OmniSciDB  c1a53651b2
 All Classes Namespaces Files Functions Variables Typedefs Enumerations Enumerator Friends Macros Groups Pages
ExecutionKernel Class Reference

#include <ExecutionKernel.h>

+ Collaboration diagram for ExecutionKernel:

Public Member Functions

 ExecutionKernel (const RelAlgExecutionUnit &ra_exe_unit, const ExecutorDeviceType chosen_device_type, int chosen_device_id, const ExecutionOptions &eo, const ColumnFetcher &column_fetcher, const QueryCompilationDescriptor &query_comp_desc, const QueryMemoryDescriptor &query_mem_desc, const FragmentsList &frag_list, const ExecutorDispatchMode kernel_dispatch_mode, RenderInfo *render_info, const int64_t rowid_lookup_key)
 
void run (Executor *executor, const size_t thread_idx, SharedKernelContext &shared_context)
 

Public Attributes

const RelAlgExecutionUnitra_exe_unit_
 

Private Member Functions

void runImpl (Executor *executor, const size_t thread_idx, SharedKernelContext &shared_context)
 

Private Attributes

const ExecutorDeviceType chosen_device_type
 
int chosen_device_id
 
const ExecutionOptionseo
 
const ColumnFetchercolumn_fetcher
 
const QueryCompilationDescriptorquery_comp_desc
 
const QueryMemoryDescriptorquery_mem_desc
 
const FragmentsList frag_list
 
const ExecutorDispatchMode kernel_dispatch_mode
 
RenderInforender_info_
 
const int64_t rowid_lookup_key
 
ResultSetPtr device_results_
 

Friends

class KernelSubtask
 

Detailed Description

Definition at line 72 of file ExecutionKernel.h.

Constructor & Destructor Documentation

ExecutionKernel::ExecutionKernel ( const RelAlgExecutionUnit ra_exe_unit,
const ExecutorDeviceType  chosen_device_type,
int  chosen_device_id,
const ExecutionOptions eo,
const ColumnFetcher column_fetcher,
const QueryCompilationDescriptor query_comp_desc,
const QueryMemoryDescriptor query_mem_desc,
const FragmentsList frag_list,
const ExecutorDispatchMode  kernel_dispatch_mode,
RenderInfo render_info,
const int64_t  rowid_lookup_key 
)
inline

Definition at line 74 of file ExecutionKernel.h.

85  : ra_exe_unit_(ra_exe_unit)
88  , eo(eo)
89  , column_fetcher(column_fetcher)
90  , query_comp_desc(query_comp_desc)
91  , query_mem_desc(query_mem_desc)
94  , render_info_(render_info)
const ExecutionOptions & eo
const ExecutorDispatchMode kernel_dispatch_mode
const RelAlgExecutionUnit & ra_exe_unit_
const int64_t rowid_lookup_key
const ExecutorDeviceType chosen_device_type
RenderInfo * render_info_
const QueryMemoryDescriptor & query_mem_desc
const QueryCompilationDescriptor & query_comp_desc
const FragmentsList frag_list
const ColumnFetcher & column_fetcher

Member Function Documentation

void ExecutionKernel::run ( Executor executor,
const size_t  thread_idx,
SharedKernelContext shared_context 
)

Definition at line 129 of file ExecutionKernel.cpp.

References DEBUG_TIMER, Executor::ERR_COLUMNAR_CONVERSION_NOT_SUPPORTED, Executor::ERR_OUT_OF_CPU_MEM, Executor::ERR_OUT_OF_GPU_MEM, Executor::ERR_OUT_OF_RENDER_MEM, Executor::ERR_STRING_CONST_IN_RESULTSET, Executor::ERR_TOO_MANY_LITERALS, QueryMemoryDescriptor::getQueryDescriptionType(), INJECT_TIMER, kernel_dispatch_mode, MultifragmentKernel, query_mem_desc, runImpl(), and OutOfHostMemory::what().

Referenced by Executor::executeUpdate(), and Executor::executeWorkUnitPerFragment().

