Skip to content

Add optional all-dims-small (AND) rule for SUP threshold #944

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Draft
jeffhammond wants to merge 1 commit into
base: master
Choose a base branch
from
+12 −0
Draft

Add optional all-dims-small (AND) rule for SUP threshold #944

Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
12 changes: 12 additions & 0 deletions frame/base/bli_cntx.h
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -185,11 +185,23 @@ BLIS_INLINE bool bli_cntx_get_ukr_prefs_dt( num_t dt, kerid_t ukr_id, const cntx

BLIS_INLINE bool bli_cntx_l3_sup_thresh_is_met( num_t dt, dim_t m, dim_t n, dim_t k, const cntx_t* cntx )
{
#ifdef BLIS_SUP_THRESH_ALL
// Take the SUP (skinny/unpacked) path only when *every* dimension is below
// its threshold, i.e. the matrix is small in all three dims. Shapes that
// are skinny in one dimension but large in another (rank-k updates, tall
// panels) are then handled by the packed native path, which is markedly
// faster for them on GB10 (measured: 4000x4000x64 native 677 vs SUP 474
// GFLOP/s). The default OR-semantics below would misroute those to SUP.
return ( m < bli_cntx_get_blksz_def_dt( dt, BLIS_MT, cntx ) &&
n < bli_cntx_get_blksz_def_dt( dt, BLIS_NT, cntx ) &&
k < bli_cntx_get_blksz_def_dt( dt, BLIS_KT, cntx ) );
#else
if ( m < bli_cntx_get_blksz_def_dt( dt, BLIS_MT, cntx ) ) return TRUE;
if ( n < bli_cntx_get_blksz_def_dt( dt, BLIS_NT, cntx ) ) return TRUE;
if ( k < bli_cntx_get_blksz_def_dt( dt, BLIS_KT, cntx ) ) return TRUE;

return FALSE;
#endif
}

// -----------------------------------------------------------------------------
Expand Down