void bli_cntx_3mh_stage( dim_t stage, cntx_t* cntx )
{
// Set the pack_t schemas as a function of the stage of execution.
if ( stage == 0 )
{
bli_cntx_set_schema_a_block( BLIS_PACKED_ROW_PANELS_RO, cntx );
bli_cntx_set_schema_b_panel( BLIS_PACKED_COL_PANELS_RO, cntx );
}
else if ( stage == 1 )
{
bli_cntx_set_schema_a_block( BLIS_PACKED_ROW_PANELS_IO, cntx );
bli_cntx_set_schema_b_panel( BLIS_PACKED_COL_PANELS_IO, cntx );
}
else // if ( stage == 2 )
{
bli_cntx_set_schema_a_block( BLIS_PACKED_ROW_PANELS_RPI, cntx );
bli_cntx_set_schema_b_panel( BLIS_PACKED_COL_PANELS_RPI, cntx );
}
}
void GENBAINAME(cntx_init)
(
ind_t method,
num_t dt,
cntx_t* cntx
)
{
func_t* funcs;
dim_t i;
// This function is designed to modify a copy of an existing native
// context to enable computation via an induced method for complex
// domain level-3 operations. It is called by bli_gks_query_ind_cntx()
// on a context after its contexts are set by copying from the
// architecture's native context.
// -- Set induced method level-3 virtual micro-kernels ---------------------
funcs = bli_cntx_l3_vir_ukrs_buf( cntx );
// 3mh, 4mh, and 4mb do not not support trsm.
bli_func_init_null( &funcs[ BLIS_GEMMTRSM_L_UKR ] );
bli_func_init_null( &funcs[ BLIS_GEMMTRSM_U_UKR ] );
bli_func_init_null( &funcs[ BLIS_TRSM_L_UKR ] );
bli_func_init_null( &funcs[ BLIS_TRSM_U_UKR ] );
if ( method == BLIS_3MH )
{
gen_func_init_co( &funcs[ BLIS_GEMM_UKR ], gemm3mh_ukr_name );
}
else if ( method == BLIS_3M1 )
{
gen_func_init_co( &funcs[ BLIS_GEMM_UKR ], gemm3m1_ukr_name );
gen_func_init_co( &funcs[ BLIS_GEMMTRSM_L_UKR ], gemmtrsm3m1_l_ukr_name );
gen_func_init_co( &funcs[ BLIS_GEMMTRSM_U_UKR ], gemmtrsm3m1_u_ukr_name );
gen_func_init_co( &funcs[ BLIS_TRSM_L_UKR ], trsm3m1_l_ukr_name );
gen_func_init_co( &funcs[ BLIS_TRSM_U_UKR ], trsm3m1_u_ukr_name );
}
else if ( method == BLIS_4MH )
{
gen_func_init_co( &funcs[ BLIS_GEMM_UKR ], gemm4mh_ukr_name );
}
else if ( method == BLIS_4M1B )
{
gen_func_init_co( &funcs[ BLIS_GEMM_UKR ], gemm4mb_ukr_name );
}
else if ( method == BLIS_4M1A )
{
gen_func_init_co( &funcs[ BLIS_GEMM_UKR ], gemm4m1_ukr_name );
gen_func_init_co( &funcs[ BLIS_GEMMTRSM_L_UKR ], gemmtrsm4m1_l_ukr_name );
gen_func_init_co( &funcs[ BLIS_GEMMTRSM_U_UKR ], gemmtrsm4m1_u_ukr_name );
gen_func_init_co( &funcs[ BLIS_TRSM_L_UKR ], trsm4m1_l_ukr_name );
gen_func_init_co( &funcs[ BLIS_TRSM_U_UKR ], trsm4m1_u_ukr_name );
}
else if ( method == BLIS_1M )
{
gen_func_init_co( &funcs[ BLIS_GEMM_UKR ], gemm1m_ukr_name );
gen_func_init_co( &funcs[ BLIS_GEMMTRSM_L_UKR ], gemmtrsm1m_l_ukr_name );
gen_func_init_co( &funcs[ BLIS_GEMMTRSM_U_UKR ], gemmtrsm1m_u_ukr_name );
gen_func_init_co( &funcs[ BLIS_TRSM_L_UKR ], trsm1m_l_ukr_name );
gen_func_init_co( &funcs[ BLIS_TRSM_U_UKR ], trsm1m_u_ukr_name );
}
else // if ( method == BLIS_NAT )
{
gen_func_init_co( &funcs[ BLIS_GEMM_UKR ], gemm_ukr_name );
gen_func_init_co( &funcs[ BLIS_GEMMTRSM_L_UKR ], gemmtrsm_l_ukr_name );
gen_func_init_co( &funcs[ BLIS_GEMMTRSM_U_UKR ], gemmtrsm_u_ukr_name );
gen_func_init_co( &funcs[ BLIS_TRSM_L_UKR ], trsm_l_ukr_name );
gen_func_init_co( &funcs[ BLIS_TRSM_U_UKR ], trsm_u_ukr_name );
}
// For 1m, we employ an optimization which requires that we copy the native
// real domain gemm ukernel function pointers to the corresponding slots in
// the virtual gemm ukernel func_t.