131  {
132  DEBUG_TIMER("ExecutionKernel::run");
133  INJECT_TIMER(kernel_run);
134  try {
135  runImpl(executor, thread_idx, shared_context);
136  } catch (const OutOfHostMemory& e) {
138  } catch (const std::bad_alloc& e) {
140  } catch (const OutOfRenderMemory& e) {
142  } catch (const OutOfMemory& e) {
143  throw QueryExecutionError(
145  e.what(),
149  } catch (const ColumnarConversionNotSupported& e) {
151  } catch (const TooManyLiterals& e) {
153  } catch (const StringConstInResultSet& e) {
155  } catch (const QueryExecutionError& e) {
156  throw e;
157  }
158 }
const ExecutorDispatchMode kernel_dispatch_mode
static const int32_t ERR_TOO_MANY_LITERALS
Definition: Execute.h:1438
static const int32_t ERR_STRING_CONST_IN_RESULTSET
Definition: Execute.h:1439
static const int32_t ERR_COLUMNAR_CONVERSION_NOT_SUPPORTED
Definition: Execute.h:1437
#define INJECT_TIMER(DESC)
Definition: measure.h:93
static const int32_t ERR_OUT_OF_RENDER_MEM
Definition: Execute.h:1432
const QueryMemoryDescriptor & query_mem_desc
static const int32_t ERR_OUT_OF_GPU_MEM
Definition: Execute.h:1429
QueryDescriptionType getQueryDescriptionType() const
void runImpl(Executor *executor, const size_t thread_idx, SharedKernelContext &shared_context)
#define DEBUG_TIMER(name)
Definition: Logger.h:411
const char * what() const noexceptfinal
Definition: checked_alloc.h:39
static const int32_t ERR_OUT_OF_CPU_MEM
Definition: Execute.h:1433

+ Here is the call graph for this function:

+ Here is the caller graph for this function:

void ExecutionKernel::runImpl ( Executor executor,
const size_t  thread_idx,
SharedKernelContext shared_context 
)
private

Definition at line 160 of file ExecutionKernel.cpp.

References gpu_enabled::accumulate(), SharedKernelContext::addDeviceResults(), ExecutionOptions::allow_runtime_query_interrupt, CHECK, CHECK_EQ, CHECK_GE, CHECK_GT, CHECK_LT, chosen_device_id, chosen_device_type, column_fetcher, QueryFragmentDescriptor::computeAllTablesFragments(), CPU, Data_Namespace::CPU_LEVEL, device_results_, dynamic_watchdog_init(), SharedKernelContext::dynamic_watchdog_set, ExecutionOptions::dynamic_watchdog_time_limit, eo, Executor::ERR_OUT_OF_CPU_MEM, Executor::ERR_OUT_OF_GPU_MEM, RelAlgExecutionUnit::estimator, ExecutionOptions::executor_type, Extern, frag_list, g_cpu_sub_task_size, QueryCompilationDescriptor::getCompilationResult(), SharedKernelContext::getFragOffsets(), QueryMemoryDescriptor::getQueryDescriptionType(), getQueryEngineCudaStreamForDevice(), QueryMemoryDescriptor::getQueryExecutionContext(), SharedKernelContext::getQueryInfos(), GPU, Data_Namespace::GPU_LEVEL, RelAlgExecutionUnit::groupby_exprs, QueryCompilationDescriptor::hoistLiterals(), logger::INFO, RelAlgExecutionUnit::input_descs, heavyai::InSituFlagsOwnerInterface::isInSitu(), kernel_dispatch_mode, KernelPerFragment, LOG, Executor::max_gpu_count, MultifragmentKernel, Native, anonymous_namespace{ExecutionKernel.cpp}::need_to_hold_chunk(), ExecutionOptions::optimize_cuda_block_and_grid_sizes, CompilationResult::output_columnar, Projection, query_comp_desc, anonymous_namespace{ExecutionKernel.cpp}::query_has_inner_join(), query_mem_desc, ra_exe_unit_, render_info_, rowid_lookup_key, run_query_external(), RelAlgExecutionUnit::scan_limit, serialize_to_sql(), QueryMemoryDescriptor::sortOnGpu(), RelAlgExecutionUnit::target_exprs, target_exprs_to_infos(), to_string(), RelAlgExecutionUnit::union_all, VLOG, and ExecutionOptions::with_dynamic_watchdog.

