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

Fixed packm variants 3 and 4 where every thread was trying to manipulate the same state

Browse Source 
Now just performed by the master thread.
This commit is contained in:
Tyler Smith
2014-03-19 11:21:16 -05:00
parent fb42983bd9
commit c0140cb752
2 changed files with 38 additions and 32 deletions
Download Patch File Download Diff File Expand all files Collapse all files

35
frame/1m/packm/bli_packm_blk_var3.c
View File

@@ -100,7 +100,7 @@ void bli_packm_blk_var3( obj_t* c,
// in the real domain.
if ( bli_is_real( dt_cp ) )
{
bli_packm_blk_var1( c, p, &BLIS_PACKM_SINGLE_THREADED );
bli_packm_blk_var1( c, p, t );
return;
}
@@ -111,23 +111,26 @@ void bli_packm_blk_var3( obj_t* c,
// real domain counterparts. (In the aforementioned situation,
// applying a real scalar is easy, but applying a complex one is
// harder, so we avoid the need altogether with the code below.)
if ( bli_obj_scalar_has_nonzero_imag( p ) )
{
// Detach the scalar.
bli_obj_scalar_detach( p, &kappa );
if ( thread_am_ochief( t ) ) {
if ( bli_obj_scalar_has_nonzero_imag( p ) )
{
// Detach the scalar.
bli_obj_scalar_detach( p, &kappa );
// Reset the attached scalar (to 1.0).
bli_obj_scalar_reset( p );
// Reset the attached scalar (to 1.0).
bli_obj_scalar_reset( p );
kappa_p = κ
}
else
{
// If the internal scalar of A has only a real component, then
// we will apply it later (in the micro-kernel), and so we will
// use BLIS_ONE to indicate no scaling during packing.
kappa_p = &BLIS_ONE;
}
kappa_p = κ
}
else
{
// If the internal scalar of A has only a real component, then
// we will apply it later (in the micro-kernel), and so we will
// use BLIS_ONE to indicate no scaling during packing.
kappa_p = &BLIS_ONE;
}
}
kappa_p = thread_obroadcast( t, kappa_p );
// Acquire the buffer to the kappa chosen above.

35
frame/1m/packm/bli_packm_blk_var4.c
View File

@@ -100,7 +100,7 @@ void bli_packm_blk_var4( obj_t* c,
// in the real domain.
if ( bli_is_real( dt_cp ) )
{
bli_packm_blk_var1( c, p, &BLIS_PACKM_SINGLE_THREADED );
bli_packm_blk_var1( c, p, t );
return;
}
@@ -111,23 +111,26 @@ void bli_packm_blk_var4( obj_t* c,
// real domain counterparts. (In the aforementioned situation,
// applying a real scalar is easy, but applying a complex one is
// harder, so we avoid the need altogether with the code below.)
if ( bli_obj_scalar_has_nonzero_imag( p ) )
{
// Detach the scalar.
bli_obj_scalar_detach( p, &kappa );
if( thread_am_ochief( t ) ) {
if ( bli_obj_scalar_has_nonzero_imag( p ) )
{
// Detach the scalar.
bli_obj_scalar_detach( p, &kappa );
// Reset the attached scalar (to 1.0).
bli_obj_scalar_reset( p );
// Reset the attached scalar (to 1.0).
bli_obj_scalar_reset( p );
kappa_p = κ
}
else
{
// If the internal scalar of A has only a real component, then
// we will apply it later (in the micro-kernel), and so we will
// use BLIS_ONE to indicate no scaling during packing.
kappa_p = &BLIS_ONE;
}
kappa_p = κ
}
else
{
// If the internal scalar of A has only a real component, then
// we will apply it later (in the micro-kernel), and so we will
// use BLIS_ONE to indicate no scaling during packing.
kappa_p = &BLIS_ONE;
}
}
kappa_p = thread_obroadcast( t, kappa_p );
// Acquire the buffer to the kappa chosen above.