if ( method == BLIS_1M )
{
func_t* gemm_nat_ukrs = bli_cntx_get_l3_nat_ukrs( BLIS_GEMM_UKR, cntx );
func_t* gemm_vir_ukrs = bli_cntx_get_l3_vir_ukrs( BLIS_GEMM_UKR, cntx );
bli_func_copy_dt( BLIS_FLOAT, gemm_nat_ukrs, BLIS_FLOAT, gemm_vir_ukrs );
bli_func_copy_dt( BLIS_DOUBLE, gemm_nat_ukrs, BLIS_DOUBLE, gemm_vir_ukrs );
}
// -- Set induced method packm kernels -------------------------------------
funcs = bli_cntx_packm_kers_buf( cntx );
// Initialize all packm kernel func_t entries to NULL.
for ( i = BLIS_PACKM_0XK_KER; i <= BLIS_PACKM_31XK_KER; ++i )
{
bli_func_init_null( &funcs[ i ] );
}
if ( method == BLIS_3MH || method == BLIS_4MH )
{
gen_func_init_co( &funcs[ BLIS_PACKM_2XK_KER ], packm_2xk_rih_ker_name );
gen_func_init_co( &funcs[ BLIS_PACKM_4XK_KER ], packm_4xk_rih_ker_name );
gen_func_init_co( &funcs[ BLIS_PACKM_6XK_KER ], packm_6xk_rih_ker_name );
gen_func_init_co( &funcs[ BLIS_PACKM_8XK_KER ], packm_8xk_rih_ker_name );
gen_func_init_co( &funcs[ BLIS_PACKM_10XK_KER ], packm_10xk_rih_ker_name );
gen_func_init_co( &funcs[ BLIS_PACKM_12XK_KER ], packm_12xk_rih_ker_name );
gen_func_init_co( &funcs[ BLIS_PACKM_14XK_KER ], packm_14xk_rih_ker_name );
gen_func_init_co( &funcs[ BLIS_PACKM_16XK_KER ], packm_16xk_rih_ker_name );
gen_func_init_co( &funcs[ BLIS_PACKM_30XK_KER ], packm_30xk_rih_ker_name );
}
else if ( method == BLIS_3M1 )
{
gen_func_init_co( &funcs[ BLIS_PACKM_2XK_KER ], packm_2xk_3mis_ker_name );
gen_func_init_co( &funcs[ BLIS_PACKM_4XK_KER ], packm_4xk_3mis_ker_name );
gen_func_init_co( &funcs[ BLIS_PACKM_6XK_KER ], packm_6xk_3mis_ker_name );
gen_func_init_co( &funcs[ BLIS_PACKM_8XK_KER ], packm_8xk_3mis_ker_name );
gen_func_init_co( &funcs[ BLIS_PACKM_10XK_KER ], packm_10xk_3mis_ker_name );
gen_func_init_co( &funcs[ BLIS_PACKM_12XK_KER ], packm_12xk_3mis_ker_name );
gen_func_init_co( &funcs[ BLIS_PACKM_14XK_KER ], packm_14xk_3mis_ker_name );
gen_func_init_co( &funcs[ BLIS_PACKM_16XK_KER ], packm_16xk_3mis_ker_name );
gen_func_init_co( &funcs[ BLIS_PACKM_30XK_KER ], packm_30xk_3mis_ker_name );
}
else if ( method == BLIS_4M1A || method == BLIS_4M1B )
{
gen_func_init_co( &funcs[ BLIS_PACKM_2XK_KER ], packm_2xk_4mi_ker_name );
gen_func_init_co( &funcs[ BLIS_PACKM_4XK_KER ], packm_4xk_4mi_ker_name );
gen_func_init_co( &funcs[ BLIS_PACKM_6XK_KER ], packm_6xk_4mi_ker_name );
gen_func_init_co( &funcs[ BLIS_PACKM_8XK_KER ], packm_8xk_4mi_ker_name );
