#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 shared::TableKey &outer_table_key, 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, const size_t thread_idx, 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 CompilationContext compilation_context, 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 bool allow_runtime_interrupt, const std::vector< int8_t * > &join_hash_tables, RenderAllocatorMap *render_allocator_map, bool optimize_cuda_block_and_grid_sizes)

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< int8_t > &join_hash_tables, const int64_t num_rows_to_process=-1)

int64_t	getAggInitValForIndex (const size_t index) const

Private Types
enum	{ COL_BUFFERS, NUM_FRAGMENTS, LITERALS, NUM_ROWS, FRAG_ROW_OFFSETS, MAX_MATCHED, TOTAL_MATCHED, INIT_AGG_VALS, GROUPBY_BUF, ERROR_CODE, NUM_TABLES, JOIN_HASH_TABLES, ROW_FUNC_MGR, KERN_PARAM_COUNT }

Private Member Functions
std::vector< int8_t * >	prepareKernelParams (const std::vector< std::vector< const int8_t * >> &col_buffers, const std::vector< int8_t > &literal_buff, const std::vector< std::vector< int64_t >> &num_rows, const std::vector< std::vector< uint64_t >> &frag_offsets, const int32_t scan_limit, const std::vector< int64_t > &init_agg_vals, const std::vector< int32_t > &error_codes, const uint32_t num_tables, const std::vector< int8_t * > &join_hash_tables, Data_Namespace::DataMgr *data_mgr, const int device_id, const bool hoist_literals, const bool is_group_by) const

ResultSetPtr	groupBufferToDeinterleavedResults (const size_t i) const

Private Attributes
std::unique_ptr< DeviceAllocator >	gpu_allocator_

QueryMemoryDescriptor	query_mem_desc_

const Executor *	executor_

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< ResultSet >	estimator_result_set_

Friends
class	Executor

Detailed Description

Definition at line 39 of file QueryExecutionContext.h.

Member Enumeration Documentation

anonymous enum

private

Enumerator
COL_BUFFERS
NUM_FRAGMENTS
LITERALS
NUM_ROWS
FRAG_ROW_OFFSETS
MAX_MATCHED
TOTAL_MATCHED
INIT_AGG_VALS
GROUPBY_BUF
ERROR_CODE
NUM_TABLES
JOIN_HASH_TABLES
ROW_FUNC_MGR
KERN_PARAM_COUNT

Definition at line 101 of file QueryExecutionContext.h.

        {
     COL_BUFFERS,
     NUM_FRAGMENTS,
     LITERALS,
     NUM_ROWS,
     FRAG_ROW_OFFSETS,
     MAX_MATCHED,
     TOTAL_MATCHED,
     INIT_AGG_VALS,
     GROUPBY_BUF,
     ERROR_CODE,
     NUM_TABLES,
     JOIN_HASH_TABLES,
     ROW_FUNC_MGR,
     KERN_PARAM_COUNT,
   };

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 shared::TableKey &	outer_table_key,
		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,
		const size_t	thread_idx,
		RenderInfo *	render_info
	)

Definition at line 33 of file QueryExecutionContext.cpp.

References CHECK, getQueryEngineCudaStreamForDevice(), GPU, gpu_allocator_, heavyai::InSituFlagsOwnerInterface::isInSitu(), query_buffers_, query_mem_desc, RenderInfo::render_allocator_map_ptr, and sort_on_gpu().

     : query_mem_desc_(query_mem_desc)
     , executor_(executor)
     , device_type_(device_type)
     , dispatch_mode_(dispatch_mode)
     , row_set_mem_owner_(row_set_mem_owner)
     , output_columnar_(output_columnar) {
   CHECK(executor);
   auto data_mgr = executor->getDataMgr();
   if (device_type == ExecutorDeviceType::GPU) {
     gpu_allocator_ = std::make_unique<CudaAllocator>(
         data_mgr, device_id, getQueryEngineCudaStreamForDevice(device_id));
   }
 
   auto render_allocator_map = render_info && render_info->isInSitu()
                                   ? render_info->render_allocator_map_ptr.get()
                                   : nullptr;
   query_buffers_ = std::make_unique<QueryMemoryInitializer>(ra_exe_unit,
                                                             query_mem_desc,
                                                             device_id,
                                                             device_type,
                                                             dispatch_mode,
                                                             output_columnar,
                                                             sort_on_gpu,
                                                             outer_table_key,
                                                             num_rows,
                                                             col_buffers,
                                                             frag_offsets,
                                                             render_allocator_map,
                                                             render_info,
                                                             row_set_mem_owner,
                                                             gpu_allocator_.get(),
                                                             thread_idx,
                                                             executor);
 }

Here is the call graph for this function:

Member Function Documentation

int64_t QueryExecutionContext::getAggInitValForIndex ( const size_t index ) const

Definition at line 152 of file QueryExecutionContext.cpp.

References CHECK, and query_buffers_.

Referenced by Executor::executePlanWithoutGroupBy().

                                                                              {
   CHECK(query_buffers_);
   return query_buffers_->getAggInitValForIndex(index);
 }

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 157 of file QueryExecutionContext.cpp.

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

Referenced by Executor::executePlanWithGroupBy().

