OmniSciDB  06b3bd477c
 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(), num_rows, 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 int8_t const int64_t * num_rows
const ExecutorDispatchMode dispatch_mode_
const ExecutorDeviceType device_type_
std::unique_ptr< QueryMemoryInitializer > query_buffers_
CHECK(cgen_state)
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
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 126 of file QueryExecutionContext.cpp.

References CHECK(), and query_buffers_.

Referenced by Executor::executePlanWithoutGroupBy().

126  {
128  return query_buffers_->getAggInitValForIndex(index);
129 }
std::unique_ptr< QueryMemoryInitializer > query_buffers_
CHECK(cgen_state)

+ Here is the call graph for this function:

+ 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 131 of file QueryExecutionContext.cpp.

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

Referenced by Executor::executePlanWithGroupBy().

133  {
134  auto timer = DEBUG_TIMER(__func__);
135  std::vector<std::pair<ResultSetPtr, std::vector<size_t>>> results_per_sm;
137  const auto group_by_buffers_size = query_buffers_->getNumBuffers();
139  CHECK_EQ(size_t(1), group_by_buffers_size);
140  return groupBufferToResults(0);
141  }
142  size_t step{query_mem_desc_.threadsShareMemory() ? executor_->blockSize() : 1};
143  for (size_t i = 0; i < group_by_buffers_size; i += step) {
144  results_per_sm.emplace_back(groupBufferToResults(i), std::vector<size_t>{});
145  }
147  return executor_->reduceMultiDeviceResults(
148  ra_exe_unit, results_per_sm, row_set_mem_owner_, query_mem_desc);
149 }
#define CHECK_EQ(x, y)
Definition: Logger.h:205
const ExecutorDeviceType device_type_
std::unique_ptr< QueryMemoryInitializer > query_buffers_
CHECK(cgen_state)
QueryMemoryDescriptor query_mem_desc_
std::shared_ptr< RowSetMemoryOwner > row_set_mem_owner_
ResultSetPtr groupBufferToResults(const size_t i) const
#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 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_
CHECK(cgen_state)
QueryMemoryDescriptor query_mem_desc_
std::shared_ptr< RowSetMemoryOwner > row_set_mem_owner_
static QueryMemoryDescriptor fixupQueryMemoryDescriptor(const QueryMemoryDescriptor &)
Definition: ResultSet.cpp:506
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 151 of file QueryExecutionContext.cpp.

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

Referenced by getRowSet().

151  {
154  }
155  return query_buffers_->getResultSetOwned(i);
156 }
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 563 of file QueryExecutionContext.cpp.