Referenced by run().

162  {
163  CHECK(executor);
164  const auto memory_level = chosen_device_type == ExecutorDeviceType::GPU
167  CHECK_GE(frag_list.size(), size_t(1));
168  // frag_list[0].table_id is how we tell which query we are running for UNION ALL.
169  const auto& outer_table_key = ra_exe_unit_.union_all
170  ? frag_list[0].table_key
171  : ra_exe_unit_.input_descs[0].getTableKey();
172  CHECK_EQ(frag_list[0].table_key, outer_table_key);
173  const auto& outer_tab_frag_ids = frag_list[0].fragment_ids;
174 
177 
178  auto data_mgr = executor->getDataMgr();
179 
180  // need to own them while query executes
181  auto chunk_iterators_ptr = std::make_shared<std::list<ChunkIter>>();
182  std::list<std::shared_ptr<Chunk_NS::Chunk>> chunks;
183  std::unique_ptr<std::lock_guard<std::mutex>> gpu_lock;
184  std::unique_ptr<CudaAllocator> device_allocator;
186  gpu_lock.reset(
187  new std::lock_guard<std::mutex>(executor->gpu_exec_mutex_[chosen_device_id]));
188  device_allocator = std::make_unique<CudaAllocator>(
189  data_mgr, chosen_device_id, getQueryEngineCudaStreamForDevice(chosen_device_id));
190  }
191  std::shared_ptr<FetchResult> fetch_result(new FetchResult);
192  try {
193  std::map<shared::TableKey, const TableFragments*> all_tables_fragments;
195  all_tables_fragments, ra_exe_unit_, shared_context.getQueryInfos());
196 
197  *fetch_result = ra_exe_unit_.union_all
198  ? executor->fetchUnionChunks(column_fetcher,
199  ra_exe_unit_,
201  memory_level,
202  all_tables_fragments,
203  frag_list,
204  *chunk_iterators_ptr,
205  chunks,
206  device_allocator.get(),
207  thread_idx,
209  : executor->fetchChunks(column_fetcher,
210  ra_exe_unit_,
212  memory_level,
213  all_tables_fragments,
214  frag_list,
215  *chunk_iterators_ptr,
216  chunks,
217  device_allocator.get(),
218  thread_idx,
220  if (fetch_result->num_rows.empty()) {
221  return;
222  }
224  !shared_context.dynamic_watchdog_set.test_and_set(std::memory_order_acquire)) {
227  LOG(INFO) << "Dynamic Watchdog budget: CPU: "
229  << std::to_string(cycle_budget) << " cycles";
230  }
231  } catch (const OutOfMemory&) {
232  throw QueryExecutionError(
238  return;
239  }
240 
242  if (ra_exe_unit_.input_descs.size() > 1) {
243  throw std::runtime_error("Joins not supported through external execution");
244  }
245  const auto query = serialize_to_sql(&ra_exe_unit_);
246  GroupByAndAggregate group_by_and_aggregate(executor,
248  ra_exe_unit_,
249  shared_context.getQueryInfos(),
250  executor->row_set_mem_owner_,
251  std::nullopt);
252  const auto query_mem_desc =
253  group_by_and_aggregate.initQueryMemoryDescriptor(false, 0, 8, nullptr, false);
255  query,
256  *fetch_result,
257  executor->plan_state_.get(),
261  executor});
262  shared_context.addDeviceResults(std::move(device_results_), outer_tab_frag_ids);
263  return;
264  }
265  const CompilationResult& compilation_result = query_comp_desc.getCompilationResult();
266  std::unique_ptr<QueryExecutionContext> query_exe_context_owned;
267  const bool do_render = render_info_ && render_info_->isInSitu();
268 
269  int64_t total_num_input_rows{-1};
271  query_mem_desc.getQueryDescriptionType() == QueryDescriptionType::Projection) {
272  total_num_input_rows = 0;
273  std::for_each(fetch_result->num_rows.begin(),
274  fetch_result->num_rows.end(),
275  [&total_num_input_rows](const std::vector<int64_t>& frag_row_count) {
276  total_num_input_rows = std::accumulate(frag_row_count.begin(),
277  frag_row_count.end(),
278  total_num_input_rows);
279  });
280  VLOG(2) << "total_num_input_rows=" << total_num_input_rows;
281  // TODO(adb): we may want to take this early out for all queries, but we are most
282  // likely to see this query pattern on the kernel per fragment path (e.g. with HAVING
283  // 0=1)
284  if (total_num_input_rows == 0) {
