Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

refactor(query): vectorized hash join #13103

Merged
merged 2 commits into from
Oct 8, 2023
+356 −364
Merged

refactor(query): vectorized hash join #13103

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
ab7139f
vectorized hash join
Dousir9 Oct 7, 2023
39e704c
Merge branch 'main' of github.com:datafuselabs/databend into vectoriz…
Dousir9 Oct 7, 2023
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
77 changes: 23 additions & 54 deletions src/common/hashtable/src/hashjoin_hashtable.rs
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -52,7 +52,7 @@ pub struct RawEntry<K> {
pub struct HashJoinHashTable<K: Keyable, A: Allocator + Clone = MmapAllocator> {
pub(crate) pointers: Box<[u64], A>,
pub(crate) atomic_pointers: *mut AtomicU64,
pub(crate) hash_mask: usize,
pub(crate) hash_mask: u64,
pub(crate) phantom: PhantomData<K>,
}

Expand All @@ -68,7 +68,7 @@ impl<K: Keyable, A: Allocator + Clone + Default> HashJoinHashTable<K, A> {
Box::new_zeroed_slice_in(capacity, Default::default()).assume_init()
},
atomic_pointers: std::ptr::null_mut(),
hash_mask: capacity - 1,
hash_mask: capacity as u64 - 1,
phantom: PhantomData,
};
hashtable.atomic_pointers = unsafe {
Expand All @@ -78,7 +78,7 @@ impl<K: Keyable, A: Allocator + Clone + Default> HashJoinHashTable<K, A> {
}

pub fn insert(&mut self, key: K, raw_entry_ptr: *mut RawEntry<K>) {
let index = key.hash() as usize & self.hash_mask;
let index = (key.hash() & self.hash_mask) as usize;
// # Safety
// `index` is less than the capacity of hash table.
let mut head = unsafe { (*self.atomic_pointers.add(index)).load(Ordering::Relaxed) };
Expand Down Expand Up @@ -107,73 +107,42 @@ where
{
type Key = K;

fn contains(&self, key_ref: &Self::Key) -> bool {
let index = key_ref.hash() as usize & self.hash_mask;
let mut raw_entry_ptr = self.pointers[index];
loop {
if raw_entry_ptr == 0 {
break;
}
let raw_entry = unsafe { &*(raw_entry_ptr as *mut RawEntry<K>) };
if key_ref == &raw_entry.key {
return true;
}
raw_entry_ptr = raw_entry.next;
}
false
// Using hashes to probe hash table and converting them in-place to pointers for memory reuse.
fn probe(&self, hashes: &mut [u64]) {
hashes
.iter_mut()
.for_each(|hash| *hash = self.pointers[(*hash & self.hash_mask) as usize]);
}

fn probe_hash_table(
&self,
key_ref: &Self::Key,
vec_ptr: *mut RowPtr,
mut occupied: usize,
capacity: usize,
) -> (usize, u64) {
let index = key_ref.hash() as usize & self.hash_mask;
let origin = occupied;
let mut raw_entry_ptr = self.pointers[index];
fn next_contains(&self, key: &Self::Key, mut ptr: u64) -> bool {
loop {
if raw_entry_ptr == 0 || occupied >= capacity {
if ptr == 0 {
break;
}
let raw_entry = unsafe { &*(raw_entry_ptr as *mut RawEntry<K>) };
if key_ref == &raw_entry.key {
// # Safety
// occupied is less than the capacity of vec_ptr.
unsafe {
std::ptr::copy_nonoverlapping(
&raw_entry.row_ptr as *const RowPtr,
vec_ptr.add(occupied),
1,
)
};
occupied += 1;
let raw_entry = unsafe { &*(ptr as *mut RawEntry<K>) };
if key == &raw_entry.key {
return true;
}
raw_entry_ptr = raw_entry.next;
}
if occupied > origin {
(occupied - origin, raw_entry_ptr)
} else {
(0, 0)
ptr = raw_entry.next;
}
false
}