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

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

574  {
575  auto timer = DEBUG_TIMER(__func__);
576  INJECT_TIMER(lauchCpuCode);
577 
579  const auto& init_agg_vals = query_buffers_->init_agg_vals_;
580 
581  std::vector<const int8_t**> multifrag_col_buffers;
582  for (auto& col_buffer : col_buffers) {
583  multifrag_col_buffers.push_back(&col_buffer[0]);
584  }
585  const int8_t*** multifrag_cols_ptr{
586  multifrag_col_buffers.empty() ? nullptr : &multifrag_col_buffers[0]};
587  const uint64_t num_fragments =
588  multifrag_cols_ptr ? static_cast<uint64_t>(col_buffers.size()) : uint64_t(0);
589  const auto num_out_frags = multifrag_cols_ptr ? num_fragments : uint64_t(0);
590 
591  const bool is_group_by{query_mem_desc_.isGroupBy()};
592  std::vector<int64_t*> out_vec;
593  if (ra_exe_unit.estimator) {
594  estimator_result_set_.reset(
595  new ResultSet(ra_exe_unit.estimator, ExecutorDeviceType::CPU, 0, nullptr));
596  out_vec.push_back(
597  reinterpret_cast<int64_t*>(estimator_result_set_->getHostEstimatorBuffer()));
598  } else {
599  if (!is_group_by) {
600  for (size_t i = 0; i < init_agg_vals.size(); ++i) {
601  auto buff = new int64_t[num_out_frags];
602  out_vec.push_back(static_cast<int64_t*>(buff));
603  }
604  }
605  }
606 
607  CHECK_EQ(num_rows.size(), col_buffers.size());
608  std::vector<int64_t> flatened_num_rows;
609  for (auto& nums : num_rows) {
610  flatened_num_rows.insert(flatened_num_rows.end(), nums.begin(), nums.end());
611  }
612  std::vector<uint64_t> flatened_frag_offsets;
613  for (auto& offsets : frag_offsets) {
614  flatened_frag_offsets.insert(
615  flatened_frag_offsets.end(), offsets.begin(), offsets.end());
616  }
617  int64_t rowid_lookup_num_rows{*error_code ? *error_code + 1 : 0};
618  auto num_rows_ptr =
619  rowid_lookup_num_rows ? &rowid_lookup_num_rows : &flatened_num_rows[0];
620  int32_t total_matched_init{0};
621 
622  std::vector<int64_t> cmpt_val_buff;
623  if (is_group_by) {
624  cmpt_val_buff =
626  init_agg_vals,
628  }
629 
630  CHECK(native_code);
631  const int64_t* join_hash_tables_ptr =
632  join_hash_tables.size() == 1
633  ? reinterpret_cast<int64_t*>(join_hash_tables[0])
634  : (join_hash_tables.size() > 1 ? &join_hash_tables[0] : nullptr);
635  if (hoist_literals) {
636  using agg_query = void (*)(const int8_t***, // col_buffers
637  const uint64_t*, // num_fragments
638  const int8_t*, // literals
639  const int64_t*, // num_rows
640  const uint64_t*, // frag_row_offsets
641  const int32_t*, // max_matched
642  int32_t*, // total_matched
643  const int64_t*, // init_agg_value
644  int64_t**, // out
645  int32_t*, // error_code
646  const uint32_t*, // num_tables
647  const int64_t*); // join_hash_tables_ptr
648  if (is_group_by) {
649  reinterpret_cast<agg_query>(native_code->func())(
650  multifrag_cols_ptr,
651  &num_fragments,
652  &literal_buff[0],
653  num_rows_ptr,
654  &flatened_frag_offsets[0],
655  &scan_limit,
656  &total_matched_init,
657  &cmpt_val_buff[0],
658  query_buffers_->getGroupByBuffersPtr(),
659  error_code,
660  &num_tables,
661  join_hash_tables_ptr);
662  } else {
663  reinterpret_cast<agg_query>(native_code->func())(multifrag_cols_ptr,
664  &num_fragments,
665  &literal_buff[0],
666  num_rows_ptr,
667  &flatened_frag_offsets[0],
668  &scan_limit,
669  &total_matched_init,
670  &init_agg_vals[0],
671  &out_vec[0],
672  error_code,
673  &num_tables,
674  join_hash_tables_ptr);
675  }
676  } else {
677  using agg_query = void (*)(const int8_t***, // col_buffers
678  const uint64_t*, // num_fragments
679  const int64_t*, // num_rows
680  const uint64_t*, // frag_row_offsets
681  const int32_t*, // max_matched
682  int32_t*, // total_matched
683  const int64_t*, // init_agg_value
684  int64_t**, // out
685  int32_t*, // error_code
686  const uint32_t*, // num_tables
687  const int64_t*); // join_hash_tables_ptr
688  if (is_group_by) {
689  reinterpret_cast<agg_query>(native_code->func())(
690  multifrag_cols_ptr,
691  &num_fragments,
692  num_rows_ptr,
693  &flatened_frag_offsets[0],
694  &scan_limit,
695  &total_matched_init,
696  &cmpt_val_buff[0],
697  query_buffers_->getGroupByBuffersPtr(),
698  error_code,
699  &num_tables,
700  join_hash_tables_ptr);
701  } else {
702  reinterpret_cast<agg_query>(native_code->func())(multifrag_cols_ptr,
703  &num_fragments,
704  num_rows_ptr,
705  &flatened_frag_offsets[0],
706  &scan_limit,
707  &total_matched_init,
708  &init_agg_vals[0],
709  &out_vec[0],
710  error_code,
711  &num_tables,
712  join_hash_tables_ptr);
713  }
714  }
715 
716  if (ra_exe_unit.estimator) {
717  return {};
718  }
719 
720  if (rowid_lookup_num_rows && *error_code < 0) {
721  *error_code = 0;
722  }
723 
725  query_buffers_->applyStreamingTopNOffsetCpu(query_mem_desc_, ra_exe_unit);
726  }
727 
730  query_buffers_->compactProjectionBuffersCpu(query_mem_desc_, total_matched_init);
731  }
732 
733  return out_vec;
734 }
#define CHECK_EQ(x, y)
Definition: Logger.h:205
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_
CHECK(cgen_state)
#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
QueryDescriptionType getQueryDescriptionType() const
QueryMemoryDescriptor query_mem_desc_
#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 179 of file QueryExecutionContext.cpp.

