amd/blis
1
0
Fork 0
mirror of https://github.com/amd/blis.git synced 2026-05-11 17:50:00 +00:00
Code Issues Packages Projects Releases Wiki Activity

Switched to using environment variables to control threading.

Browse Source 
The environment variables all follow the format BLIS_X_NT,
where X is the index of the loop as described in our paper
Anatomy of High Performance Many-Threaded Matrix Multiplication.
These indices are IR, JR, IC, KC, and JC.

Also enabled parallelism for hemm and symm, but these are currently untested.
This commit is contained in:
Tyler Smith
2014-03-17 11:39:32 -05:00
parent 92233cf642
commit c720b14156
6 changed files with 57 additions and 26 deletions
Download Patch File Download Diff File Expand all files Collapse all files

4
frame/3/gemm/bli_gemm_cntl.c
View File

@@ -55,11 +55,9 @@ gemm_t* gemm_cntl_vl_mm;
gemm_t* gemm_cntl;
dim_t gemm_caucuses_at_level[5] = {1, 1, 2, 1, 1};
gemm_thrinfo_t* bli_gemm_cntl_get_thrinfos()
{
return bli_create_gemm_thrinfo_paths( gemm_caucuses_at_level, 5 );
return bli_create_gemm_thrinfo_paths( );
}
void bli_gemm_cntl_free_thrinfos(thrinfo_t* tofree)

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

@@ -90,5 +90,7 @@ void bli_gemm_front( obj_t* alpha,
cntl,
&infos[omp_id] );
}
bli_gemm_cntl_free_thrinfos( infos );
}

23
frame/3/gemm/bli_gemm_threading.c
View File

@@ -84,15 +84,24 @@ gemm_thrinfo_t* bli_create_gemm_thrinfo_node( thread_comm_t* ocomm, dim_t ocomm_
return thread;
}
gemm_thrinfo_t* bli_create_gemm_thrinfo_paths( dim_t* threads_at_level, dim_t n_levels )
dim_t read_env( char* env )
{
assert(n_levels == 5);
dim_t number = 1;
char* str = getenv( env );
if( str != NULL )
{
number = strtol( str, NULL, 10 );
}
return number;
}
dim_t jc_way = threads_at_level[0];
dim_t kc_way = threads_at_level[1];
dim_t ic_way = threads_at_level[2];
dim_t jr_way = threads_at_level[3];
dim_t ir_way = threads_at_level[4];
gemm_thrinfo_t* bli_create_gemm_thrinfo_paths( )
{
dim_t jc_way = read_env( "BLIS_JC_NT" );
dim_t kc_way = read_env( "BLIS_KC_NT" );
dim_t ic_way = read_env( "BLIS_IC_NT" );
dim_t jr_way = read_env( "BLIS_JR_NT" );
dim_t ir_way = read_env( "BLIS_IR_NT" );
dim_t global_num_threads = jc_way * kc_way * ic_way * jr_way * ir_way;
assert( global_num_threads != 0 );

2
frame/3/gemm/bli_gemm_threading.h
View File

@@ -53,7 +53,7 @@ typedef struct gemm_thrinfo_s gemm_thrinfo_t;
#define gemm_thread_sub_opackm( thread ) thread->opackm
#define gemm_thread_sub_ipackm( thread ) thread->ipackm
gemm_thrinfo_t* bli_create_gemm_thrinfo_paths( dim_t* threads_at_level, dim_t n_levels );
gemm_thrinfo_t* bli_create_gemm_thrinfo_paths( );
void bli_setup_gemm_thrinfo_node( gemm_thrinfo_t* thread,
thread_comm_t* ocomm, dim_t ocomm_id,

27
frame/3/hemm/bli_hemm_front.c
View File

@@ -80,13 +80,24 @@ void bli_hemm_front( side_t side,
bli_obj_swap( a_local, b_local );
}
// Invoke the internal back-end.
bli_gemm_int( alpha,
&a_local,
&b_local,
beta,
&c_local,
cntl,
&BLIS_GEMM_SINGLE_THREADED );
gemm_thrinfo_t* infos = bli_gemm_cntl_get_thrinfos();
dim_t n_threads = thread_num_threads( (&infos[0]) );
// Invoke the internal back-end.
_Pragma( "omp parallel num_threads(n_threads)" )
{
dim_t omp_id = omp_get_thread_num();
// Invoke the internal back-end.
bli_gemm_int( alpha,
&a_local,
&b_local,
beta,
&c_local,
cntl,
&infos[omp_id] );
}
bli_gemm_cntl_free_thrinfos( infos );
}

25
frame/3/symm/bli_symm_front.c
View File

@@ -79,13 +79,24 @@ void bli_symm_front( side_t side,
bli_obj_swap( a_local, b_local );
}
gemm_thrinfo_t* infos = bli_gemm_cntl_get_thrinfos();
dim_t n_threads = thread_num_threads( (&infos[0]) );
// Invoke the internal back-end.
bli_gemm_int( alpha,
&a_local,
&b_local,
beta,
&c_local,
cntl,
&BLIS_GEMM_SINGLE_THREADED );
_Pragma( "omp parallel num_threads(n_threads)" )
{
dim_t omp_id = omp_get_thread_num();
bli_gemm_int( alpha,
&a_local,
&b_local,
beta,
&c_local,
cntl,
&infos[omp_id] );
}
bli_gemm_cntl_free_thrinfos( infos );
}