amd/blis
1
0
Fork 0
mirror of https://github.com/amd/blis.git synced 2026-04-19 23:28:52 +00:00
Code Issues Packages Projects Releases Wiki Activity

thread: free global communicator after parallel region completes in p…

Browse Source 
* thread: free global communicator after parallel region completes in pthreads decorator

Avoid potential data race by deferring  free until all threads have joined. Previously, chief thread could free  inside  while non-chief threads still held pointers. Now,  frees  after the parallel region, following barrier and joins.

Files:
- frame/thread/bli_l3_sup_decor_pthreads.c
- frame/thread/bli_l3_decor_pthreads.c

* AMD-Internal: [CPUPL-7694]
This commit is contained in:
Varaganti, Kiran
2025-12-09 19:15:52 +05:30
committed by GitHub
parent eff1b561c5
commit bbb7edcb22
2 changed files with 28 additions and 0 deletions
Download Patch File Download Diff File Expand all files Collapse all files

14
frame/thread/bli_l3_decor_pthreads.c
View File

@@ -126,6 +126,12 @@ void* bli_l3_thread_entry( void* data_void )
// Free the thread's local control tree.
bli_l3_cntl_free( rntm_p, cntl_use, thread );
// NOTE: The barrier here is very important as it prevents memory being
// released by the chief of some thread sub-group before its peers are done
// using it. See PR #702 for more info [1].
// [1] https://github.com/flame/blis/pull/702
bli_thread_barrier( thread );
// Free the current thread's thrinfo_t structure.
bli_l3_thrinfo_free( rntm_p, thread );
@@ -239,6 +245,14 @@ void bli_l3_thread_decorator
// mutual exclusion.
bli_sba_checkin_array( array );
// Now global communicator is not freed in bli_l3_thrinfo_free().
// Free the global communicator after the parallel region completes.
// This ensures that no thread can be using gl_comm when it is freed,
// avoiding a potential data race where the chief thread would free
// gl_comm inside bli_thrinfo_free() while non-chief threads might
// still hold pointers to it.
bli_thrcomm_free( rntm, gl_comm );
#ifdef BLIS_ENABLE_MEM_TRACING
printf( "bli_l3_thread_decorator().pth: " );
#endif

14
frame/thread/bli_l3_sup_decor_pthreads.c
View File

@@ -103,6 +103,12 @@ void* bli_l3_sup_thread_entry( void* data_void )
thread
);
// NOTE: The barrier here is very important as it prevents memory being
// released by the chief of some thread sub-group before its peers are done
// using it. See PR #702 for more info [1].
// [1] https://github.com/flame/blis/pull/702
bli_thread_barrier( thread );
// Free the current thread's thrinfo_t structure.
bli_l3_sup_thrinfo_free( rntm_p, thread );
@@ -201,6 +207,14 @@ err_t bli_l3_sup_thread_decorator
// mutual exclusion.
bli_sba_checkin_array( array );
// Now global communicator is not freed in bli_l3_thrinfo_free().
// Free the global communicator after the parallel region completes.
// This ensures that no thread can be using gl_comm when it is freed,
// avoiding a potential data race where the chief thread would free
// gl_comm inside bli_thrinfo_free() while non-chief threads might
// still hold pointers to it.
bli_thrcomm_free( rntm, gl_comm );
#ifdef BLIS_ENABLE_MEM_TRACING
printf( "bli_l3_thread_decorator().pth: " );
#endif