References agg_col_count, 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(), INJECT_TIMER, inplace_sort_gpu(), QueryMemoryDescriptor::isGroupBy(), max_matched, num_rows, 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().

196  {
197  auto timer = DEBUG_TIMER(__func__);
198  INJECT_TIMER(lauchGpuCode);
199 #ifdef HAVE_CUDA
202  const auto& init_agg_vals = query_buffers_->init_agg_vals_;
203 
204  bool is_group_by{query_mem_desc_.isGroupBy()};
205 
206  RenderAllocator* render_allocator = nullptr;
207  if (render_allocator_map) {
208  render_allocator = render_allocator_map->getRenderAllocator(device_id);
209  }
210 
211  CHECK(cu_functions);
212  const auto native_code = cu_functions->getNativeCode(device_id);
213  auto cu_func = static_cast<CUfunction>(native_code.first);
214  std::vector<int64_t*> out_vec;
215  uint32_t num_fragments = col_buffers.size();
216  std::vector<int32_t> error_codes(grid_size_x * block_size_x);
217 
218  CUevent start0, stop0; // preparation
219  cuEventCreate(&start0, 0);
220  cuEventCreate(&stop0, 0);
221  CUevent start1, stop1; // cuLaunchKernel
222  cuEventCreate(&start1, 0);
223  cuEventCreate(&stop1, 0);
224  CUevent start2, stop2; // finish
225  cuEventCreate(&start2, 0);
226  cuEventCreate(&stop2, 0);
227 
229  cuEventRecord(start0, 0);
230  }
231 
233  initializeDynamicWatchdog(native_code.second, device_id);
234  }
235 
237  initializeRuntimeInterrupter(native_code.second, device_id);
238  }
239 
240  auto kernel_params = prepareKernelParams(col_buffers,
241  literal_buff,
242  num_rows,
243  frag_offsets,
244  scan_limit,
245  init_agg_vals,
246  error_codes,
247  num_tables,
249  data_mgr,
250  device_id,
251  hoist_literals,
252  is_group_by);
253 
254  CHECK_EQ(static_cast<size_t>(KERN_PARAM_COUNT), kernel_params.size());
255  CHECK_EQ(CUdeviceptr(0), kernel_params[GROUPBY_BUF]);
256 
257  const unsigned block_size_y = 1;
258  const unsigned block_size_z = 1;
259  const unsigned grid_size_y = 1;
260  const unsigned grid_size_z = 1;
261  const auto total_thread_count = block_size_x * grid_size_x;
262  const auto err_desc = kernel_params[ERROR_CODE];
263 
264  if (is_group_by) {
265  CHECK(!(query_buffers_->getGroupByBuffersSize() == 0) || render_allocator);
266  bool can_sort_on_gpu = query_mem_desc_.sortOnGpu();
267  auto gpu_group_by_buffers =
268  query_buffers_->createAndInitializeGroupByBufferGpu(ra_exe_unit,
270  kernel_params[INIT_AGG_VALS],
271  device_id,
273  block_size_x,
274  grid_size_x,
275  executor_->warpSize(),
276  can_sort_on_gpu,
278  render_allocator);
279  if (ra_exe_unit.use_bump_allocator) {
280  const auto max_matched = static_cast<int32_t>(gpu_group_by_buffers.entry_count);
281  copy_to_gpu(data_mgr,
282  kernel_params[MAX_MATCHED],
283  &max_matched,
284  sizeof(max_matched),
285  device_id);
286  }
287 
288  kernel_params[GROUPBY_BUF] = gpu_group_by_buffers.first;
289  std::vector<void*> param_ptrs;
290  for (auto& param : kernel_params) {