                                                        {
   auto timer = DEBUG_TIMER(__func__);
   std::vector<std::pair<ResultSetPtr, std::vector<size_t>>> results_per_sm;
   CHECK(query_buffers_);
   const auto group_by_buffers_size = query_buffers_->getNumBuffers();
   if (device_type_ == ExecutorDeviceType::CPU) {
     const size_t expected_num_buffers = query_mem_desc.hasVarlenOutput() ? 2 : 1;
     CHECK_EQ(expected_num_buffers, group_by_buffers_size);
     return groupBufferToResults(0);
   }
   const size_t step{query_mem_desc_.threadsShareMemory() ? executor_->blockSize() : 1};
   const size_t group_by_output_buffers_size =
       group_by_buffers_size - (query_mem_desc.hasVarlenOutput() ? 1 : 0);
   for (size_t i = 0; i < group_by_output_buffers_size; i += step) {
     results_per_sm.emplace_back(groupBufferToResults(i), std::vector<size_t>{});
   }
   CHECK(device_type_ == ExecutorDeviceType::GPU);
   return executor_->reduceMultiDeviceResults(
       ra_exe_unit, results_per_sm, row_set_mem_owner_, query_mem_desc);
 }

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 84 of file QueryExecutionContext.cpp.

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

Referenced by groupBufferToResults().

                           {
   CHECK(!output_columnar_);
   const auto& result_set = query_buffers_->getResultSet(i);
   auto deinterleaved_query_mem_desc =
       ResultSet::fixupQueryMemoryDescriptor(query_mem_desc_);
   deinterleaved_query_mem_desc.setHasInterleavedBinsOnGpu(false);
   deinterleaved_query_mem_desc.useConsistentSlotWidthSize(8);
 
   auto deinterleaved_result_set =
       std::make_shared<ResultSet>(result_set->getTargetInfos(),
                                   std::vector<ColumnLazyFetchInfo>{},
                                   std::vector<std::vector<const int8_t*>>{},
                                   std::vector<std::vector<int64_t>>{},
                                   std::vector<int64_t>{},
                                   ExecutorDeviceType::CPU,
                                   -1,
                                   deinterleaved_query_mem_desc,
                                   row_set_mem_owner_,
                                   executor_->blockSize(),
                                   executor_->gridSize());
   auto deinterleaved_storage =
       deinterleaved_result_set->allocateStorage(executor_->plan_state_->init_agg_vals_);
   auto deinterleaved_buffer =
       reinterpret_cast<int64_t*>(deinterleaved_storage->getUnderlyingBuffer());
   const auto rows_ptr = result_set->getStorage()->getUnderlyingBuffer();
   size_t deinterleaved_buffer_idx = 0;
   const size_t agg_col_count{query_mem_desc_.getSlotCount()};
   auto do_work = [&](const size_t bin_base_off) {
     std::vector<int64_t> agg_vals(agg_col_count, 0);
     memcpy(&agg_vals[0],
            &executor_->plan_state_->init_agg_vals_[0],
            agg_col_count * sizeof(agg_vals[0]));
     ResultSetStorage::reduceSingleRow(rows_ptr + bin_base_off,
                                       executor_->warpSize(),
                                       false,
                                       true,
                                       agg_vals,
                                       query_mem_desc_,
                                       result_set->getTargetInfos(),
                                       executor_->plan_state_->init_agg_vals_);
     for (size_t agg_idx = 0; agg_idx < agg_col_count;
          ++agg_idx, ++deinterleaved_buffer_idx) {
       deinterleaved_buffer[deinterleaved_buffer_idx] = agg_vals[agg_idx];
     }
   };
   if (g_enable_non_kernel_time_query_interrupt) {
     for (size_t bin_base_off = query_mem_desc_.getColOffInBytes(0), bin_idx = 0;
          bin_idx < result_set->entryCount();
          ++bin_idx, bin_base_off += query_mem_desc_.getColOffInBytesInNextBin(0)) {
       if (UNLIKELY((bin_idx & 0xFFFF) == 0 &&
                    executor_->checkNonKernelTimeInterrupted())) {
         throw std::runtime_error(
             "Query execution has interrupted during result set reduction");
       }
       do_work(bin_base_off);
     }
   } else {
     for (size_t bin_base_off = query_mem_desc_.getColOffInBytes(0), bin_idx = 0;
          bin_idx < result_set->entryCount();
          ++bin_idx, bin_base_off += query_mem_desc_.getColOffInBytesInNextBin(0)) {
       do_work(bin_base_off);
     }
   }
   query_buffers_->resetResultSet(i);
   return deinterleaved_result_set;
 }

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 180 of file QueryExecutionContext.cpp.

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

Referenced by getRowSet().

                                                                              {
   if (query_mem_desc_.interleavedBins(device_type_)) {
     return groupBufferToDeinterleavedResults(i);
   }
   return query_buffers_->getResultSetOwned(i);
 }

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< int8_t * > &	join_hash_tables,
		const int64_t	num_rows_to_process = `-1`
	)

Definition at line 556 of file QueryExecutionContext.cpp.

