JoinHashTableQueryRuntime.cpp File Reference

#include <cmath>
#include <cstdio>
#include <cstdlib>
#include "Geospatial/CompressionRuntime.h"
#include "QueryEngine/CompareKeysInl.h"
#include "QueryEngine/MurmurHash.h"

Include dependency graph for JoinHashTableQueryRuntime.cpp:

This graph shows which files directly or indirectly include this file:

Go to the source code of this file.

Classes
struct	BufferRange
struct	Bounds

Namespaces
	anonymous_namespace{JoinHashTableQueryRuntime.cpp}

Functions
DEVICE bool	compare_to_key (const int8_t *entry, const int8_t *key, const size_t key_bytes)
template<class T >
DEVICE int64_t	get_matching_slot (const int8_t *hash_buff, const uint32_t h, const int8_t *key, const size_t key_bytes)
template<class T >
FORCE_INLINE DEVICE int64_t	baseline_hash_join_idx_impl (const int8_t *hash_buff, const int8_t *key, const size_t key_bytes, const size_t entry_count)
RUNTIME_EXPORT NEVER_INLINE DEVICE int64_t	baseline_hash_join_idx_32 (const int8_t *hash_buff, const int8_t *key, const size_t key_bytes, const size_t entry_count)
RUNTIME_EXPORT NEVER_INLINE DEVICE int64_t	baseline_hash_join_idx_64 (const int8_t *hash_buff, const int8_t *key, const size_t key_bytes, const size_t entry_count)
template<typename T >
FORCE_INLINE DEVICE int64_t	get_bucket_key_for_value_impl (const T value, const double bucket_size)
RUNTIME_EXPORT NEVER_INLINE DEVICE int64_t	get_bucket_key_for_range_double (const int8_t *range_bytes, const size_t range_component_index, const double bucket_size)
FORCE_INLINE DEVICE int64_t	get_bucket_key_for_range_compressed_impl (const int8_t *range, const size_t range_component_index, const double bucket_size)
RUNTIME_EXPORT NEVER_INLINE DEVICE int64_t	get_bucket_key_for_range_compressed (const int8_t *range, const size_t range_component_index, const double bucket_size)
template<typename T >
FORCE_INLINE DEVICE int64_t	get_composite_key_index_impl (const T *key, const size_t key_component_count, const T *composite_key_dict, const size_t entry_count)
RUNTIME_EXPORT NEVER_INLINE DEVICE int64_t	get_composite_key_index_32 (const int32_t *key, const size_t key_component_count, const int32_t *composite_key_dict, const size_t entry_count)
RUNTIME_EXPORT NEVER_INLINE DEVICE int64_t	get_composite_key_index_64 (const int64_t *key, const size_t key_component_count, const int64_t *composite_key_dict, const size_t entry_count)
RUNTIME_EXPORT NEVER_INLINE DEVICE int32_t	insert_sorted (int32_t *arr, size_t elem_count, int32_t elem)
RUNTIME_EXPORT ALWAYS_INLINE DEVICE int64_t	bbox_intersect_hash_join_idx (int64_t hash_buff, const int64_t key, const int64_t min_key, const int64_t max_key)
ALWAYS_INLINE DEVICE BufferRange	get_row_id_buffer_ptr (int64_t *hash_table_ptr, const int64_t *key, const int64_t key_component_count, const int64_t entry_count, const int64_t offset_buffer_ptr_offset, const int64_t sub_buff_size)
RUNTIME_EXPORT NEVER_INLINE DEVICE int64_t	get_candidate_rows (int32_t *out_arr, int32_t *error_code, const uint32_t max_arr_size, const int8_t *range_bytes, const int32_t range_component_index, const double bucket_size_x, const double bucket_size_y, const int32_t keys_count, const int64_t key_component_count, int64_t *hash_table_ptr, const int64_t entry_count, const int64_t offset_buffer_ptr_offset, const int64_t sub_buff_size, const int32_t max_bbox_overlaps_error_code)
RUNTIME_EXPORT NEVER_INLINE DEVICE int32_t	get_num_buckets_for_bounds (const int8_t *range_bytes, const int32_t range_component_index, const double bucket_size_x, const double bucket_size_y)