291  param_ptrs.push_back(&param);
292  }
293 
295  cuEventRecord(stop0, 0);
296  cuEventSynchronize(stop0);
297  float milliseconds0 = 0;
298  cuEventElapsedTime(&milliseconds0, start0, stop0);
299  VLOG(1) << "Device " << std::to_string(device_id)
300  << ": launchGpuCode: group-by prepare: " << std::to_string(milliseconds0)
301  << " ms";
302  cuEventRecord(start1, 0);
303  }
304 
305  if (hoist_literals) {
306  checkCudaErrors(cuLaunchKernel(cu_func,
307  grid_size_x,
308  grid_size_y,
309  grid_size_z,
310  block_size_x,
311  block_size_y,
312  block_size_z,
313  shared_memory_size,
314  nullptr,
315  &param_ptrs[0],
316  nullptr));
317  } else {
318  param_ptrs.erase(param_ptrs.begin() + LITERALS); // TODO(alex): remove
319  checkCudaErrors(cuLaunchKernel(cu_func,
320  grid_size_x,
321  grid_size_y,
322  grid_size_z,
323  block_size_x,
324  block_size_y,
325  block_size_z,
326  shared_memory_size,
327  nullptr,
328  &param_ptrs[0],
329  nullptr));
330  }
332  executor_->registerActiveModule(native_code.second, device_id);
333  cuEventRecord(stop1, 0);
334  cuEventSynchronize(stop1);
335  executor_->unregisterActiveModule(native_code.second, device_id);
336  float milliseconds1 = 0;
337  cuEventElapsedTime(&milliseconds1, start1, stop1);
338  VLOG(1) << "Device " << std::to_string(device_id)
339  << ": launchGpuCode: group-by cuLaunchKernel: "
340  << std::to_string(milliseconds1) << " ms";
341  cuEventRecord(start2, 0);
342  }
343 
344  gpu_allocator_->copyFromDevice(reinterpret_cast<int8_t*>(error_codes.data()),
345  reinterpret_cast<int8_t*>(err_desc),
346  error_codes.size() * sizeof(error_codes[0]));
347  *error_code = aggregate_error_codes(error_codes);
348  if (*error_code > 0) {
349  return {};
350  }
351 
352  if (!render_allocator) {
354  query_buffers_->applyStreamingTopNOffsetGpu(data_mgr,
356  gpu_group_by_buffers,
357  ra_exe_unit,
358  total_thread_count,
359  device_id);
360  } else {
361  if (use_speculative_top_n(ra_exe_unit, query_mem_desc_)) {
362  try {
365  gpu_group_by_buffers,
366  data_mgr,
367  device_id);
368  } catch (const std::bad_alloc&) {
369  throw SpeculativeTopNFailed("Failed during in-place GPU sort.");
370  }
371  }
375  query_buffers_->compactProjectionBuffersGpu(
377  data_mgr,
378  gpu_group_by_buffers,
380  data_mgr, kernel_params[TOTAL_MATCHED], device_id),
381  device_id);
382  } else {
383  size_t num_allocated_rows{0};
384  if (ra_exe_unit.use_bump_allocator) {
385  num_allocated_rows = get_num_allocated_rows_from_gpu(
386  data_mgr, kernel_params[TOTAL_MATCHED], device_id);
387  // First, check the error code. If we ran out of slots, don't copy data back
388  // into the ResultSet or update ResultSet entry count
389  if (*error_code < 0) {
390  return {};
391  }
392  }
393  query_buffers_->copyGroupByBuffersFromGpu(
394  data_mgr,
396  ra_exe_unit.use_bump_allocator ? num_allocated_rows
398  gpu_group_by_buffers,
399  &ra_exe_unit,
400  block_size_x,
401  grid_size_x,
