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

Added early returns for "object is zeros" case.

Browse Source 
Details:
- Added some logic to packm_init(), pack_int() and gemm_int() so that
  (a) objects marked as BLIS_ZEROS are not packed, and (b) those
  objects are not computed with. This functionality is not currently
  needed by any existing implementations, but may be used in the
  future.
This commit is contained in:
Field G. Van Zee
2014-02-04 09:15:19 -06:00
parent 6bbd4be769
commit 5cf58f7c2d
3 changed files with 21 additions and 26 deletions
Download Patch File Download Diff File Expand all files Collapse all files

8
frame/1m/packm/bli_packm_init.c
View File

@@ -95,6 +95,14 @@ void bli_packm_init( obj_t* a,
return;
}
// If the object is marked as being filled with zeros, then we can skip
// the packm operation entirely and alias.
if ( bli_obj_is_zeros( *a ) )
{
bli_obj_alias_to( *a, *p );
return;
}
// Now, if we are not skipping the pack operation, then the only question
// left is whether we are to typecast matrix a before packing.
if ( bli_obj_datatype( *a ) != bli_obj_target_datatype( *a ) )

30
frame/1m/packm/bli_packm_int.c
View File

@@ -102,34 +102,12 @@ void bli_packm_int( obj_t* a,
return;
}
/*
// The value for kappa we use will depend on whether the scalar
// attached to A has a nonzero imaginary component. If it does,
// then we will apply the scalar during packing to facilitate
// implementing complex domain micro-kernels in terms of their
// 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( a ) )
// If the object is marked as being filled with zeros, then we can skip
// the packm operation entirely.
if ( bli_obj_is_zeros( *a ) )
{
bli_check_error_code( BLIS_NOT_YET_IMPLEMENTED );
// Detach the scalar.
bli_obj_scalar_detach( a, &kappa );
// Reset the attached scalar (to 1.0).
bli_obj_scalar_reset( a );
kappa_p = κ
return;
}
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;
}
*/
// Extract the variant number and implementation type.

9
frame/3/gemm/bli_gemm_int.c
View File

@@ -81,6 +81,15 @@ void bli_gemm_int( obj_t* alpha,
return;
}
// If A or B is marked as being filled with zeros, scale C by beta and
// return early.
if ( bli_obj_is_zeros( *a ) ||
bli_obj_is_zeros( *b ) )
{
bli_scalm( beta, c );
return;
}
// Alias A and B in case we need to update attached scalars.
bli_obj_alias_to( *a, a_local );
bli_obj_alias_to( *b, b_local );