285  return;
286  }
287 
289  total_num_input_rows *= ra_exe_unit_.input_descs.size();
290  }
291  }
292 
293  uint32_t start_rowid{0};
294  if (rowid_lookup_key >= 0) {
295  if (!frag_list.empty()) {
296  const auto& all_frag_row_offsets = shared_context.getFragOffsets();
297  start_rowid = rowid_lookup_key -
298  all_frag_row_offsets[frag_list.begin()->fragment_ids.front()];
299  }
300  }
301 
302 #ifdef HAVE_TBB
303  bool can_run_subkernels = shared_context.getThreadPool() != nullptr;
304 
305  // Sub-tasks are supported for groupby queries and estimators only for now.
306  bool is_groupby =
307  (ra_exe_unit_.groupby_exprs.size() > 1) ||
308  (ra_exe_unit_.groupby_exprs.size() == 1 && ra_exe_unit_.groupby_exprs.front());
309  can_run_subkernels = can_run_subkernels && (is_groupby || ra_exe_unit_.estimator);
310 
311  // In case some column is lazily fetched, we cannot mix different fragments in a single
312  // ResultSet.
313  can_run_subkernels =
314  can_run_subkernels && !executor->hasLazyFetchColumns(ra_exe_unit_.target_exprs);
315 
316  // TODO: Use another structure to hold chunks. Currently, ResultSet holds them, but with
317  // sub-tasks chunk can be referenced by many ResultSets. So, some outer structure to
318  // hold all ResultSets and all chunks is required.
319  can_run_subkernels =
320  can_run_subkernels &&
322  chunks, ra_exe_unit_, std::vector<ColumnLazyFetchInfo>(), chosen_device_type);
323 
324  // TODO: check for literals? We serialize literals before execution and hold them in
325  // result sets. Can we simply do it once and holdin an outer structure?
326  if (can_run_subkernels) {
327  size_t total_rows = fetch_result->num_rows[0][0];
328  size_t sub_size = g_cpu_sub_task_size;
329 
330  for (size_t sub_start = start_rowid; sub_start < total_rows; sub_start += sub_size) {
331  sub_size = (sub_start + sub_size > total_rows) ? total_rows - sub_start : sub_size;
332  auto subtask = std::make_shared<KernelSubtask>(*this,
333  shared_context,
334  fetch_result,
335  chunk_iterators_ptr,
336  total_num_input_rows,
337  sub_start,
338  sub_size,
339  thread_idx);
340  shared_context.getThreadPool()->run(
341  [subtask, executor] { subtask->run(executor); });
342  }
343 
344  return;
345  }
346 #endif // HAVE_TBB
347 
349  try {
350  // std::unique_ptr<QueryExecutionContext> query_exe_context_owned
351  // has std::unique_ptr<QueryMemoryInitializer> query_buffers_
352  // has std::vector<std::unique_ptr<ResultSet>> result_sets_
353  // has std::unique_ptr<ResultSetStorage> storage_
354  // which are initialized and possibly allocated here.
355  query_exe_context_owned =
356  query_mem_desc.getQueryExecutionContext(ra_exe_unit_,
357  executor,
361  outer_table_key,
362  total_num_input_rows,
363  fetch_result->col_buffers,
364  fetch_result->frag_offsets,
365  executor->getRowSetMemoryOwner(),
366  compilation_result.output_columnar,
367  query_mem_desc.sortOnGpu(),
368  thread_idx,
369  do_render ? render_info_ : nullptr);
370  } catch (const OutOfHostMemory& e) {
371  throw QueryExecutionError(Executor::ERR_OUT_OF_CPU_MEM);
372  }
373  }
374  QueryExecutionContext* query_exe_context{query_exe_context_owned.get()};
375  CHECK(query_exe_context);
376  int32_t err{0};
377  bool optimize_cuda_block_and_grid_sizes =
380 
381  if (ra_exe_unit_.groupby_exprs.empty()) {
382  err = executor->executePlanWithoutGroupBy(ra_exe_unit_,
383  compilation_result,
388  fetch_result->col_buffers,
389  query_exe_context,
390  fetch_result->num_rows,
391  fetch_result->frag_offsets,
392  data_mgr,
394  start_rowid,
395  ra_exe_unit_.input_descs.size(),
397  do_render ? render_info_ : nullptr,
398  optimize_cuda_block_and_grid_sizes);
399  } else {
400  if (ra_exe_unit_.union_all) {
401  VLOG(1) << "outer_table_key=" << outer_table_key