402  device_id,
403  can_sort_on_gpu && query_mem_desc_.hasKeylessHash());
404  if (num_allocated_rows) {
405  CHECK(ra_exe_unit.use_bump_allocator);
406  CHECK(!query_buffers_->result_sets_.empty());
407  query_buffers_->result_sets_.front()->updateStorageEntryCount(
408  num_allocated_rows);
409  }
410  }
411  } else {
412  query_buffers_->copyGroupByBuffersFromGpu(
413  data_mgr,
416  gpu_group_by_buffers,
417  &ra_exe_unit,
418  block_size_x,
419  grid_size_x,
420  device_id,
421  can_sort_on_gpu && query_mem_desc_.hasKeylessHash());
422  }
423  }
424  }
425  } else {
426  std::vector<CUdeviceptr> out_vec_dev_buffers;
427  const size_t agg_col_count{ra_exe_unit.estimator ? size_t(1) : init_agg_vals.size()};
428  // by default, non-grouped aggregate queries generate one result per available thread
429  // in the lifetime of (potentially multi-fragment) kernel execution.
430  // We can reduce these intermediate results internally in the device and hence have
431  // only one result per device, if GPU shared memory optimizations are enabled.
432  const auto num_results_per_agg_col =
433  shared_memory_size ? 1 : block_size_x * grid_size_x * num_fragments;
434  const auto output_buffer_size_per_agg = num_results_per_agg_col * sizeof(int64_t);
435  if (ra_exe_unit.estimator) {
436  estimator_result_set_.reset(new ResultSet(
437  ra_exe_unit.estimator, ExecutorDeviceType::GPU, device_id, data_mgr));
438  out_vec_dev_buffers.push_back(reinterpret_cast<CUdeviceptr>(
439  estimator_result_set_->getDeviceEstimatorBuffer()));
440  } else {
441  for (size_t i = 0; i < agg_col_count; ++i) {
442  CUdeviceptr out_vec_dev_buffer =
443  num_fragments ? reinterpret_cast<CUdeviceptr>(
444  gpu_allocator_->alloc(output_buffer_size_per_agg))
445  : 0;
446  out_vec_dev_buffers.push_back(out_vec_dev_buffer);
447  if (shared_memory_size) {
448  CHECK_EQ(output_buffer_size_per_agg, size_t(8));
449  gpu_allocator_->copyToDevice(reinterpret_cast<int8_t*>(out_vec_dev_buffer),
450  reinterpret_cast<const int8_t*>(&init_agg_vals[i]),
451  output_buffer_size_per_agg);
452  }
453  }
454  }
455  auto out_vec_dev_ptr = gpu_allocator_->alloc(agg_col_count * sizeof(CUdeviceptr));
456  gpu_allocator_->copyToDevice(out_vec_dev_ptr,
457  reinterpret_cast<int8_t*>(out_vec_dev_buffers.data()),
458  agg_col_count * sizeof(CUdeviceptr));
459  kernel_params[GROUPBY_BUF] = reinterpret_cast<CUdeviceptr>(out_vec_dev_ptr);
460  std::vector<void*> param_ptrs;
461  for (auto& param : kernel_params) {
462  param_ptrs.push_back(&param);
463  }
464 
466  cuEventRecord(stop0, 0);
467  cuEventSynchronize(stop0);
468  float milliseconds0 = 0;
469  cuEventElapsedTime(&milliseconds0, start0, stop0);
470  VLOG(1) << "Device " << std::to_string(device_id)
471  << ": launchGpuCode: prepare: " << std::to_string(milliseconds0) << " ms";
472  cuEventRecord(start1, 0);
473  }
474 
475  if (hoist_literals) {
476  checkCudaErrors(cuLaunchKernel(cu_func,
477  grid_size_x,
