_->copyToDevice(out_vec_dev_ptr,
462  reinterpret_cast<int8_t*>(out_vec_dev_buffers.data()),
463  agg_col_count * sizeof(CUdeviceptr));
464  kernel_params[GROUPBY_BUF] = reinterpret_cast<CUdeviceptr>(out_vec_dev_ptr);
465  std::vector<void*> param_ptrs;
466  for (auto& param : kernel_params) {
467  param_ptrs.push_back(&param);
468  }
469 
471  (g_enable_runtime_query_interrupt && !render_allocator)) {
472  cuEventRecord(stop0, 0);
473  cuEventSynchronize(stop0);
474  float milliseconds0 = 0;
475  cuEventElapsedTime(&milliseconds0, start0, stop0);
476  VLOG(1) << "Device " << std::to_string(device_id)
477  << ": launchGpuCode: prepare: " << std::to_string(milliseconds0) << " ms";
478  cuEventRecord(start1, 0);
479  }
480 
481  if (hoist_literals) {
482  checkCudaErrors(cuLaunchKernel(cu_func,
483  grid_size_x,
484  grid_size_y,
485  grid_size_z,
486  block_size_x,
487  block_size_y,
488  block_size_z,
489  shared_memory_size,
490  nullptr,
491  &param_ptrs[0],
492  nullptr));
493  } else {
494  param_ptrs.erase(param_ptrs.begin() + LITERALS); // TODO(alex): remove
495  checkCudaErrors(cuLaunchKernel(cu_func,
496  grid_size_x,
497  grid_size_y,
498  grid_size_z,
499  block_size_x,
500  block_size_y,
501  block_size_z,
502  shared_memory_size,
503  nullptr,
504  &param_ptrs[0],
505  nullptr));
506  }
507 
509  (g_enable_runtime_query_interrupt && !render_allocator)) {
510  executor_->registerActiveModule(native_code.second, device_id);
511  cuEventRecord(stop1, 0);
512  cuEventSynchronize(stop1);
513  executor_->unregisterActiveModule(native_code.second, device_id);
514  float milliseconds1 = 0;
515  cuEventElapsedTime(&milliseconds1, start1, stop1);
516  VLOG(1) << "Device " << std::to_string(device_id)
517  << ": launchGpuCode: cuLaunchKernel: " << std::to_string(milliseconds1)
518  << " ms";
519  cuEventRecord(start2, 0);
520  }
521 
522  copy_from_gpu(data_mgr,
523  &error_codes[0],
524  err_desc,
525  error_codes.size() * sizeof(error_codes[0]),
526  device_id);
527  *error_code = aggregate_error_codes(error_codes);
528  if (*error_code > 0) {
529  return {};
530  }
531  if (ra_exe_unit.estimator) {
533  estimator_result_set_->syncEstimatorBuffer();
534  return {};
535  }
536  for (size_t i = 0; i < agg_col_count; ++i) {
537  int64_t* host_out_vec = new int64_t[output_buffer_size_per_agg];
538  copy_from_gpu(data_mgr,
539  host_out_vec,
540  out_vec_dev_buffers[i],
541  output_buffer_size_per_agg,
542  device_id);
543  out_vec.push_back(host_out_vec);
544  }
545  }
546  const auto count_distinct_bitmap_mem = query_buffers_->getCountDistinctBitmapPtr();
547  if (count_distinct_bitmap_mem) {
548  copy_from_gpu(data_mgr,
549  query_buffers_->getCountDistinctHostPtr(),
550  count_distinct_bitmap_mem,
551  query_buffers_->getCountDistinctBitmapBytes(),
552  device_id);
553  }
554 
556  (g_enable_runtime_query_interrupt && !render_allocator)) {
557  cuEventRecord(stop2, 0);
558  cuEventSynchronize(stop2);
559  float milliseconds2 = 0;
560  cuEventElapsedTime(&milliseconds2, start2, stop2);
561  VLOG(1) << "Device " << std::to_string(device_id)
562  << ": launchGpuCode: finish: " << std::to_string(milliseconds2) << " ms";
563  }
564 
565  return out_vec;
566 #else
567  return {};
568 #endif
569 }
#define CHECK_EQ(x, y)
Definition: Logger.h:205
RenderAllocator * getRenderAllocator(size_t device_id)
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:77
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::pair< void *, void * > getNativeCode(const size_t device_id) const
Definition: NvidiaKernel.h:81
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:30
const SortInfo sort_info
#define INJECT_TIMER(DESC)
Definition: measure.h:93
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
QueryDescriptionType getQueryDescriptionType() const
QueryMemoryDescriptor query_mem_desc_
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:197
#define DEBUG_TIMER(name)
Definition: Logger.h:313
std::unique_ptr< ResultSet > estimator_result_set_
bool g_enable_runtime_query_interrupt
Definition: Execute.cpp:111
#define VLOG(n)
Definition: Logger.h:291

+ Here is the call graph for this function:

+ Here is the caller graph for this function:

Friends And Related Function Documentation

friend class Executor
friend

Definition at line 146 of file QueryExecutionContext.h.

Member Data Documentation

const ExecutorDeviceType QueryExecutionContext::device_type_
private

Definition at line 139 of file QueryExecutionContext.h.

Referenced by getRowSet(), and groupBufferToResults().

const ExecutorDispatchMode QueryExecutionContext::dispatch_mode_
private

Definition at line 140 of file QueryExecutionContext.h.

Referenced by launchGpuCode().

std::unique_ptr<ResultSet> QueryExecutionContext::estimator_result_set_
mutableprivate
const Executor* QueryExecutionContext::executor_
private
std::unique_ptr<CudaAllocator> QueryExecutionContext::gpu_allocator_
private

Definition at line 134 of file QueryExecutionContext.h.

Referenced by launchGpuCode(), and QueryExecutionContext().

const bool QueryExecutionContext::output_columnar_
private

Definition at line 142 of file QueryExecutionContext.h.

Referenced by groupBufferToDeinterleavedResults(), and launchGpuCode().

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

Definition at line 141 of file QueryExecutionContext.h.

Referenced by getRowSet(), and groupBufferToDeinterleavedResults().


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