Variables
const int	anonymous_namespace{JoinHashTableQueryRuntime.cpp}::kNoMatch = -1
const int	anonymous_namespace{JoinHashTableQueryRuntime.cpp}::kNotPresent = -2

Function Documentation

RUNTIME_EXPORT NEVER_INLINE DEVICE int64_t baseline_hash_join_idx_32	(	const int8_t *	hash_buff,
		const int8_t *	key,
		const size_t	key_bytes,
		const size_t	entry_count
	)

Definition at line 84 of file JoinHashTableQueryRuntime.cpp.

                                                    {
  return baseline_hash_join_idx_impl<int32_t>(hash_buff, key, key_bytes, entry_count);
}

RUNTIME_EXPORT NEVER_INLINE DEVICE int64_t baseline_hash_join_idx_64	(	const int8_t *	hash_buff,
		const int8_t *	key,
		const size_t	key_bytes,
		const size_t	entry_count
	)

Definition at line 92 of file JoinHashTableQueryRuntime.cpp.

                                                    {
  return baseline_hash_join_idx_impl<int64_t>(hash_buff, key, key_bytes, entry_count);
}

template<class T >

FORCE_INLINE DEVICE int64_t baseline_hash_join_idx_impl	(	const int8_t *	hash_buff,
		const int8_t *	key,
		const size_t	key_bytes,
		const size_t	entry_count
	)

Definition at line 60 of file JoinHashTableQueryRuntime.cpp.

References anonymous_namespace{JoinHashTableQueryRuntime.cpp}::kNoMatch, and MurmurHash1().

                                                                                  {
  if (!entry_count) {
    return kNoMatch;
  }
  const uint32_t h = MurmurHash1(key, key_bytes, 0) % entry_count;
  int64_t matching_slot = get_matching_slot<T>(hash_buff, h, key, key_bytes);
  if (matching_slot != kNoMatch) {
    return matching_slot;
  }
  uint32_t h_probe = (h + 1) % entry_count;
  while (h_probe != h) {
    matching_slot = get_matching_slot<T>(hash_buff, h_probe, key, key_bytes);
    if (matching_slot != kNoMatch) {
      return matching_slot;
    }
    h_probe = (h_probe + 1) % entry_count;
  }
  return kNoMatch;
}

Here is the call graph for this function:

RUNTIME_EXPORT ALWAYS_INLINE DEVICE int64_t bbox_intersect_hash_join_idx	(	int64_t	hash_buff,
		const int64_t	key,
		const int64_t	min_key,
		const int64_t	max_key
	)

Definition at line 203 of file JoinHashTableQueryRuntime.cpp.

Referenced by get_row_id_buffer_ptr().

                                                    {
  if (key >= min_key && key <= max_key) {
    return *(reinterpret_cast<int32_t*>(hash_buff) + (key - min_key));
  }
  return -1;
}

Here is the caller graph for this function:

DEVICE bool compare_to_key	(	const int8_t *	entry,
		const int8_t *	key,
		const size_t	key_bytes
	)

Definition at line 25 of file JoinHashTableQueryRuntime.cpp.

Referenced by get_matching_slot().

                                                   {
  for (size_t i = 0; i < key_bytes; ++i) {
    if (entry[i] != key[i]) {
      return false;
    }
  }
  return true;
}

Here is the caller graph for this function:

RUNTIME_EXPORT NEVER_INLINE DEVICE int64_t get_bucket_key_for_range_compressed	(	const int8_t *	range,
		const size_t	range_component_index,
		const double	bucket_size
	)

Definition at line 130 of file JoinHashTableQueryRuntime.cpp.

References get_bucket_key_for_range_compressed_impl().

                                                              {
  return get_bucket_key_for_range_compressed_impl(
      range, range_component_index, bucket_size);
}

Here is the call graph for this function:

FORCE_INLINE DEVICE int64_t get_bucket_key_for_range_compressed_impl	(	const int8_t *	range,
		const size_t	range_component_index,
		const double	bucket_size
	)

Definition at line 114 of file JoinHashTableQueryRuntime.cpp.

References Geospatial::decompress_latitude_coord_geoint32(), Geospatial::decompress_longitude_coord_geoint32(), and get_bucket_key_for_value_impl().

Referenced by get_bucket_key_for_range_compressed().

                                                                   {
  const auto range_ptr = reinterpret_cast<const int32_t*>(range);
  if (range_component_index % 2 == 0) {
    return get_bucket_key_for_value_impl(
        Geospatial::decompress_longitude_coord_geoint32(range_ptr[range_component_index]),
        bucket_size);
  } else {
    return get_bucket_key_for_value_impl(
        Geospatial::decompress_latitude_coord_geoint32(range_ptr[range_component_index]),
        bucket_size);
  }
}

Here is the call graph for this function:

Here is the caller graph for this function:

RUNTIME_EXPORT NEVER_INLINE DEVICE int64_t get_bucket_key_for_range_double	(	const int8_t *	range_bytes,
		const size_t	range_component_index,
		const double	bucket_size
	)

Definition at line 106 of file JoinHashTableQueryRuntime.cpp.

References get_bucket_key_for_value_impl().

                                                          {
  const auto range = reinterpret_cast<const double*>(range_bytes);
  return get_bucket_key_for_value_impl(range[range_component_index], bucket_size);
}

Here is the call graph for this function:

template<typename T >

FORCE_INLINE DEVICE int64_t get_bucket_key_for_value_impl	(	const T	value,
		const double	bucket_size
	)

Definition at line 100 of file JoinHashTableQueryRuntime.cpp.

Referenced by get_bucket_key_for_range_compressed_impl(), and get_bucket_key_for_range_double().

                                                                                    {
  return static_cast<int64_t>(floor(static_cast<double>(value) * bucket_size));
}

Here is the caller graph for this function:

RUNTIME_EXPORT NEVER_INLINE DEVICE int64_t get_candidate_rows	(	int32_t *	out_arr,
		int32_t *	error_code,
		const uint32_t	max_arr_size,
		const int8_t *	range_bytes,
		const int32_t	range_component_index,
		const double	bucket_size_x,
		const double	bucket_size_y,
		const int32_t	keys_count,
		const int64_t	key_component_count,
		int64_t *	hash_table_ptr,
		const int64_t	entry_count,
		const int64_t	offset_buffer_ptr_offset,
		const int64_t	sub_buff_size,
		const int32_t	max_bbox_overlaps_error_code
	)

Getting overlapping candidates for the bounding box intersection algorithm works as follows:

1. Take the bounds of the Polygon and use the bucket sizes to split the bounding box into the hash keys it falls into.
2. Iterate over the keys and look them up in the hash table.
3. When looking up the values of each key, we use a series of offsets to get to the array of row ids.
4. Since it is possible (likely) we encounter the same row id in several buckets, we need to ensure we remove the duplicates we encounter. A simple ordered insertion is used which ignores duplicate values. Since the N elements we insert can be considered relatively small (N < 200) this exhibits a good tradeoff to conserve space since we are constrained by the stack size on the GPU.

RETURNS: Unique Row IDs are placed on the fixed size stack array that is passed in as out_arr. The number of row ids in this array is returned.

Definition at line 290 of file JoinHashTableQueryRuntime.cpp.

References get_row_id_buffer_ptr(), and insert_sorted().

                                                               {
  const auto range = reinterpret_cast<const double*>(range_bytes);

  size_t elem_count = 0;

  const auto bounds = Bounds{range[0], range[1], range[2], range[3]};

  for (int64_t x = floor(bounds.min_X * bucket_size_x);
       x <= floor(bounds.max_X * bucket_size_x);
       x++) {
    for (int64_t y = floor(bounds.min_Y * bucket_size_y);
         y <= floor(bounds.max_Y * bucket_size_y);
         y++) {
      int64_t cur_key[2];
      cur_key[0] = static_cast<int64_t>(x);
      cur_key[1] = static_cast<int64_t>(y);

      const auto buffer_range = get_row_id_buffer_ptr(hash_table_ptr,
                                                      cur_key,
                                                      key_component_count,
                                                      entry_count,
                                                      offset_buffer_ptr_offset,
                                                      sub_buff_size);

      for (int64_t j = 0; j < buffer_range.element_count; j++) {
        const auto rowid = buffer_range.buffer[j];
        elem_count += insert_sorted(out_arr, elem_count, rowid);
        if (elem_count > max_arr_size) {
          *error_code = max_bbox_overlaps_error_code;
          return 0;
        }
      }
    }
  }

  return elem_count;
}

Here is the call graph for this function:

RUNTIME_EXPORT NEVER_INLINE DEVICE int64_t get_composite_key_index_32	(	const int32_t *	key,
		const size_t	key_component_count,
		const int32_t *	composite_key_dict,
		const size_t	entry_count
	)

Definition at line 162 of file JoinHashTableQueryRuntime.cpp.

References get_composite_key_index_impl().

                                                     {
  return get_composite_key_index_impl(
      key, key_component_count, composite_key_dict, entry_count);
}

Here is the call graph for this function:

RUNTIME_EXPORT NEVER_INLINE DEVICE int64_t get_composite_key_index_64	(	const int64_t *	key,
		const size_t	key_component_count,
		const int64_t *	composite_key_dict,
		const size_t	entry_count
	)

Definition at line 171 of file JoinHashTableQueryRuntime.cpp.

References get_composite_key_index_impl().

Referenced by get_row_id_buffer_ptr().

                                                     {
  return get_composite_key_index_impl(
      key, key_component_count, composite_key_dict, entry_count);
}

Here is the call graph for this function:

Here is the caller graph for this function:

template<typename T >

FORCE_INLINE DEVICE int64_t get_composite_key_index_impl	(	const T *	key,
		const size_t	key_component_count,
		const T *	composite_key_dict,
		const size_t	entry_count
	)

Definition at line 138 of file JoinHashTableQueryRuntime.cpp.

References get_invalid_key(), keys_are_equal(), MurmurHash1(), SUFFIX, and heavydb.dtypes::T.

Referenced by get_composite_key_index_32(), and get_composite_key_index_64().

                                                                                   {
  const uint32_t h = MurmurHash1(key, key_component_count * sizeof(T), 0) % entry_count;
  uint32_t off = h * key_component_count;
  if (keys_are_equal(&composite_key_dict[off], key, key_component_count)) {
    return h;
  }
  uint32_t h_probe = (h + 1) % entry_count;
  while (h_probe != h) {
    off = h_probe * key_component_count;
    if (keys_are_equal(&composite_key_dict[off], key, key_component_count)) {
      return h_probe;
    }
    if (composite_key_dict[off] == SUFFIX(get_invalid_key) < T > ()) {
      return -1;
    }
    h_probe = (h_probe + 1) % entry_count;
  }
  return -1;
}

Here is the call graph for this function:

Here is the caller graph for this function:

template<class T >

DEVICE int64_t get_matching_slot	(	const int8_t *	hash_buff,
		const uint32_t	h,
		const int8_t *	key,
		const size_t	key_bytes
	)

Definition at line 44 of file JoinHashTableQueryRuntime.cpp.

References compare_to_key(), get_invalid_key(), anonymous_namespace{JoinHashTableQueryRuntime.cpp}::kNoMatch, anonymous_namespace{JoinHashTableQueryRuntime.cpp}::kNotPresent, SUFFIX, and heavydb.dtypes::T.