478  grid_size_y,
479  grid_size_z,
480  block_size_x,
481  block_size_y,
482  block_size_z,
483  shared_memory_size,
484  nullptr,
485  &param_ptrs[0],
486  nullptr));
487  } else {
488  param_ptrs.erase(param_ptrs.begin() + LITERALS); // TODO(alex): remove
489  checkCudaErrors(cuLaunchKernel(cu_func,
490  grid_size_x,
491  grid_size_y,
492  grid_size_z,
493  block_size_x,
494  block_size_y,
495  block_size_z,
496  shared_memory_size,
497  nullptr,
498  &param_ptrs[0],
499  nullptr));
500  }
501 
503  executor_->registerActiveModule(native_code.second, device_id);
504  cuEventRecord(stop1, 0);
505  cuEventSynchronize(stop1);
506  executor_->unregisterActiveModule(native_code.second, device_id);
507  float milliseconds1 = 0;
508  cuEventElapsedTime(&milliseconds1, start1, stop1);
509  VLOG(1) << "Device " << std::to_string(device_id)
510  << ": launchGpuCode: cuLaunchKernel: " << std::to_string(milliseconds1)
511  << " ms";
512  cuEventRecord(start2, 0);
513  }
514 
515  copy_from_gpu(data_mgr,
516  &error_codes[0],
517  err_desc,
518  error_codes.size() * sizeof(error_codes[0]),
519  device_id);
520  *error_code = aggregate_error_codes(error_codes);
521  if (*error_code > 0) {
522  return {};
523  }
524  if (ra_exe_unit.estimator) {
526  estimator_result_set_->syncEstimatorBuffer();
527  return {};
528  }
529  for (size_t i = 0; i < agg_col_count; ++i) {
530  int64_t* host_out_vec = new int64_t[output_buffer_size_per_agg];
531  copy_from_gpu(data_mgr,
532  host_out_vec,
533  out_vec_dev_buffers[i],
534  output_buffer_size_per_agg,
535  device_id);
536  out_vec.push_back(host_out_vec);
537  }
538  }
539  const auto count_distinct_bitmap_mem = query_buffers_->getCountDistinctBitmapPtr();
540  if (count_distinct_bitmap_mem) {
541  copy_from_gpu(data_mgr,
542  query_buffers_->getCountDistinctHostPtr(),
543  count_distinct_bitmap_mem,
544  query_buffers_->getCountDistinctBitmapBytes(),
545  device_id);
546  }
547 
549  cuEventRecord(stop2, 0);
550  cuEventSynchronize(stop2);
551  float milliseconds2 = 0;
552  cuEventElapsedTime(&milliseconds2, start2, stop2);
553  VLOG(1) << "Device " << std::to_string(device_id)
554  << ": launchGpuCode: finish: " << std::to_string(milliseconds2) << " ms";
555  }
556 
557  return out_vec;
558 #else
559  return {};
560 #endif
561 }
#define CHECK_EQ(x, y)
Definition: Logger.h:205
RenderAllocator * getRenderAllocator(size_t device_id)
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
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:75
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_
CHECK(cgen_state)
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: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
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 DEBUG_TIMER(name)
Definition: Logger.h:313
std::unique_ptr< ResultSet > estimator_result_set_
bool g_enable_runtime_query_interrupt
Definition: Execute.cpp:108
#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: