amd/blis
1
0
Fork 0
mirror of https://github.com/amd/blis.git synced 2026-05-11 09:39:59 +00:00
Code Issues Packages Projects Releases Wiki Activity

Fixed a type of race condition exposed by pthreads implementation.

Browse Source 
Lead thread of the inner thread communicator could exit subproblem, move on the next iteration of the loop and modify a1_pack, b1_pack, or c1_pack while other threads were still using those.

Barriers were inserted to fix this.
This commit is contained in:
Tyler Smith
2014-11-26 18:00:56 -06:00
parent 76bde44411
commit bef24e67e0
20 changed files with 109 additions and 37 deletions
Download Patch File Download Diff File Expand all files Collapse all files

2
frame/3/gemm/bli_gemm_blk_var1f.c
View File

@@ -130,6 +130,8 @@ void bli_gemm_blk_var1f( obj_t* a,
c1_pack,
cntl_sub_gemm( cntl ),
gemm_thread_sub_gemm( thread ) );
thread_ibarrier( thread );
// Unpack C1 (if C1 was packed).
// Currently must be done by 1 thread

2
frame/3/gemm/bli_gemm_blk_var2f.c
View File

@@ -129,6 +129,8 @@ void bli_gemm_blk_var2f( obj_t* a,
c1_pack,
cntl_sub_gemm( cntl ),
gemm_thread_sub_gemm( thread ) );
thread_ibarrier( thread );
// Unpack C1 (if C1 was packed).
// Currently must be done by 1 thread

2
frame/3/gemm/bli_gemm_blk_var3f.c
View File

@@ -131,7 +131,7 @@ void bli_gemm_blk_var3f( obj_t* a,
// And since c_pack is a local obj_t, we can simply overwrite the
// internal beta scalar with BLIS_ONE once it has been used in the
// first iteration.
if ( i == 0 ) thread_ibarrier( thread );
thread_ibarrier( thread );
if ( i == 0 && thread_am_ichief( thread ) ) bli_obj_scalar_reset( c_pack );
}

2
frame/3/herk/bli_herk_blk_var1f.c
View File

@@ -127,6 +127,8 @@ void bli_herk_blk_var1f( obj_t* a,
cntl_sub_gemm( cntl ),
herk_thread_sub_herk( thread ) );
thread_ibarrier( thread );
// Unpack C1 (if C1 was packed).
bli_unpackm_int( c1_pack, &c1,
cntl_sub_unpackm_c( cntl ),

2
frame/3/herk/bli_herk_blk_var2f.c
View File

@@ -142,6 +142,8 @@ void bli_herk_blk_var2f( obj_t* a,
cntl_sub_gemm( cntl ),
herk_thread_sub_herk( thread ) );
thread_ibarrier( thread );
// Unpack C1 (if C1 was packed).
bli_unpackm_int( c1S_pack, &c1S,
cntl_sub_unpackm_c( cntl ),

2
frame/3/herk/bli_herk_blk_var3f.c
View File

@@ -128,7 +128,7 @@ void bli_herk_blk_var3f( obj_t* a,
// And since c_pack is a local obj_t, we can simply overwrite the
// internal beta scalar with BLIS_ONE once it has been used in the
// first iteration.
if ( i == 0 ) thread_ibarrier( thread );
thread_ibarrier( thread );
if ( i == 0 && thread_am_ichief( thread ) ) bli_obj_scalar_reset( c_pack );
}

44
frame/3/herk/bli_herk_threading.c
View File

@@ -160,38 +160,56 @@ herk_thrinfo_t** bli_create_herk_thrinfo_paths( )
dim_t jc_comm_id = b*kc_nt + kc_comm_id;
dim_t global_comm_id = a*jc_nt + jc_comm_id;
// Macrokernel loops
herk_thrinfo_t* ir_info = bli_create_herk_thrinfo_node( jr_comm, jr_comm_id,
ir_comm, ir_comm_id,
ir_way, e,
ir_way, e,
NULL, NULL, NULL);
herk_thrinfo_t* jr_info = bli_create_herk_thrinfo_node( ic_comm, ic_comm_id,
jr_comm, jr_comm_id,
jr_way, d,
jr_way, d,
NULL, NULL, ir_info);
packm_thrinfo_t* packb = bli_create_packm_thread_info( kc_comm, kc_comm_id,
ic_comm, ic_comm_id,
kc_nt, kc_comm_id );
packm_thrinfo_t* packa = bli_create_packm_thread_info( ic_comm, ic_comm_id,
//blk_var_1
packm_thrinfo_t* pack_ic_in = bli_create_packm_thread_info( ic_comm, ic_comm_id,
jr_comm, jr_comm_id,
ic_nt, ic_comm_id );
packm_thrinfo_t* pack_ic_out = bli_create_packm_thread_info( kc_comm, kc_comm_id,
ic_comm, ic_comm_id,
kc_nt, kc_comm_id );
herk_thrinfo_t* ic_info = bli_create_herk_thrinfo_node( kc_comm, kc_comm_id,
ic_comm, ic_comm_id,
ic_way, c,
packb, packa, jr_info);
ic_way, c,
pack_ic_out, pack_ic_in, jr_info);
//blk_var_3
packm_thrinfo_t* pack_kc_in = bli_create_packm_thread_info( kc_comm, kc_comm_id,
ic_comm, ic_comm_id,
kc_nt, kc_comm_id );
packm_thrinfo_t* pack_kc_out = bli_create_packm_thread_info( jc_comm, jc_comm_id,
jc_comm, jc_comm_id,
jc_nt, jc_comm_id );
herk_thrinfo_t* kc_info = bli_create_herk_thrinfo_node( jc_comm, jc_comm_id,
kc_comm, kc_comm_id,
kc_way, b,
NULL, NULL, ic_info);
pack_kc_out, pack_kc_in, ic_info);
//blk_var_2
packm_thrinfo_t* pack_jc_in = bli_create_packm_thread_info( jc_comm, jc_comm_id,
kc_comm, kc_comm_id,
jc_nt, jc_comm_id );
packm_thrinfo_t* pack_jc_out = bli_create_packm_thread_info( global_comm, global_comm_id,
jc_comm, jc_comm_id,
global_num_threads, global_comm_id );
herk_thrinfo_t* jc_info = bli_create_herk_thrinfo_node( global_comm, global_comm_id,
jc_comm, jc_comm_id,
jc_way, a,
NULL, NULL, kc_info);
jc_way, a,
pack_jc_out, pack_jc_in, kc_info);
paths[global_comm_id] = jc_info;
}

1
frame/3/trmm/bli_trmm_blk_var1f.c
View File

@@ -138,6 +138,7 @@ void bli_trmm_blk_var1f( obj_t* a,
c1_pack,
cntl_sub_gemm( cntl ),
trmm_thread_sub_trmm( thread ) );
thread_ibarrier( thread );
// Unpack C1 (if C1 was packed).
bli_unpackm_int( c1_pack, &c1,

1
frame/3/trmm/bli_trmm_blk_var2b.c
View File

@@ -126,6 +126,7 @@ void bli_trmm_blk_var2b( obj_t* a,
c1_pack,
cntl_sub_gemm( cntl ),
trmm_thread_sub_trmm( thread ) );
thread_ibarrier( thread );
// Unpack C1 (if C1 was packed).
bli_unpackm_int( c1_pack, &c1,

1
frame/3/trmm/bli_trmm_blk_var2f.c
View File

@@ -126,6 +126,7 @@ void bli_trmm_blk_var2f( obj_t* a,
c1_pack,
cntl_sub_gemm( cntl ),
trmm_thread_sub_trmm( thread ) );
thread_ibarrier( thread );
// Unpack C1 (if C1 was packed).
bli_unpackm_int( c1_pack, &c1,

1
frame/3/trmm/bli_trmm_blk_var3b.c
View File

@@ -124,6 +124,7 @@ void bli_trmm_blk_var3b( obj_t* a,
c_pack,
cntl_sub_gemm( cntl ),
trmm_thread_sub_trmm( thread ) );
thread_ibarrier( thread );
}
thread_obarrier( thread );

1
frame/3/trmm/bli_trmm_blk_var3f.c
View File

@@ -124,6 +124,7 @@ void bli_trmm_blk_var3f( obj_t* a,
c_pack,
cntl_sub_gemm( cntl ),
trmm_thread_sub_trmm( thread ) );
thread_ibarrier( thread );
}
thread_obarrier( thread );

38
frame/3/trmm/bli_trmm_threading.c
View File

@@ -164,7 +164,8 @@ trmm_thrinfo_t** bli_create_trmm_thrinfo_paths( bool_t jc_dependency )
dim_t kc_comm_id = c*ic_nt + ic_comm_id;
dim_t jc_comm_id = b*kc_nt + kc_comm_id;
dim_t global_comm_id = a*jc_nt + jc_comm_id;
// Macrokernel loops
trmm_thrinfo_t* ir_info = bli_create_trmm_thrinfo_node( jr_comm, jr_comm_id,
ir_comm, ir_comm_id,
ir_way, e,
@@ -174,29 +175,46 @@ trmm_thrinfo_t** bli_create_trmm_thrinfo_paths( bool_t jc_dependency )
jr_comm, jr_comm_id,
jr_way, d,
NULL, NULL, ir_info);
packm_thrinfo_t* packb = bli_create_packm_thread_info( kc_comm, kc_comm_id,
ic_comm, ic_comm_id,
kc_nt, kc_comm_id );
packm_thrinfo_t* packa = bli_create_packm_thread_info( ic_comm, ic_comm_id,
//blk_var_1
packm_thrinfo_t* pack_ic_in = bli_create_packm_thread_info( ic_comm, ic_comm_id,
jr_comm, jr_comm_id,
ic_nt, ic_comm_id );
packm_thrinfo_t* pack_ic_out = bli_create_packm_thread_info( kc_comm, kc_comm_id,
ic_comm, ic_comm_id,
kc_nt, kc_comm_id );
trmm_thrinfo_t* ic_info = bli_create_trmm_thrinfo_node( kc_comm, kc_comm_id,
ic_comm, ic_comm_id,
ic_way, c,
packb, packa, jr_info);
pack_ic_out, pack_ic_in, jr_info);
//blk_var_3
packm_thrinfo_t* pack_kc_in = bli_create_packm_thread_info( kc_comm, kc_comm_id,
ic_comm, ic_comm_id,
kc_nt, kc_comm_id );
packm_thrinfo_t* pack_kc_out = bli_create_packm_thread_info( jc_comm, jc_comm_id,
jc_comm, jc_comm_id,
jc_nt, jc_comm_id );
trmm_thrinfo_t* kc_info = bli_create_trmm_thrinfo_node( jc_comm, jc_comm_id,
kc_comm, kc_comm_id,
kc_way, b,
NULL, NULL, ic_info);
pack_kc_out, pack_kc_in, ic_info);
//blk_var_2
packm_thrinfo_t* pack_jc_in = bli_create_packm_thread_info( jc_comm, jc_comm_id,
kc_comm, kc_comm_id,
jc_nt, jc_comm_id );
packm_thrinfo_t* pack_jc_out = bli_create_packm_thread_info( global_comm, global_comm_id,
jc_comm, jc_comm_id,
global_num_threads, global_comm_id );
trmm_thrinfo_t* jc_info = bli_create_trmm_thrinfo_node( global_comm, global_comm_id,
jc_comm, jc_comm_id,
jc_way, a,
NULL, NULL, kc_info);
pack_jc_out, pack_jc_in, kc_info);
paths[global_comm_id] = jc_info;
}
}

1
frame/3/trsm/bli_trsm_blk_var1b.c
View File

@@ -121,6 +121,7 @@ void bli_trsm_blk_var1b( obj_t* a,
&c1,
cntl_sub_trsm( cntl ),
trsm_thread_sub_trsm( thread ) );
thread_ibarrier( thread );
}
// If any packing buffers were acquired within packm, release them back

1
frame/3/trsm/bli_trsm_blk_var1f.c
View File

@@ -120,6 +120,7 @@ void bli_trsm_blk_var1f( obj_t* a,
&c1,
cntl_sub_trsm( cntl ),
trsm_thread_sub_trsm( thread ) );
thread_ibarrier( thread );
}
// If any packing buffers were acquired within packm, release them back

1
frame/3/trsm/bli_trsm_blk_var2b.c
View File

@@ -128,6 +128,7 @@ void bli_trsm_blk_var2b( obj_t* a,
c1_pack,
cntl_sub_trsm( cntl ),
trsm_thread_sub_trsm( thread ) );
thread_ibarrier( thread );
// Unpack C1 (if C1 was packed).
bli_unpackm_int( c1_pack, &c1,

1
frame/3/trsm/bli_trsm_blk_var2f.c
View File

@@ -128,6 +128,7 @@ void bli_trsm_blk_var2f( obj_t* a,
c1_pack,
cntl_sub_trsm( cntl ),
trsm_thread_sub_trsm( thread ) );
thread_ibarrier( thread );
// Unpack C1 (if C1 was packed).
bli_unpackm_int( c1_pack, &c1,

2
frame/3/trsm/bli_trsm_blk_var3b.c
View File

@@ -129,7 +129,7 @@ void bli_trsm_blk_var3b( obj_t* a,
// This variant executes multiple rank-k updates. Therefore, if the
// internal alpha scalars on A/B and C are non-zero, we must ensure
// that they are only used in the first iteration.
if ( i == 0 ) thread_ibarrier( thread );
thread_ibarrier( thread );
if ( i == 0 && thread_am_ichief( thread ) ) {
bli_obj_scalar_reset( a );
bli_obj_scalar_reset( b );

2
frame/3/trsm/bli_trsm_blk_var3f.c
View File

@@ -129,7 +129,7 @@ void bli_trsm_blk_var3f( obj_t* a,
// This variant executes multiple rank-k updates. Therefore, if the
// internal alpha scalars on A/B and C are non-zero, we must ensure
// that they are only used in the first iteration.
if ( i == 0 ) thread_ibarrier( thread );
thread_ibarrier( thread );
if ( i == 0 && thread_am_ichief( thread ) ) {
bli_obj_scalar_reset( a );
bli_obj_scalar_reset( b );

39
frame/3/trsm/bli_trsm_threading.c
View File

@@ -167,6 +167,8 @@ trsm_thrinfo_t** bli_create_trsm_thrinfo_paths( bool_t right_sided )
dim_t jc_comm_id = b*kc_nt + kc_comm_id;
dim_t global_comm_id = a*jc_nt + jc_comm_id;
// Macrokernel loops
trsm_thrinfo_t* ir_info = bli_create_trsm_thrinfo_node( jr_comm, jr_comm_id,
ir_comm, ir_comm_id,
ir_way, e,
@@ -176,29 +178,46 @@ trsm_thrinfo_t** bli_create_trsm_thrinfo_paths( bool_t right_sided )
jr_comm, jr_comm_id,
jr_way, d,
NULL, NULL, ir_info);
//blk_var_1
packm_thrinfo_t* pack_ic_in = bli_create_packm_thread_info( ic_comm, ic_comm_id,
jr_comm, jr_comm_id,
ic_nt, ic_comm_id );
packm_thrinfo_t* packb = bli_create_packm_thread_info( kc_comm, kc_comm_id,
ic_comm, ic_comm_id,
kc_nt, kc_comm_id );
packm_thrinfo_t* packa = bli_create_packm_thread_info( ic_comm, ic_comm_id,
jr_comm, jr_comm_id,
ic_nt, ic_comm_id );
packm_thrinfo_t* pack_ic_out = bli_create_packm_thread_info( kc_comm, kc_comm_id,
ic_comm, ic_comm_id,
kc_nt, kc_comm_id );
trsm_thrinfo_t* ic_info = bli_create_trsm_thrinfo_node( kc_comm, kc_comm_id,
ic_comm, ic_comm_id,
ic_way, c,
packb, packa, jr_info);
pack_ic_out, pack_ic_in, jr_info);
//blk_var_3
packm_thrinfo_t* pack_kc_in = bli_create_packm_thread_info( kc_comm, kc_comm_id,
ic_comm, ic_comm_id,
kc_nt, kc_comm_id );
packm_thrinfo_t* pack_kc_out = bli_create_packm_thread_info( jc_comm, jc_comm_id,
jc_comm, jc_comm_id,
jc_nt, jc_comm_id );
trsm_thrinfo_t* kc_info = bli_create_trsm_thrinfo_node( jc_comm, jc_comm_id,
kc_comm, kc_comm_id,
kc_way, b,
NULL, NULL, ic_info);
pack_kc_out, pack_kc_in, ic_info);
//blk_var_2
packm_thrinfo_t* pack_jc_in = bli_create_packm_thread_info( jc_comm, jc_comm_id,
kc_comm, kc_comm_id,
jc_nt, jc_comm_id );
packm_thrinfo_t* pack_jc_out = bli_create_packm_thread_info( global_comm, global_comm_id,
jc_comm, jc_comm_id,
global_num_threads, global_comm_id );
trsm_thrinfo_t* jc_info = bli_create_trsm_thrinfo_node( global_comm, global_comm_id,
jc_comm, jc_comm_id,
jc_way, a,
NULL, NULL, kc_info);
pack_jc_out, pack_jc_in, kc_info);
paths[global_comm_id] = jc_info;
}
}