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

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

                                        {
   auto timer = DEBUG_TIMER(__func__);
   INJECT_TIMER(lauchCpuCode);
 
   CHECK(query_buffers_);
   const auto& init_agg_vals = query_buffers_->init_agg_vals_;
 
   std::vector<const int8_t**> multifrag_col_buffers;
   for (auto& col_buffer : col_buffers) {
     multifrag_col_buffers.push_back(col_buffer.empty() ? nullptr : col_buffer.data());
   }
   const int8_t*** multifrag_cols_ptr{
       multifrag_col_buffers.empty() ? nullptr : &multifrag_col_buffers[0]};
   const uint64_t num_fragments =
       multifrag_cols_ptr ? static_cast<uint64_t>(col_buffers.size()) : uint64_t(0);
   const auto num_out_frags = multifrag_cols_ptr ? num_fragments : uint64_t(0);
 
   const bool is_group_by{query_mem_desc_.isGroupBy()};
   std::vector<int64_t*> out_vec;
   if (ra_exe_unit.estimator) {
     // Subfragments collect the result from multiple runs in a single
     // result set.
     if (!estimator_result_set_) {
       estimator_result_set_.reset(
           new ResultSet(ra_exe_unit.estimator, ExecutorDeviceType::CPU, 0, nullptr));
     }
     out_vec.push_back(
         reinterpret_cast<int64_t*>(estimator_result_set_->getHostEstimatorBuffer()));
   } else {
     if (!is_group_by) {
       for (size_t i = 0; i < init_agg_vals.size(); ++i) {
         auto buff = new int64_t[num_out_frags];
         out_vec.push_back(static_cast<int64_t*>(buff));
       }
     }
   }
 
   CHECK_EQ(num_rows.size(), col_buffers.size());
   std::vector<int64_t> flatened_num_rows;
   for (auto& nums : num_rows) {
     flatened_num_rows.insert(flatened_num_rows.end(), nums.begin(), nums.end());
   }
   std::vector<uint64_t> flatened_frag_offsets;
   for (auto& offsets : frag_offsets) {
     flatened_frag_offsets.insert(
         flatened_frag_offsets.end(), offsets.begin(), offsets.end());
   }
   int64_t rowid_lookup_num_rows{*error_code ? *error_code + 1 : 0};
   int64_t* num_rows_ptr;
   if (num_rows_to_process > 0) {
     flatened_num_rows[0] = num_rows_to_process;
     num_rows_ptr = flatened_num_rows.data();
   } else {
     num_rows_ptr =
         rowid_lookup_num_rows ? &rowid_lookup_num_rows : flatened_num_rows.data();
   }
   int32_t total_matched_init{0};
 
   std::vector<int64_t> cmpt_val_buff;
   if (is_group_by) {
     cmpt_val_buff =
         compact_init_vals(align_to_int64(query_mem_desc_.getColsSize()) / sizeof(int64_t),
                           init_agg_vals,
                           query_mem_desc_);
   }
 
   RowFunctionManager mgr(this->executor_, ra_exe_unit);
   int8_t* row_func_mgr_ptr = reinterpret_cast<int8_t*>(&mgr);
 
   CHECK(native_code);
   const int64_t* join_hash_tables_ptr =
       join_hash_tables.size() == 1
           ? reinterpret_cast<const int64_t*>(join_hash_tables[0])
           : (join_hash_tables.size() > 1
                  ? reinterpret_cast<const int64_t*>(&join_hash_tables[0])
                  : nullptr);
   if (hoist_literals) {
     using agg_query = void (*)(const int8_t***,  // col_buffers
                                const uint64_t*,  // num_fragments
                                const int8_t*,    // literals
                                const int64_t*,   // num_rows
                                const uint64_t*,  // frag_row_offsets
                                const int32_t*,   // max_matched
                                int32_t*,         // total_matched
                                const int64_t*,   // init_agg_value
                                int64_t**,        // out
                                int32_t*,         // error_code
                                const uint32_t*,  // num_tables
                                const int64_t*,   // join_hash_tables_ptr
                                const int8_t*);   // row_func_mgr
     if (is_group_by) {
       reinterpret_cast<agg_query>(native_code->func())(
           multifrag_cols_ptr,
           &num_fragments,
           literal_buff.data(),
           num_rows_ptr,
           flatened_frag_offsets.data(),
           &scan_limit,
           &total_matched_init,
           cmpt_val_buff.data(),
           query_buffers_->getGroupByBuffersPtr(),
           error_code,
           &num_tables,
           join_hash_tables_ptr,
           row_func_mgr_ptr);
     } else {
       reinterpret_cast<agg_query>(native_code->func())(multifrag_cols_ptr,
                                                        &num_fragments,
                                                        literal_buff.data(),
                                                        num_rows_ptr,
                                                        flatened_frag_offsets.data(),
                                                        &scan_limit,
                                                        &total_matched_init,
                                                        init_agg_vals.data(),
                                                        out_vec.data(),
                                                        error_code,
                                                        &num_tables,
                                                        join_hash_tables_ptr,
                                                        row_func_mgr_ptr);
     }
   } else {
     using agg_query = void (*)(const int8_t***,  // col_buffers
                                const uint64_t*,  // num_fragments
                                const int64_t*,   // num_rows
                                const uint64_t*,  // frag_row_offsets
                                const int32_t*,   // max_matched
                                int32_t*,         // total_matched
                                const int64_t*,   // init_agg_value
                                int64_t**,        // out
                                int32_t*,         // error_code
                                const uint32_t*,  // num_tables
                                const int64_t*,   // join_hash_tables_ptr
                                const int8_t*);   // row_func_mgr
     if (is_group_by) {
       reinterpret_cast<agg_query>(native_code->func())(
           multifrag_cols_ptr,
           &num_fragments,
           num_rows_ptr,
           flatened_frag_offsets.data(),
           &scan_limit,
           &total_matched_init,
           cmpt_val_buff.data(),
           query_buffers_->getGroupByBuffersPtr(),
           error_code,
           &num_tables,
           join_hash_tables_ptr,
           row_func_mgr_ptr);
     } else {
       reinterpret_cast<agg_query>(native_code->func())(multifrag_cols_ptr,
                                                        &num_fragments,
                                                        num_rows_ptr,
                                                        flatened_frag_offsets.data(),
                                                        &scan_limit,
                                                        &total_matched_init,
                                                        init_agg_vals.data(),
                                                        out_vec.data(),
                                                        error_code,
                                                        &num_tables,
                                                        join_hash_tables_ptr,
                                                        row_func_mgr_ptr);
     }
   }
 