                                                         {
  const auto lookup_result_ptr = hash_buff + h * (key_bytes + sizeof(T));
  if (compare_to_key(lookup_result_ptr, key, key_bytes)) {
    return *reinterpret_cast<const T*>(lookup_result_ptr + key_bytes);
  }
  if (*reinterpret_cast<const T*>(lookup_result_ptr) ==
      SUFFIX(get_invalid_key) < T > ()) {
    return kNotPresent;
  }
  return kNoMatch;
}

Here is the call graph for this function:

RUNTIME_EXPORT NEVER_INLINE DEVICE int32_t get_num_buckets_for_bounds	(	const int8_t *	range_bytes,
		const int32_t	range_component_index,
		const double	bucket_size_x,
		const double	bucket_size_y
	)

Definition at line 345 of file JoinHashTableQueryRuntime.cpp.

                                                       {
  const auto range = reinterpret_cast<const double*>(range_bytes);

  const auto bounds_min_x = range[0];
  const auto bounds_min_y = range[1];
  const auto bounds_max_x = range[2];
  const auto bounds_max_y = range[3];

  const auto num_x =
      floor(bounds_max_x * bucket_size_x) - floor(bounds_min_x * bucket_size_x);
  const auto num_y =
      floor(bounds_max_y * bucket_size_y) - floor(bounds_min_y * bucket_size_y);
  const auto num_buckets = (num_x + 1) * (num_y + 1);

  return num_buckets;
}

ALWAYS_INLINE DEVICE BufferRange get_row_id_buffer_ptr	(	int64_t *	hash_table_ptr,
		const int64_t *	key,
		const int64_t	key_component_count,
		const int64_t	entry_count,
		const int64_t	offset_buffer_ptr_offset,
		const int64_t	sub_buff_size
	)

Definition at line 219 of file JoinHashTableQueryRuntime.cpp.

References bbox_intersect_hash_join_idx(), and get_composite_key_index_64().

Referenced by get_candidate_rows().

                                                   {
  const int64_t min_key = 0;
  const int64_t max_key = entry_count - 1;

  auto key_idx =
      get_composite_key_index_64(key, key_component_count, hash_table_ptr, entry_count);

  if (key_idx < -1) {
    return BufferRange{nullptr, 0};
  }

  int8_t* one_to_many_ptr = reinterpret_cast<int8_t*>(hash_table_ptr);
  one_to_many_ptr += offset_buffer_ptr_offset;

  // Returns an index used to fetch row count and row ids.
  const auto slot = bbox_intersect_hash_join_idx(
      reinterpret_cast<int64_t>(one_to_many_ptr), key_idx, min_key, max_key);
  if (slot < 0) {
    return BufferRange{nullptr, 0};
  }

  // Offset into the row count section of buffer
  int8_t* count_ptr = one_to_many_ptr + sub_buff_size;

  const int64_t matched_row_count = bbox_intersect_hash_join_idx(
      reinterpret_cast<int64_t>(count_ptr), key_idx, min_key, max_key);

  // Offset into payload section, containing an array of row ids
  int32_t* rowid_buffer = (int32_t*)(one_to_many_ptr + 2 * sub_buff_size);
  const auto rowidoff_ptr = &rowid_buffer[slot];

  return BufferRange{rowidoff_ptr, matched_row_count};
}

Here is the call graph for this function:

Here is the caller graph for this function:

RUNTIME_EXPORT NEVER_INLINE DEVICE int32_t insert_sorted	(	int32_t *	arr,
		size_t	elem_count,
		int32_t	elem
	)

Definition at line 179 of file JoinHashTableQueryRuntime.cpp.

Referenced by get_candidate_rows().

                                                                                  {
  for (size_t i = 0; i < elem_count; i++) {
    if (elem == arr[i]) {
      return 0;
    }

    if (elem > arr[i]) {
      continue;
    }

    for (size_t j = elem_count; i < j; j--) {
      arr[j] = arr[j - 1];
    }
    arr[i] = elem;
    return 1;
  }

  arr[elem_count] = elem;
  return 1;
}

Here is the caller graph for this function: