void bli_trmm_cntl_init()
{
// Create control tree objects for packm operations (left side).
trmm_l_packa_cntl
=
bli_packm_cntl_obj_create( BLIS_BLOCKED,
BLIS_VARIANT1,
// IMPORTANT: Unlike trsm, trmm does not require a
// "k" dim multiple equal to mr.
gemm_mr,
gemm_kr,
FALSE, // do NOT invert diagonal
FALSE, // reverse iteration if upper?
FALSE, // reverse iteration if lower?
BLIS_PACKED_ROW_PANELS,
BLIS_BUFFER_FOR_A_BLOCK );
trmm_l_packb_cntl
=
bli_packm_cntl_obj_create( BLIS_BLOCKED,
BLIS_VARIANT1,
// IMPORTANT: Unlike trsm, trmm does not require a
// "k" dim multiple equal to mr.
gemm_kr,
gemm_nr,
FALSE, // do NOT invert diagonal
FALSE, // reverse iteration if upper?
FALSE, // reverse iteration if lower?
BLIS_PACKED_COL_PANELS,
BLIS_BUFFER_FOR_B_PANEL );
// Create control tree object for lowest-level block-panel kernel.
trmm_cntl_bp_ke
=
bli_trmm_cntl_obj_create( BLIS_UNB_OPT,
BLIS_VARIANT2,
NULL,
gemm_ukrs,
NULL, NULL, NULL, NULL,
NULL, NULL, NULL );
// Create control tree object for outer panel (to block-panel)
// problem (left side).
trmm_l_cntl_op_bp
=
bli_trmm_cntl_obj_create( BLIS_BLOCKED,
BLIS_VARIANT1,
gemm_mc,
gemm_ukrs,
NULL,
trmm_l_packa_cntl,
trmm_l_packb_cntl,
NULL,
trmm_cntl_bp_ke,
gemm_cntl_bp_ke,
NULL );
// Create control tree object for general problem via multiple
// rank-k (outer panel) updates (left side).
trmm_l_cntl_mm_op
=
bli_trmm_cntl_obj_create( BLIS_BLOCKED,
BLIS_VARIANT3,
gemm_kc,
gemm_ukrs,
NULL,
NULL,
NULL,
NULL,
trmm_l_cntl_op_bp,
NULL,
NULL );
// Create control tree object for very large problem via multiple
// general problems (left side).
trmm_l_cntl_vl_mm
=
bli_trmm_cntl_obj_create( BLIS_BLOCKED,
BLIS_VARIANT2,
gemm_nc,
gemm_ukrs,
NULL,
NULL,
NULL,
NULL,
trmm_l_cntl_mm_op,
NULL,
NULL );
// Alias the "master" trmm control trees to shorter names.
trmm_l_cntl = trmm_l_cntl_vl_mm;
}
void bli_trmm3m_cntl_init()
{
// Create control tree objects for packm operations (left side).
trmm3m_l_packa_cntl
=
bli_packm_cntl_obj_create( BLIS_BLOCKED,
BLIS_VARIANT3,
// IMPORTANT: for consistency with trsm, "k" dim
// multiple is set to mr.
gemm3m_mr,
gemm3m_mr,
TRUE, // densify
FALSE, // do NOT invert diagonal
FALSE, // reverse iteration if upper?
FALSE, // reverse iteration if lower?
BLIS_PACKED_ROW_PANELS_3M,
BLIS_BUFFER_FOR_A_BLOCK );
trmm3m_l_packb_cntl
=
bli_packm_cntl_obj_create( BLIS_BLOCKED,
BLIS_VARIANT3,
// IMPORTANT: m dim multiple here must be mr
// since "k" dim multiple is set to mr above.
gemm3m_mr,
gemm3m_nr,
FALSE, // already dense
FALSE, // do NOT invert diagonal
FALSE, // reverse iteration if upper?
FALSE, // reverse iteration if lower?
BLIS_PACKED_COL_PANELS_3M,
BLIS_BUFFER_FOR_B_PANEL );
// Create control tree objects for packm operations (right side).
trmm3m_r_packa_cntl
=
bli_packm_cntl_obj_create( BLIS_BLOCKED,
BLIS_VARIANT3,
// IMPORTANT: for consistency with trsm, "k" dim
// multiple is set to nr.
gemm3m_mr,
gemm3m_nr,
FALSE, // already dense
FALSE, // do NOT invert diagonal
FALSE, // reverse iteration if upper?
FALSE, // reverse iteration if lower?
BLIS_PACKED_ROW_PANELS_3M,
BLIS_BUFFER_FOR_A_BLOCK );
trmm3m_r_packb_cntl
=
bli_packm_cntl_obj_create( BLIS_BLOCKED,
BLIS_VARIANT3,
// IMPORTANT: m dim multiple here must be nr
// since "k" dim multiple is set to nr above.
gemm3m_nr,
gemm3m_nr,
TRUE, // densify
FALSE, // do NOT invert diagonal
FALSE, // reverse iteration if upper?
FALSE, // reverse iteration if lower?
BLIS_PACKED_COL_PANELS_3M,
BLIS_BUFFER_FOR_B_PANEL );
// Create control tree object for lowest-level block-panel kernel.
trmm3m_cntl_bp_ke
=
bli_trmm_cntl_obj_create( BLIS_UNB_OPT,
BLIS_VARIANT2,
NULL,
gemm3m_ukrs,
NULL, NULL, NULL, NULL,
NULL, NULL, NULL );
// Create control tree object for outer panel (to block-panel)
// problem (left side).
trmm3m_l_cntl_op_bp
=
bli_trmm_cntl_obj_create( BLIS_BLOCKED,
BLIS_VARIANT1,
gemm3m_mc,
NULL,
NULL,
trmm3m_l_packa_cntl,
trmm3m_l_packb_cntl,
NULL,
trmm3m_cntl_bp_ke,
gemm3m_cntl_bp_ke,
NULL );
// Create control tree object for general problem via multiple
// rank-k (outer panel) updates (left side).
trmm3m_l_cntl_mm_op
=
bli_trmm_cntl_obj_create( BLIS_BLOCKED,
BLIS_VARIANT3,
gemm3m_kc,
NULL,
NULL,
NULL,
NULL,
NULL,
trmm3m_l_cntl_op_bp,
NULL,
NULL );
// Create control tree object for very large problem via multiple
// general problems (left side).
trmm3m_l_cntl_vl_mm
=
bli_trmm_cntl_obj_create( BLIS_BLOCKED,
BLIS_VARIANT2,
gemm3m_nc,
NULL,
NULL,
NULL,
NULL,
NULL,
trmm3m_l_cntl_mm_op,
NULL,
NULL );
// Create control tree object for outer panel (to block-panel)
// problem (right side).
trmm3m_r_cntl_op_bp
=
bli_trmm_cntl_obj_create( BLIS_BLOCKED,
BLIS_VARIANT1,
gemm3m_mc,
NULL,
NULL,
trmm3m_r_packa_cntl,
trmm3m_r_packb_cntl,
NULL,
trmm3m_cntl_bp_ke,
gemm3m_cntl_bp_ke,
NULL );
// Create control tree object for general problem via multiple
// rank-k (outer panel) updates (right side).
trmm3m_r_cntl_mm_op
=
bli_trmm_cntl_obj_create( BLIS_BLOCKED,
BLIS_VARIANT3,
gemm3m_kc,
NULL,
NULL,
NULL,
NULL,
NULL,
trmm3m_r_cntl_op_bp,
NULL,
NULL );
// Create control tree object for very large problem via multiple
// general problems (right side).
trmm3m_r_cntl_vl_mm
=
bli_trmm_cntl_obj_create( BLIS_BLOCKED,
BLIS_VARIANT2,
gemm3m_nc,
NULL,
NULL,
NULL,
NULL,
NULL,
trmm3m_r_cntl_mm_op,
NULL,
NULL );
// Alias the "master" trmm control trees to shorter names.
trmm3m_l_cntl = trmm3m_l_cntl_vl_mm;
trmm3m_r_cntl = trmm3m_r_cntl_vl_mm;
}