   if (ra_exe_unit.estimator) {
     return {};
   }
 
   if (rowid_lookup_num_rows && *error_code < 0) {
     *error_code = 0;
   }
 
   if (query_mem_desc_.useStreamingTopN()) {
     query_buffers_->applyStreamingTopNOffsetCpu(query_mem_desc_, ra_exe_unit);
   }
 
   if (query_mem_desc_.didOutputColumnar() &&
       query_mem_desc_.getQueryDescriptionType() == QueryDescriptionType::Projection) {
     query_buffers_->compactProjectionBuffersCpu(query_mem_desc_, total_matched_init);
   }
   return out_vec;
 }

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 CompilationContext *	compilation_context,
		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 bool	allow_runtime_interrupt,
		const std::vector< int8_t * > &	join_hash_tables,
		RenderAllocatorMap *	render_allocator_map,
		bool	optimize_cuda_block_and_grid_sizes
	)

Definition at line 206 of file QueryExecutionContext.cpp.

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

                                              {
   auto timer = DEBUG_TIMER(__func__);
   INJECT_TIMER(lauchGpuCode);
   CHECK(gpu_allocator_);
   CHECK(query_buffers_);
   CHECK(compilation_context);
   const auto& init_agg_vals = query_buffers_->init_agg_vals_;
 
   bool is_group_by{query_mem_desc_.isGroupBy()};
 
   RenderAllocator* render_allocator = nullptr;
   if (render_allocator_map) {
     render_allocator = render_allocator_map->getRenderAllocator(device_id);
   }
 
   auto kernel = create_device_kernel(compilation_context, device_id);
 
   std::vector<int64_t*> out_vec;
   uint32_t num_fragments = col_buffers.size();
   std::vector<int32_t> error_codes(grid_size_x * block_size_x);
 
   auto prepareClock = kernel->make_clock();
   auto launchClock = kernel->make_clock();
   auto finishClock = kernel->make_clock();
 
   if (g_enable_dynamic_watchdog || (allow_runtime_interrupt && !render_allocator)) {
     prepareClock->start();
   }
 
   if (g_enable_dynamic_watchdog) {
     kernel->initializeDynamicWatchdog(
         executor_->interrupted_.load(),
         executor_->deviceCycles(g_dynamic_watchdog_time_limit));
   }
 
   if (allow_runtime_interrupt && !render_allocator) {
     kernel->initializeRuntimeInterrupter(device_id);
   }
 
   auto kernel_params = prepareKernelParams(col_buffers,
                                            literal_buff,
                                            num_rows,
                                            frag_offsets,
                                            scan_limit,
                                            init_agg_vals,
                                            error_codes,
                                            num_tables,
                                            join_hash_tables,
                                            data_mgr,
                                            device_id,
                                            hoist_literals,
                                            is_group_by);
 
   CHECK_EQ(static_cast<size_t>(KERN_PARAM_COUNT), kernel_params.size());
   CHECK(!kernel_params[GROUPBY_BUF]);
 
   const unsigned block_size_y = 1;
   const unsigned block_size_z = 1;
   const unsigned grid_size_y = 1;
   const unsigned grid_size_z = 1;
   const auto total_thread_count = block_size_x * grid_size_x;
   const auto err_desc = kernel_params[ERROR_CODE];
   if (is_group_by) {
     CHECK(!(query_buffers_->getGroupByBuffersSize() == 0) || render_allocator);
     bool can_sort_on_gpu = query_mem_desc_.sortOnGpu();
     auto gpu_group_by_buffers =
         query_buffers_->createAndInitializeGroupByBufferGpu(ra_exe_unit,
                                                             query_mem_desc_,
                                                             kernel_params[INIT_AGG_VALS],
                                                             device_id,
                                                             dispatch_mode_,
                                                             block_size_x,
                                                             grid_size_x,
                                                             executor_->warpSize(),
                                                             can_sort_on_gpu,
                                                             output_columnar_,
                                                             render_allocator);
     const auto max_matched = static_cast<int32_t>(gpu_group_by_buffers.entry_count);
     gpu_allocator_->copyToDevice(
         kernel_params[MAX_MATCHED], &max_matched, sizeof(max_matched));
 
     kernel_params[GROUPBY_BUF] = gpu_group_by_buffers.ptrs;
     std::vector<void*> param_ptrs;
     for (auto& param : kernel_params) {
       param_ptrs.push_back(&param);
     }
 
     if (g_enable_dynamic_watchdog || (allow_runtime_interrupt && !render_allocator)) {
       auto prepareTime = prepareClock->stop();
       VLOG(1) << "Device " << std::to_string(device_id)
               << ": launchGpuCode: group-by prepare: " << std::to_string(prepareTime)
               << " ms";
       launchClock->start();
     }
 
     if (hoist_literals) {
       kernel->launch(grid_size_x,
                      grid_size_y,
                      grid_size_z,
                      block_size_x,
                      block_size_y,
                      block_size_z,
                      shared_memory_size,
                      &param_ptrs[0],
                      optimize_cuda_block_and_grid_sizes);
     } else {
       param_ptrs.erase(param_ptrs.begin() + LITERALS);  // TODO(alex): remove
       kernel->launch(grid_size_x,
                      grid_size_y,
                      grid_size_z,
                      block_size_x,
                      block_size_y,
                      block_size_z,
                      shared_memory_size,
                      &param_ptrs[0],
                      optimize_cuda_block_and_grid_sizes);
     }
     if (g_enable_dynamic_watchdog || (allow_runtime_interrupt && !render_allocator)) {
       auto launchTime = launchClock->stop();
       VLOG(1) << "Device " << std::to_string(device_id)
               << ": launchGpuCode: group-by cuLaunchKernel: "
               << std::to_string(launchTime) << " ms";
       finishClock->start();
     }
 
     gpu_allocator_->copyFromDevice(reinterpret_cast<int8_t*>(error_codes.data()),
                                    reinterpret_cast<int8_t*>(err_desc),
                                    error_codes.size() * sizeof(error_codes[0]));
     *error_code = aggregate_error_codes(error_codes);
     if (*error_code > 0) {
       return {};
     }
 
     if (!render_allocator) {
       if (query_mem_desc_.useStreamingTopN()) {
         query_buffers_->applyStreamingTopNOffsetGpu(data_mgr,
                                                     query_mem_desc_,
                                                     gpu_group_by_buffers,
                                                     ra_exe_unit,
                                                     total_thread_count,
                                                     device_id);
       } else {
         if (use_speculative_top_n(ra_exe_unit, query_mem_desc_)) {
           try {
             inplace_sort_gpu(ra_exe_unit.sort_info.order_entries,
                              query_mem_desc_,
                              gpu_group_by_buffers,
                              data_mgr,
                              device_id);
           } catch (const std::bad_alloc&) {
             throw SpeculativeTopNFailed("Failed during in-place GPU sort.");
           }
         }
         if (query_mem_desc_.getQueryDescriptionType() ==
             QueryDescriptionType::Projection) {
           if (query_mem_desc_.didOutputColumnar()) {
             query_buffers_->compactProjectionBuffersGpu(
                 query_mem_desc_,
                 data_mgr,
                 gpu_group_by_buffers,
                 get_num_allocated_rows_from_gpu(
                     *gpu_allocator_, kernel_params[TOTAL_MATCHED], device_id),
                 device_id);
           } else {
             size_t num_allocated_rows{0};
             if (ra_exe_unit.use_bump_allocator) {
               num_allocated_rows = get_num_allocated_rows_from_gpu(
                   *gpu_allocator_, kernel_params[TOTAL_MATCHED], device_id);
               // First, check the error code. If we ran out of slots, don't copy data back
               // into the ResultSet or update ResultSet entry count
               if (*error_code < 0) {
                 return {};
               }
             }
             query_buffers_->copyGroupByBuffersFromGpu(
                 *gpu_allocator_,
                 query_mem_desc_,
                 ra_exe_unit.use_bump_allocator ? num_allocated_rows
                                                : query_mem_desc_.getEntryCount(),
                 gpu_group_by_buffers,
                 &ra_exe_unit,
                 block_size_x,
                 grid_size_x,
                 device_id,
                 can_sort_on_gpu && query_mem_desc_.hasKeylessHash());
             if (num_allocated_rows) {
               CHECK(ra_exe_unit.use_bump_allocator);
               CHECK(!query_buffers_->result_sets_.empty());
               query_buffers_->result_sets_.front()->updateStorageEntryCount(
                   num_allocated_rows);
             }
           }
         } else {
           query_buffers_->copyGroupByBuffersFromGpu(
               *gpu_allocator_,
               query_mem_desc_,
               query_mem_desc_.getEntryCount(),
               gpu_group_by_buffers,
               &ra_exe_unit,
               block_size_x,
               grid_size_x,
               device_id,
               can_sort_on_gpu && query_mem_desc_.hasKeylessHash());
         }
       }
     }
   } else {
     std::vector<int8_t*> out_vec_dev_buffers;
     const size_t agg_col_count{ra_exe_unit.estimator ? size_t(1) : init_agg_vals.size()};
     // by default, non-grouped aggregate queries generate one result per available thread
     // in the lifetime of (potentially multi-fragment) kernel execution.
     // We can reduce these intermediate results internally in the device and hence have
     // only one result per device, if GPU shared memory optimizations are enabled.
     const auto num_results_per_agg_col =
         shared_memory_size ? 1 : block_size_x * grid_size_x * num_fragments;
     const auto output_buffer_size_per_agg = num_results_per_agg_col * sizeof(int64_t);
     if (ra_exe_unit.estimator) {
       estimator_result_set_.reset(new ResultSet(
           ra_exe_unit.estimator, ExecutorDeviceType::GPU, device_id, data_mgr));
       out_vec_dev_buffers.push_back(estimator_result_set_->getDeviceEstimatorBuffer());
     } else {
       for (size_t i = 0; i < agg_col_count; ++i) {
         int8_t* out_vec_dev_buffer =
             num_fragments ? gpu_allocator_->alloc(output_buffer_size_per_agg) : nullptr;
         out_vec_dev_buffers.push_back(out_vec_dev_buffer);
         if (shared_memory_size) {
           CHECK_EQ(output_buffer_size_per_agg, size_t(8));
           gpu_allocator_->copyToDevice(reinterpret_cast<int8_t*>(out_vec_dev_buffer),
                                        reinterpret_cast<const int8_t*>(&init_agg_vals[i]),
                                        output_buffer_size_per_agg);
         }
       }
     }
     auto out_vec_dev_ptr = gpu_allocator_->alloc(agg_col_count * sizeof(int8_t*));
     gpu_allocator_->copyToDevice(out_vec_dev_ptr,
                                  reinterpret_cast<int8_t*>(out_vec_dev_buffers.data()),
                                  agg_col_count * sizeof(int8_t*));
     kernel_params[GROUPBY_BUF] = out_vec_dev_ptr;
     std::vector<void*> param_ptrs;
     for (auto& param : kernel_params) {
       param_ptrs.push_back(&param);
     }
 