402  << " ra_exe_unit_.scan_limit=" << ra_exe_unit_.scan_limit;
403  }
404  err = executor->executePlanWithGroupBy(ra_exe_unit_,
405  compilation_result,
409  fetch_result->col_buffers,
410  outer_tab_frag_ids,
411  query_exe_context,
412  fetch_result->num_rows,
413  fetch_result->frag_offsets,
414  data_mgr,
416  outer_table_key,
418  start_rowid,
419  ra_exe_unit_.input_descs.size(),
421  do_render ? render_info_ : nullptr,
422  optimize_cuda_block_and_grid_sizes);
423  }
424  if (device_results_) {
425  std::list<std::shared_ptr<Chunk_NS::Chunk>> chunks_to_hold;
426  for (const auto& chunk : chunks) {
427  if (need_to_hold_chunk(chunk.get(),
428  ra_exe_unit_,
429  device_results_->getLazyFetchInfo(),
431  chunks_to_hold.push_back(chunk);
432  }
433  }
434  device_results_->holdChunks(chunks_to_hold);
435  device_results_->holdChunkIterators(chunk_iterators_ptr);
436  } else {
437  VLOG(1) << "null device_results.";
438  }
439  if (err) {
440  throw QueryExecutionError(err);
441  }
442  shared_context.addDeviceResults(std::move(device_results_), outer_tab_frag_ids);
443 }
bool need_to_hold_chunk(const Chunk_NS::Chunk *chunk, const RelAlgExecutionUnit &ra_exe_unit, const std::vector< ColumnLazyFetchInfo > &lazy_fetch_info, const ExecutorDeviceType device_type)
std::vector< Analyzer::Expr * > target_exprs
#define CHECK_EQ(x, y)
Definition: Logger.h:301
std::atomic_flag dynamic_watchdog_set
const ExecutionOptions & eo
size_t g_cpu_sub_task_size
Definition: Execute.cpp:83
const std::vector< uint64_t > & getFragOffsets()
static const int max_gpu_count
Definition: Execute.h:1350
const std::optional< bool > union_all
#define LOG(tag)
Definition: Logger.h:285
const ExecutorDispatchMode kernel_dispatch_mode
const RelAlgExecutionUnit & ra_exe_unit_
const int64_t rowid_lookup_key
void addDeviceResults(ResultSetPtr &&device_results, std::vector< size_t > outer_table_fragment_ids)
std::vector< InputDescriptor > input_descs
const ExecutorDeviceType chosen_device_type
#define CHECK_GE(x, y)
Definition: Logger.h:306
const std::list< std::shared_ptr< Analyzer::Expr > > groupby_exprs
std::unique_ptr< ResultSet > run_query_external(const ExecutionUnitSql &sql, const FetchResult &fetch_result, const PlanState *plan_state, const ExternalQueryOutputSpec &output_spec)
RenderInfo * render_info_
#define CHECK_GT(x, y)
Definition: Logger.h:305
std::string to_string(char const *&&v)
ExecutorType executor_type
const QueryMemoryDescriptor & query_mem_desc
DEVICE auto accumulate(ARGS &&...args)
Definition: gpu_enabled.h:42
const QueryCompilationDescriptor & query_comp_desc
static void computeAllTablesFragments(std::map< shared::TableKey, const TableFragments * > &all_tables_fragments, const RelAlgExecutionUnit &ra_exe_unit, const std::vector< InputTableInfo > &query_infos)
const std::shared_ptr< Analyzer::Estimator > estimator
static const int32_t ERR_OUT_OF_GPU_MEM
Definition: Execute.h:1429
QueryDescriptionType getQueryDescriptionType() const
RUNTIME_EXPORT uint64_t dynamic_watchdog_init(unsigned ms_budget)
#define CHECK_LT(x, y)
Definition: Logger.h:303
const FragmentsList frag_list
ExecutionUnitSql serialize_to_sql(const RelAlgExecutionUnit *ra_exe_unit)
CUstream getQueryEngineCudaStreamForDevice(int device_num)
Definition: QueryEngine.cpp:7
bool optimize_cuda_block_and_grid_sizes
bool query_has_inner_join(const RelAlgExecutionUnit &ra_exe_unit)
const std::vector< InputTableInfo > & getQueryInfos() const
ResultSetPtr device_results_
#define CHECK(condition)
Definition: Logger.h:291
std::vector< TargetInfo > target_exprs_to_infos(const std::vector< Analyzer::Expr * > &targets, const QueryMemoryDescriptor &query_mem_desc)
unsigned dynamic_watchdog_time_limit
static const int32_t ERR_OUT_OF_CPU_MEM
Definition: Execute.h:1433
#define VLOG(n)
Definition: Logger.h:387
const ColumnFetcher & column_fetcher

