QueryMemoryInitializer.cpp File Reference

#include "QueryMemoryInitializer.h"
#include "Execute.h"
#include "GpuInitGroups.h"
#include "Logger/Logger.h"
#include "OutputBufferInitialization.h"
#include "QueryEngine/QueryEngine.h"
#include "Shared/checked_alloc.h"
#include "StreamingTopN.h"
#include "Utils/FlatBuffer.h"

Include dependency graph for QueryMemoryInitializer.cpp:

Go to the source code of this file.

Classes
struct	anonymous_namespace{QueryMemoryInitializer.cpp}::AddNbytes

Namespaces
	anonymous_namespace{QueryMemoryInitializer.cpp}

Macros
#define	MAX_BUFFER_SIZE   0x100000000000ll

Functions
void	anonymous_namespace{QueryMemoryInitializer.cpp}::check_total_bitmap_memory (const QueryMemoryDescriptor &query_mem_desc)
std::pair< int64_t *, bool >	anonymous_namespace{QueryMemoryInitializer.cpp}::alloc_group_by_buffer (const size_t numBytes, RenderAllocatorMap *render_allocator_map, const size_t thread_idx, RowSetMemoryOwner *mem_owner, const bool reuse_existing_buffer_for_thread)
int64_t	anonymous_namespace{QueryMemoryInitializer.cpp}::get_consistent_frag_size (const std::vector< uint64_t > &frag_offsets)
std::vector< int64_t >	anonymous_namespace{QueryMemoryInitializer.cpp}::get_consistent_frags_sizes (const std::vector< std::vector< uint64_t >> &frag_offsets)
std::vector< int64_t >	anonymous_namespace{QueryMemoryInitializer.cpp}::get_consistent_frags_sizes (const std::vector< Analyzer::Expr * > &target_exprs, const std::vector< int64_t > &table_frag_sizes)
std::vector< std::vector < int64_t > >	anonymous_namespace{QueryMemoryInitializer.cpp}::get_col_frag_offsets (const std::vector< Analyzer::Expr * > &target_exprs, const std::vector< std::vector< uint64_t >> &table_frag_offsets)
int	anonymous_namespace{QueryMemoryInitializer.cpp}::get_input_idx (RelAlgExecutionUnit const &ra_exe_unit, const shared::TableKey &outer_table_key)
void	anonymous_namespace{QueryMemoryInitializer.cpp}::check_count_distinct_expr_metadata (const QueryMemoryDescriptor &query_mem_desc, const RelAlgExecutionUnit &ra_exe_unit)
QueryMemoryInitializer::TargetAggOpsMetadata	anonymous_namespace{QueryMemoryInitializer.cpp}::collect_target_expr_metadata (const QueryMemoryDescriptor &query_mem_desc, const RelAlgExecutionUnit &ra_exe_unit)
	num_buffers_ (1)
	varlen_output_buffer_ (0)
	varlen_output_buffer_host_ptr_ (nullptr)
	count_distinct_bitmap_device_mem_ptr_ (0)
	count_distinct_bitmap_mem_size_ (0)
	count_distinct_bitmap_host_crt_ptr_ (nullptr)
	count_distinct_bitmap_host_mem_ptr_ (nullptr)
	device_allocator_ (device_allocator)
	thread_idx_ (0)
template<typename T >
int8_t *	anonymous_namespace{QueryMemoryInitializer.cpp}::initColumnarBuffer (T *buffer_ptr, const T init_val, const uint32_t entry_count)
void	anonymous_namespace{QueryMemoryInitializer.cpp}::compact_projection_buffer_for_cpu_columnar (const QueryMemoryDescriptor &query_mem_desc, int8_t *projection_buffer, const size_t projection_count)

Variables
int64_t	g_bitmap_memory_limit {8LL * 1000 * 1000 * 1000}
	query_mem_desc

Macro Definition Documentation

#define MAX_BUFFER_SIZE   0x100000000000ll

Referenced by GroupByAndAggregate::getColRangeInfo(), and thread_idx_().

Function Documentation

count_distinct_bitmap_device_mem_ptr_

(

0

)

count_distinct_bitmap_host_crt_ptr_

(

nullptr

)

count_distinct_bitmap_host_mem_ptr_

(

nullptr

)

count_distinct_bitmap_mem_size_

(

0

)

device_allocator_

(

device_allocator

)

num_buffers_

(

1

)

Referenced by thread_idx_().

Here is the caller graph for this function:

thread_idx_

(

0

)

Definition at line 488 of file QueryMemoryInitializer.cpp.

References align_to_int64(), anonymous_namespace{QueryMemoryInitializer.cpp}::alloc_group_by_buffer(), CHECK, CHECK_EQ, anonymous_namespace{QueryMemoryInitializer.cpp}::get_col_frag_offsets(), anonymous_namespace{QueryMemoryInitializer.cpp}::get_consistent_frags_sizes(), MAX_BUFFER_SIZE, num_buffers_(), query_mem_desc, and target_exprs_to_infos().

Referenced by ResultSet::getThreadIdx().

                     {
  // Table functions output columnar, basically treat this as a projection
  const auto& consistent_frag_sizes = get_consistent_frags_sizes(frag_offsets);
  if (consistent_frag_sizes.empty()) {
    // No fragments in the input, no underlying buffers will be needed.
    return;
  }

  const size_t num_columns =
      query_mem_desc.getBufferColSlotCount();  // shouldn't we use getColCount() ???
  size_t total_group_by_buffer_size{0};
  for (size_t i = 0; i < num_columns; ++i) {
    auto ti = exe_unit.target_exprs[i]->get_type_info();
    if (ti.usesFlatBuffer()) {
      // See TableFunctionManager.h for info regarding flatbuffer
      // memory managment.
      auto slot_idx = query_mem_desc.getSlotIndexForSingleSlotCol(i);
      CHECK(query_mem_desc.checkSlotUsesFlatBufferFormat(slot_idx));
      checked_int64_t flatbuffer_size = query_mem_desc.getFlatBufferSize(slot_idx);
      try {
        total_group_by_buffer_size = align_to_int64(
            static_cast<int64_t>(total_group_by_buffer_size + flatbuffer_size));
      } catch (...) {
        throw OutOfHostMemory(std::numeric_limits<int64_t>::max() / 8);
      }
    } else {
      const checked_int64_t col_width = ti.get_size();
      try {
        const checked_int64_t group_buffer_size = col_width * num_rows_;
        total_group_by_buffer_size = align_to_int64(
            static_cast<int64_t>(group_buffer_size + total_group_by_buffer_size));
      } catch (...) {
        throw OutOfHostMemory(std::numeric_limits<int64_t>::max() / 8);
      }
    }
  }

#ifdef __SANITIZE_ADDRESS__
  // AddressSanitizer will reject allocation sizes above 1 TiB
#define MAX_BUFFER_SIZE 0x10000000000ll
#else
  // otherwise, we'll set the limit to 16 TiB, feel free to increase
  // the limit if needed
#define MAX_BUFFER_SIZE 0x100000000000ll
#endif

  if (total_group_by_buffer_size >= MAX_BUFFER_SIZE) {
    throw OutOfHostMemory(total_group_by_buffer_size);
  }

  CHECK_EQ(num_buffers_, size_t(1));
  auto group_by_buffer = alloc_group_by_buffer(total_group_by_buffer_size,
                                               nullptr,
                                               thread_idx_,
                                               row_set_mem_owner.get(),
                                               false)
                             .first;
  group_by_buffers_.push_back(group_by_buffer);

  const auto column_frag_offsets =
      get_col_frag_offsets(exe_unit.target_exprs, frag_offsets);
  const auto column_frag_sizes =
      get_consistent_frags_sizes(exe_unit.target_exprs, consistent_frag_sizes);
  result_sets_.emplace_back(
      new ResultSet(target_exprs_to_infos(exe_unit.target_exprs, query_mem_desc),
                    /*col_lazy_fetch_info=*/{},
                    col_buffers,
                    column_frag_offsets,
                    column_frag_sizes,
                    device_type,
                    device_id,
                    -1, /*thread_idx*/
                    ResultSet::fixupQueryMemoryDescriptor(query_mem_desc),
                    row_set_mem_owner_,
                    executor->blockSize(),
                    executor->gridSize()));
  result_sets_.back()->allocateStorage(reinterpret_cast<int8_t*>(group_by_buffer),
                                       init_agg_vals_);
}

Here is the call graph for this function:

Here is the caller graph for this function:

varlen_output_buffer_

(

0

)

varlen_output_buffer_host_ptr_

(

nullptr

)

Variable Documentation

int64_t g_bitmap_memory_limit {8LL * 1000 * 1000 * 1000}

Definition at line 28 of file QueryMemoryInitializer.cpp.

Referenced by anonymous_namespace{QueryMemoryInitializer.cpp}::check_total_bitmap_memory(), CommandLineOptions::fillDeveloperOptions(), anonymous_namespace{GroupByAndAggregate.cpp}::init_count_distinct_descriptors(), and InValuesBitmap::InValuesBitmap().

query_mem_desc

Definition at line 479 of file QueryMemoryInitializer.cpp.

Referenced by TableFunctionManager::allocate_output_buffers(), SpeculativeTopNMap::asRows(), anonymous_namespace{Execute.cpp}::build_row_for_empty_input(), GroupByAndAggregate::codegen(), Executor::codegenJoinLoops(), QueryCompilationDescriptor::compile(), Executor::createKernels(), RelAlgExecutor::executeLogicalValues(), Executor::executeTableFunction(), Executor::executeUpdate(), ResultSet::fixupQueryMemoryDescriptor(), anonymous_namespace{ResultSetReduction.cpp}::get_matching_group_value_reduction(), GroupByAndAggregate::initQueryMemoryDescriptor(), TableFunctionExecutionContext::launchGpuCode(), Executor::launchKernelsViaResourceMgr(), QueryExecutionContext::QueryExecutionContext(), ResultSetManager::reduce(), Executor::reduceMultiDeviceResultSets(), SqliteMemDatabase::runSelect(), and thread_idx_().