     if (g_enable_dynamic_watchdog || (allow_runtime_interrupt && !render_allocator)) {
       auto prepareTime = prepareClock->stop();
 
       VLOG(1) << "Device " << std::to_string(device_id)
               << ": launchGpuCode: prepare: " << std::to_string(prepareTime) << " ms";
       launchClock->start();
     }
 
     if (hoist_literals) {
       kernel->launch(grid_size_x,
                      grid_size_y,
                      grid_size_z,
                      block_size_x,
                      block_size_y,
                      block_size_z,
                      shared_memory_size,
                      &param_ptrs[0],
                      optimize_cuda_block_and_grid_sizes);
     } else {
       param_ptrs.erase(param_ptrs.begin() + LITERALS);  // TODO(alex): remove
       kernel->launch(grid_size_x,
                      grid_size_y,
                      grid_size_z,
                      block_size_x,
                      block_size_y,
                      block_size_z,
                      shared_memory_size,
                      &param_ptrs[0],
                      optimize_cuda_block_and_grid_sizes);
     }
 
     if (g_enable_dynamic_watchdog || (allow_runtime_interrupt && !render_allocator)) {
       auto launchTime = launchClock->stop();
       VLOG(1) << "Device " << std::to_string(device_id)
               << ": launchGpuCode: cuLaunchKernel: " << std::to_string(launchTime)
               << " ms";
       finishClock->start();
     }
 
     gpu_allocator_->copyFromDevice(
         &error_codes[0], err_desc, error_codes.size() * sizeof(error_codes[0]));
     *error_code = aggregate_error_codes(error_codes);
     if (*error_code > 0) {
       return {};
     }
     if (ra_exe_unit.estimator) {
       CHECK(estimator_result_set_);
       estimator_result_set_->syncEstimatorBuffer();
       return {};
     }
     for (size_t i = 0; i < agg_col_count; ++i) {
       int64_t* host_out_vec = new int64_t[output_buffer_size_per_agg];
       gpu_allocator_->copyFromDevice(
           host_out_vec, out_vec_dev_buffers[i], output_buffer_size_per_agg);
       out_vec.push_back(host_out_vec);
     }
   }
   const auto count_distinct_bitmap_mem = query_buffers_->getCountDistinctBitmapPtr();
   if (count_distinct_bitmap_mem) {
     gpu_allocator_->copyFromDevice(query_buffers_->getCountDistinctHostPtr(),
                                    reinterpret_cast<void*>(count_distinct_bitmap_mem),
                                    query_buffers_->getCountDistinctBitmapBytes());
   }
 
   const auto varlen_output_gpu_buf = query_buffers_->getVarlenOutputPtr();
   if (varlen_output_gpu_buf) {
     CHECK(query_mem_desc_.varlenOutputBufferElemSize());
     const size_t varlen_output_buf_bytes =
         query_mem_desc_.getEntryCount() *
         query_mem_desc_.varlenOutputBufferElemSize().value();
     CHECK(query_buffers_->getVarlenOutputHostPtr());
     gpu_allocator_->copyFromDevice(query_buffers_->getVarlenOutputHostPtr(),
                                    reinterpret_cast<void*>(varlen_output_gpu_buf),
                                    varlen_output_buf_bytes);
   }
 
   if (g_enable_dynamic_watchdog || (allow_runtime_interrupt && !render_allocator)) {
     if (allow_runtime_interrupt) {
       kernel->resetRuntimeInterrupter(device_id);
     }
     auto finishTime = finishClock->stop();
     VLOG(1) << "Device " << std::to_string(device_id)
             << ": launchGpuCode: finish: " << std::to_string(finishTime) << " ms";
   }
 
   return out_vec;
 }

Here is the call graph for this function:

Here is the caller graph for this function:

std::vector< int8_t * > QueryExecutionContext::prepareKernelParams	(	const std::vector< std::vector< const int8_t * >> &	col_buffers,
		const std::vector< int8_t > &	literal_buff,
		const std::vector< std::vector< int64_t >> &	num_rows,
		const std::vector< std::vector< uint64_t >> &	frag_offsets,
		const int32_t	scan_limit,
		const std::vector< int64_t > &	init_agg_vals,
		const std::vector< int32_t > &	error_codes,
		const uint32_t	num_tables,
		const std::vector< int8_t * > &	join_hash_tables,
		Data_Namespace::DataMgr *	data_mgr,
		const int	device_id,
		const bool	hoist_literals,
		const bool	is_group_by
	)		const

private

Definition at line 750 of file QueryExecutionContext.cpp.

References align_to_int64(), CHECK, CHECK_EQ, COL_BUFFERS, compact_init_vals(), ERROR_CODE, FRAG_ROW_OFFSETS, QueryMemoryDescriptor::getColsSize(), gpu_allocator_, INIT_AGG_VALS, JOIN_HASH_TABLES, KERN_PARAM_COUNT, LITERALS, MAX_MATCHED, NUM_FRAGMENTS, NUM_ROWS, NUM_TABLES, output_columnar_, query_buffers_, query_mem_desc_, ROW_FUNC_MGR, and TOTAL_MATCHED.

Referenced by launchGpuCode().

                                   {
   CHECK(gpu_allocator_);
   std::vector<int8_t*> params(KERN_PARAM_COUNT, 0);
   const uint64_t num_fragments = static_cast<uint64_t>(col_buffers.size());
   const size_t col_count{num_fragments > 0 ? col_buffers.front().size() : 0};
   if (col_count) {
     std::vector<int8_t*> multifrag_col_dev_buffers;
     for (auto frag_col_buffers : col_buffers) {
       std::vector<const int8_t*> col_dev_buffers;
       for (auto col_buffer : frag_col_buffers) {
         col_dev_buffers.push_back((int8_t*)col_buffer);
       }
       auto col_buffers_dev_ptr = gpu_allocator_->alloc(col_count * sizeof(int8_t*));
       gpu_allocator_->copyToDevice(
           col_buffers_dev_ptr, &col_dev_buffers[0], col_count * sizeof(int8_t*));
       multifrag_col_dev_buffers.push_back(col_buffers_dev_ptr);
     }
     params[COL_BUFFERS] = gpu_allocator_->alloc(num_fragments * sizeof(int8_t*));
 
     gpu_allocator_->copyToDevice(params[COL_BUFFERS],
                                  &multifrag_col_dev_buffers[0],
                                  num_fragments * sizeof(int8_t*));
   }
   params[NUM_FRAGMENTS] = gpu_allocator_->alloc(sizeof(uint64_t));
   gpu_allocator_->copyToDevice(params[NUM_FRAGMENTS], &num_fragments, sizeof(uint64_t));
 
   int8_t* literals_and_addr_mapping =
       gpu_allocator_->alloc(literal_buff.size() + 2 * sizeof(int64_t));
   CHECK_EQ(0, (int64_t)literals_and_addr_mapping % 8);
   std::vector<int64_t> additional_literal_bytes;
   const auto count_distinct_bitmap_mem = query_buffers_->getCountDistinctBitmapPtr();
   if (count_distinct_bitmap_mem) {
     // Store host and device addresses
     const auto count_distinct_bitmap_host_mem = query_buffers_->getCountDistinctHostPtr();
     CHECK(count_distinct_bitmap_host_mem);
     additional_literal_bytes.push_back(
         reinterpret_cast<int64_t>(count_distinct_bitmap_host_mem));
     additional_literal_bytes.push_back(static_cast<int64_t>(count_distinct_bitmap_mem));
     gpu_allocator_->copyToDevice(
         literals_and_addr_mapping,
         &additional_literal_bytes[0],
         additional_literal_bytes.size() * sizeof(additional_literal_bytes[0]));
   }
   params[LITERALS] = literals_and_addr_mapping + additional_literal_bytes.size() *
                                                      sizeof(additional_literal_bytes[0]);
   if (!literal_buff.empty()) {
     CHECK(hoist_literals);
     gpu_allocator_->copyToDevice(params[LITERALS], &literal_buff[0], literal_buff.size());
   }
   CHECK_EQ(num_rows.size(), col_buffers.size());
   std::vector<int64_t> flatened_num_rows;
   for (auto& nums : num_rows) {
     CHECK_EQ(nums.size(), num_tables);
     flatened_num_rows.insert(flatened_num_rows.end(), nums.begin(), nums.end());
   }
   params[NUM_ROWS] = gpu_allocator_->alloc(sizeof(int64_t) * flatened_num_rows.size());
   gpu_allocator_->copyToDevice(params[NUM_ROWS],
                                &flatened_num_rows[0],
                                sizeof(int64_t) * flatened_num_rows.size());
 
   CHECK_EQ(frag_offsets.size(), col_buffers.size());
   std::vector<int64_t> flatened_frag_offsets;
   for (auto& offsets : frag_offsets) {
     CHECK_EQ(offsets.size(), num_tables);
     flatened_frag_offsets.insert(
         flatened_frag_offsets.end(), offsets.begin(), offsets.end());
   }
   params[FRAG_ROW_OFFSETS] =
       gpu_allocator_->alloc(sizeof(int64_t) * flatened_frag_offsets.size());
   gpu_allocator_->copyToDevice(params[FRAG_ROW_OFFSETS],
                                &flatened_frag_offsets[0],
                                sizeof(int64_t) * flatened_num_rows.size());
 