+ Here is the call graph for this function:

+ Here is the caller graph for this function:

Friends And Related Function Documentation

friend class KernelSubtask
friend

Definition at line 121 of file ExecutionKernel.h.

Member Data Documentation

int ExecutionKernel::chosen_device_id
private

Definition at line 105 of file ExecutionKernel.h.

Referenced by runImpl().

const ExecutorDeviceType ExecutionKernel::chosen_device_type
private

Definition at line 104 of file ExecutionKernel.h.

Referenced by runImpl().

const ColumnFetcher& ExecutionKernel::column_fetcher
private

Definition at line 107 of file ExecutionKernel.h.

Referenced by runImpl().

ResultSetPtr ExecutionKernel::device_results_
private

Definition at line 115 of file ExecutionKernel.h.

Referenced by runImpl().

const ExecutionOptions& ExecutionKernel::eo
private

Definition at line 106 of file ExecutionKernel.h.

Referenced by runImpl().

const FragmentsList ExecutionKernel::frag_list
private

Definition at line 110 of file ExecutionKernel.h.

Referenced by runImpl().

const ExecutorDispatchMode ExecutionKernel::kernel_dispatch_mode
private

Definition at line 111 of file ExecutionKernel.h.

Referenced by run(), and runImpl().

const QueryCompilationDescriptor& ExecutionKernel::query_comp_desc
private

Definition at line 108 of file ExecutionKernel.h.

Referenced by runImpl().

const QueryMemoryDescriptor& ExecutionKernel::query_mem_desc
private

Definition at line 109 of file ExecutionKernel.h.

Referenced by run(), and runImpl().

const RelAlgExecutionUnit& ExecutionKernel::ra_exe_unit_

Definition at line 101 of file ExecutionKernel.h.

Referenced by runImpl().

RenderInfo* ExecutionKernel::render_info_
private

Definition at line 112 of file ExecutionKernel.h.

Referenced by runImpl().

const int64_t ExecutionKernel::rowid_lookup_key
private

Definition at line 113 of file ExecutionKernel.h.

Referenced by runImpl().


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