fn next_incomplete_ptr(
fn next_probe(
&self,
key_ref: &Self::Key,
mut incomplete_ptr: u64,
key: &Self::Key,
mut ptr: u64,
vec_ptr: *mut RowPtr,
mut occupied: usize,
capacity: usize,
) -> (usize, u64) {
let origin = occupied;
loop {
if incomplete_ptr == 0 || occupied >= capacity {
if ptr == 0 || occupied >= capacity {
break;
}
let raw_entry = unsafe { &*(incomplete_ptr as *mut RawEntry<K>) };
if key_ref == &raw_entry.key {
let raw_entry = unsafe { &*(ptr as *mut RawEntry<K>) };
if key == &raw_entry.key {
// # Safety
// occupied is less than the capacity of vec_ptr.
unsafe {
Expand All @@ -185,10 +154,10 @@ where
};
occupied += 1;
}
incomplete_ptr = raw_entry.next;
ptr = raw_entry.next;
}
if occupied > origin {
(occupied - origin, incomplete_ptr)
(occupied - origin, ptr)
} else {
(0, 0)
}
Expand Down
101 changes: 29 additions & 72 deletions src/common/hashtable/src/hashjoin_string_hashtable.rs
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -34,7 +34,7 @@ pub struct StringRawEntry {
pub struct HashJoinStringHashTable<A: Allocator + Clone = MmapAllocator> {
pub(crate) pointers: Box<[u64], A>,
pub(crate) atomic_pointers: *mut AtomicU64,
pub(crate) hash_mask: usize,
pub(crate) hash_mask: u64,
}

unsafe impl<A: Allocator + Clone + Send> Send for HashJoinStringHashTable<A> {}
Expand All @@ -49,7 +49,7 @@ impl<A: Allocator + Clone + Default> HashJoinStringHashTable<A> {
Box::new_zeroed_slice_in(capacity, Default::default()).assume_init()
},
atomic_pointers: std::ptr::null_mut(),
hash_mask: capacity - 1,
hash_mask: capacity as u64 - 1,
};
hashtable.atomic_pointers = unsafe {
std::mem::transmute::<*mut u64, *mut AtomicU64>(hashtable.pointers.as_mut_ptr())
Expand All @@ -58,7 +58,7 @@ impl<A: Allocator + Clone + Default> HashJoinStringHashTable<A> {
}

pub fn insert(&mut self, key: &[u8], raw_entry_ptr: *mut StringRawEntry) {
let index = key.fast_hash() as usize & self.hash_mask;
let index = (key.fast_hash() & self.hash_mask) as usize;
// # Safety
// `index` is less than the capacity of hash table.
let mut head = unsafe { (*self.atomic_pointers.add(index)).load(Ordering::Relaxed) };
Expand All @@ -85,23 +85,28 @@ where A: Allocator + Clone + 'static
{
type Key = [u8];

fn contains(&self, key_ref: &Self::Key) -> bool {
let index = key_ref.fast_hash() as usize & self.hash_mask;
let mut raw_entry_ptr = self.pointers[index];
// Using hashes to probe hash table and converting them in-place to pointers for memory reuse.
fn probe(&self, hashes: &mut [u64]) {
hashes
.iter_mut()
.for_each(|hash| *hash = self.pointers[(*hash & self.hash_mask) as usize]);
}

fn next_contains(&self, key: &Self::Key, mut ptr: u64) -> bool {
loop {
if raw_entry_ptr == 0 {
if ptr == 0 {
break;
}
let raw_entry = unsafe { &*(raw_entry_ptr as *mut StringRawEntry) };
let raw_entry = unsafe { &*(ptr as *mut StringRawEntry) };
// Compare `early` and the length of the string, the size of `early` is 4.
let min_len = std::cmp::min(
STRING_EARLY_SIZE,
std::cmp::min(key_ref.len(), raw_entry.length as usize),
std::cmp::min(key.len(), raw_entry.length as usize),
);
if raw_entry.length as usize == key_ref.len()
&& key_ref[0..min_len] == raw_entry.early[0..min_len]
if raw_entry.length as usize == key.len()
&& key[0..min_len] == raw_entry.early[0..min_len]
{
let key = unsafe {
let key_ref = unsafe {
std::slice::from_raw_parts(
raw_entry.key as *const u8,
raw_entry.length as usize,
Expand All @@ -111,79 +116,31 @@ where A: Allocator + Clone + 'static
return true;
}
}
raw_entry_ptr = raw_entry.next;
ptr = raw_entry.next;
}
false
}

fn probe_hash_table(
&self,
key_ref: &Self::Key,
vec_ptr: *mut RowPtr,
mut occupied: usize,
capacity: usize,
) -> (usize, u64) {
let index = key_ref.fast_hash() as usize & self.hash_mask;
let origin = occupied;
let mut raw_entry_ptr = self.pointers[index];
loop {
if raw_entry_ptr == 0 || occupied >= capacity {
break;
}
let raw_entry = unsafe { &*(raw_entry_ptr as *mut StringRawEntry) };
// Compare `early` and the length of the string, the size of `early` is 4.
let min_len = std::cmp::min(STRING_EARLY_SIZE, key_ref.len());
if raw_entry.length as usize == key_ref.len()
&& key_ref[0..min_len] == raw_entry.early[0..min_len]
{
let key = unsafe {
std::slice::from_raw_parts(
raw_entry.key as *const u8,
raw_entry.length as usize,
)
};
if key == key_ref {
// # Safety
// occupied is less than the capacity of vec_ptr.
unsafe {
std::ptr::copy_nonoverlapping(
&raw_entry.row_ptr as *const RowPtr,
vec_ptr.add(occupied),
1,
)
};
occupied += 1;
}
}
raw_entry_ptr = raw_entry.next;
}
if occupied > origin {
(occupied - origin, raw_entry_ptr)
} else {
(0, 0)
}
}

fn next_incomplete_ptr(
fn next_probe(
&self,
key_ref: &Self::Key,
mut incomplete_ptr: u64,
key: &Self::Key,
mut ptr: u64,
vec_ptr: *mut RowPtr,
mut occupied: usize,
capacity: usize,
) -> (usize, u64) {
let origin = occupied;
loop {
if incomplete_ptr == 0 || occupied >= capacity {
if ptr == 0 || occupied >= capacity {
break;
}
let raw_entry = unsafe { &*(incomplete_ptr as *mut StringRawEntry) };
let raw_entry = unsafe { &*(ptr as *mut StringRawEntry) };
// Compare `early` and the length of the string, the size of `early` is 4.
let min_len = std::cmp::min(STRING_EARLY_SIZE, key_ref.len());
if raw_entry.length as usize == key_ref.len()
&& key_ref[0..min_len] == raw_entry.early[0..min_len]
let min_len = std::cmp::min(STRING_EARLY_SIZE, key.len());
if raw_entry.length as usize == key.len()
&& key[0..min_len] == raw_entry.early[0..min_len]
{
let key = unsafe {
let key_ref = unsafe {
std::slice::from_raw_parts(
raw_entry.key as *const u8,
raw_entry.length as usize,
Expand All @@ -202,10 +159,10 @@ where A: Allocator + Clone + 'static
occupied += 1;
}
}
incomplete_ptr = raw_entry.next;
ptr = raw_entry.next;
}
if occupied > origin {
(occupied - origin, incomplete_ptr)
(occupied - origin, ptr)
} else {
(0, 0)
}
Expand Down
17 changes: 6 additions & 11 deletions src/common/hashtable/src/traits.rs
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -436,20 +436,15 @@ pub trait HashtableLike {
pub trait HashJoinHashtableLike {
type Key: ?Sized;

fn contains(&self, key_ref: &Self::Key) -> bool;
// Using hashes to probe hash table and converting them in-place to pointers for memory reuse.
fn probe(&self, hashes: &mut [u64]);

fn probe_hash_table(
&self,
key_ref: &Self::Key,
vec_ptr: *mut RowPtr,
occupied: usize,
capacity: usize,
) -> (usize, u64);
fn next_contains(&self, key: &Self::Key, ptr: u64) -> bool;

fn next_incomplete_ptr(
fn next_probe(
&self,
key_ref: &Self::Key,
incomplete_ptr: u64,
key: &Self::Key,
ptr: u64,
vec_ptr: *mut RowPtr,
occupied: usize,
capacity: usize,
Expand Down
22 changes: 12 additions & 10 deletions src/query/expression/src/kernels/group_by.rs
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -49,18 +49,20 @@ impl DataBlock {
hash_key_types: &[DataType],
efficiently_memory: bool,
) -> Result<HashMethodKind> {
if hash_key_types.len() == 1 {
let typ = hash_key_types[0].clone();
if matches!(typ, DataType::String | DataType::Variant) {
return Ok(HashMethodKind::SingleString(
HashMethodSingleString::default(),
));
}
if hash_key_types.len() == 1
&& matches!(
hash_key_types[0],
DataType::String | DataType::Variant | DataType::Bitmap
)
{
return Ok(HashMethodKind::SingleString(
HashMethodSingleString::default(),
));
}

let mut group_key_len = 0;
for typ in hash_key_types {
let not_null_type = typ.remove_nullable();
for hash_key_type in hash_key_types {
let not_null_type = hash_key_type.remove_nullable();

if not_null_type.is_numeric()
|| not_null_type.is_date_or_date_time()
Expand All @@ -69,7 +71,7 @@ impl DataBlock {
group_key_len += not_null_type.numeric_byte_size().unwrap();

// extra one byte for null flag
if typ.is_nullable() {
if hash_key_type.is_nullable() {
group_key_len += 1;
}
} else {
Expand Down
Loading
Loading