   // Note that this will be overwritten if we are setting the entry count during group by
   // buffer allocation and initialization
   int32_t max_matched{scan_limit};
   params[MAX_MATCHED] = gpu_allocator_->alloc(sizeof(max_matched));
   gpu_allocator_->copyToDevice(params[MAX_MATCHED], &max_matched, sizeof(max_matched));
 
   int32_t total_matched{0};
   params[TOTAL_MATCHED] = gpu_allocator_->alloc(sizeof(total_matched));
   gpu_allocator_->copyToDevice(
       params[TOTAL_MATCHED], &total_matched, sizeof(total_matched));
 
   if (is_group_by && !output_columnar_) {
     auto cmpt_sz = align_to_int64(query_mem_desc_.getColsSize()) / sizeof(int64_t);
     auto cmpt_val_buff = compact_init_vals(cmpt_sz, init_agg_vals, query_mem_desc_);
     params[INIT_AGG_VALS] = gpu_allocator_->alloc(cmpt_sz * sizeof(int64_t));
     gpu_allocator_->copyToDevice(
         params[INIT_AGG_VALS], &cmpt_val_buff[0], cmpt_sz * sizeof(int64_t));
   } else {
     params[INIT_AGG_VALS] = gpu_allocator_->alloc(init_agg_vals.size() * sizeof(int64_t));
     gpu_allocator_->copyToDevice(
         params[INIT_AGG_VALS], &init_agg_vals[0], init_agg_vals.size() * sizeof(int64_t));
   }
 
   params[ERROR_CODE] = gpu_allocator_->alloc(error_codes.size() * sizeof(error_codes[0]));
   gpu_allocator_->copyToDevice(
       params[ERROR_CODE], &error_codes[0], error_codes.size() * sizeof(error_codes[0]));
 
   params[NUM_TABLES] = gpu_allocator_->alloc(sizeof(uint32_t));
   gpu_allocator_->copyToDevice(params[NUM_TABLES], &num_tables, sizeof(uint32_t));
 
   const auto hash_table_count = join_hash_tables.size();
   switch (hash_table_count) {
     case 0: {
       params[JOIN_HASH_TABLES] = 0;
       break;
     }
     case 1:
       params[JOIN_HASH_TABLES] = join_hash_tables[0];
       break;
     default: {
       params[JOIN_HASH_TABLES] =
           gpu_allocator_->alloc(hash_table_count * sizeof(int64_t));
       gpu_allocator_->copyToDevice(params[JOIN_HASH_TABLES],
                                    &join_hash_tables[0],
                                    hash_table_count * sizeof(int64_t));
       break;
     }
   }
 
   // RowFunctionManager is not supported in GPU. We just keep the argument
   // to avoid diverging from CPU generated code
   params[ROW_FUNC_MGR] = nullptr;
 
   return params;
 }

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 147 of file QueryExecutionContext.h.

Member Data Documentation

const ExecutorDeviceType QueryExecutionContext::device_type_

private

Definition at line 140 of file QueryExecutionContext.h.

Referenced by getRowSet(), and groupBufferToResults().

const ExecutorDispatchMode QueryExecutionContext::dispatch_mode_

private

Definition at line 141 of file QueryExecutionContext.h.

Referenced by launchGpuCode().

std::unique_ptr<ResultSet> QueryExecutionContext::estimator_result_set_

mutableprivate

Definition at line 145 of file QueryExecutionContext.h.

Referenced by Executor::executePlanWithoutGroupBy(), launchCpuCode(), and launchGpuCode().

const Executor* QueryExecutionContext::executor_

private

Definition at line 139 of file QueryExecutionContext.h.

Referenced by getRowSet(), groupBufferToDeinterleavedResults(), launchCpuCode(), and launchGpuCode().

std::unique_ptr<DeviceAllocator> QueryExecutionContext::gpu_allocator_

private

Definition at line 135 of file QueryExecutionContext.h.

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

const bool QueryExecutionContext::output_columnar_

private

Definition at line 143 of file QueryExecutionContext.h.

Referenced by groupBufferToDeinterleavedResults(), launchGpuCode(), and prepareKernelParams().

std::unique_ptr<QueryMemoryInitializer> QueryExecutionContext::query_buffers_

private

Definition at line 144 of file QueryExecutionContext.h.

Referenced by Executor::executePlanWithGroupBy(), Executor::executePlanWithoutGroupBy(), getAggInitValForIndex(), getRowSet(), groupBufferToDeinterleavedResults(), groupBufferToResults(), launchCpuCode(), launchGpuCode(), prepareKernelParams(), and QueryExecutionContext().

QueryMemoryDescriptor QueryExecutionContext::query_mem_desc_

private

Definition at line 138 of file QueryExecutionContext.h.

Referenced by Executor::executePlanWithGroupBy(), Executor::executePlanWithoutGroupBy(), getRowSet(), groupBufferToDeinterleavedResults(), groupBufferToResults(), launchCpuCode(), launchGpuCode(), and prepareKernelParams().

std::shared_ptr<RowSetMemoryOwner> QueryExecutionContext::row_set_mem_owner_

private

Definition at line 142 of file QueryExecutionContext.h.

Referenced by getRowSet(), and groupBufferToDeinterleavedResults().

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

/home/jenkins-slave/workspace/core-os-doxygen/QueryEngine/QueryExecutionContext.h
/home/jenkins-slave/workspace/core-os-doxygen/QueryEngine/QueryExecutionContext.cpp

Public Member Functions

Private Types

Private Member Functions

Private Attributes

Friends

Detailed Description

Member Enumeration Documentation

Constructor & Destructor Documentation

Member Function Documentation

Friends And Related Function Documentation

Member Data Documentation