Skip to content

MDEV-40827: prefetch MHNSW neighbours during search #5571

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

Open
chanztuying wants to merge 2 commits into
base: main
Choose a base branch
from
+33 −0
Open

MDEV-40827: prefetch MHNSW neighbours during search #5571

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
8c0ba52
MDEV-38721: prefetch MHNSW neighbours during search
chanztuying Aug 18, 2026
7868523
MDEV-38721: make the MHNSW prefetch portable to MSVC
chanztuying Aug 21, 2026
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
19 changes: 19 additions & 0 deletions include/my_cpu.h
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -110,6 +110,25 @@ static inline void MY_RELAX_CPU(void)
}


/*
Hint the CPU to start loading a cache line for reading.

Like YieldProcessor() above, PreFetchCacheLine() is architecture-independent:
winnt.h picks the instruction per target, unlike _mm_prefetch().
clang-cl defines __clang__ but not __GNUC__, so it must take the first branch.
*/
static inline void my_prefetch(const void *addr)
{
#if defined(__GNUC__) || defined(__clang__)
__builtin_prefetch(addr, 0, 3);
#elif defined(_WIN32)
PreFetchCacheLine(PF_TEMPORAL_LEVEL_1, addr);
#else
(void) addr;
#endif
}


#ifdef HAVE_PAUSE_INSTRUCTION
# ifdef __cplusplus
extern "C" {
Expand Down
14 changes: 14 additions & 0 deletions sql/vector_mhnsw.cc
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -22,6 +22,7 @@
#include "vector_mhnsw.h"
#include <scope.h>
#include <my_atomic_wrapper.h>
#include <my_cpu.h> // my_prefetch()
#include "bloom_filters.h"

// distance can be a little bit < 0 because of fast math
Expand Down Expand Up @@ -1347,6 +1348,19 @@ static int search_layer(MHNSW_param *p, const FVector *target, float threshold,
if (res == 0xff)
continue;

// A node and its vector share one allocation. Prefetch unseen nodes before
// computing distances so later loads can overlap with work on earlier ones.
for (size_t i= 0; i < 8; i++)
{
FVectorNode *link= links[i];
if (!(res & (1 << i)) && link)
{
my_prefetch(link);
my_prefetch(reinterpret_cast<const char*>(link)
+ CPU_LEVEL1_DCACHE_LINESIZE);
}
}

for (size_t i= 0; i < 8; i++)
{
if (res & (1 << i))
Expand Down