gen_func_init_co( &funcs[ BLIS_PACKM_10XK_KER ], packm_10xk_4mi_ker_name );
gen_func_init_co( &funcs[ BLIS_PACKM_12XK_KER ], packm_12xk_4mi_ker_name );
gen_func_init_co( &funcs[ BLIS_PACKM_14XK_KER ], packm_14xk_4mi_ker_name );
gen_func_init_co( &funcs[ BLIS_PACKM_16XK_KER ], packm_16xk_4mi_ker_name );
gen_func_init_co( &funcs[ BLIS_PACKM_30XK_KER ], packm_30xk_4mi_ker_name );
}
else if ( method == BLIS_1M )
{
gen_func_init_co( &funcs[ BLIS_PACKM_2XK_KER ], packm_2xk_1er_ker_name );
gen_func_init_co( &funcs[ BLIS_PACKM_4XK_KER ], packm_4xk_1er_ker_name );
gen_func_init_co( &funcs[ BLIS_PACKM_6XK_KER ], packm_6xk_1er_ker_name );
gen_func_init_co( &funcs[ BLIS_PACKM_8XK_KER ], packm_8xk_1er_ker_name );
gen_func_init_co( &funcs[ BLIS_PACKM_10XK_KER ], packm_10xk_1er_ker_name );
gen_func_init_co( &funcs[ BLIS_PACKM_12XK_KER ], packm_12xk_1er_ker_name );
gen_func_init_co( &funcs[ BLIS_PACKM_14XK_KER ], packm_14xk_1er_ker_name );
gen_func_init_co( &funcs[ BLIS_PACKM_16XK_KER ], packm_16xk_1er_ker_name );
gen_func_init_co( &funcs[ BLIS_PACKM_30XK_KER ], packm_30xk_1er_ker_name );
}
else // if ( method == BLIS_NAT )
{
gen_func_init( &funcs[ BLIS_PACKM_2XK_KER ], packm_2xk_ker_name );
gen_func_init( &funcs[ BLIS_PACKM_3XK_KER ], packm_3xk_ker_name );
gen_func_init( &funcs[ BLIS_PACKM_4XK_KER ], packm_4xk_ker_name );
gen_func_init( &funcs[ BLIS_PACKM_6XK_KER ], packm_6xk_ker_name );
gen_func_init( &funcs[ BLIS_PACKM_8XK_KER ], packm_8xk_ker_name );
gen_func_init( &funcs[ BLIS_PACKM_10XK_KER ], packm_10xk_ker_name );
gen_func_init( &funcs[ BLIS_PACKM_12XK_KER ], packm_12xk_ker_name );
gen_func_init( &funcs[ BLIS_PACKM_14XK_KER ], packm_14xk_ker_name );
gen_func_init( &funcs[ BLIS_PACKM_16XK_KER ], packm_16xk_ker_name );
gen_func_init( &funcs[ BLIS_PACKM_24XK_KER ], packm_24xk_ker_name );
gen_func_init( &funcs[ BLIS_PACKM_30XK_KER ], packm_30xk_ker_name );
}
// -- Set induced method cache and register blocksizes ---------------------
// Modify the context with cache and register blocksizes (and multiples)
// appropriate for the current induced method.
if ( method == BLIS_3MH )
{
bli_cntx_set_ind_blkszs
(
method, 6,
BLIS_NC, 1.0, 1.0,
BLIS_KC, 1.0, 1.0,
BLIS_MC, 1.0, 1.0,
BLIS_NR, 1.0, 1.0,
BLIS_MR, 1.0, 1.0,
BLIS_KR, 1.0, 1.0,
cntx
);
}
else if ( method == BLIS_3M1 )
{
bli_cntx_set_ind_blkszs
(
method, 6,
BLIS_NC, 1.0, 1.0,
BLIS_KC, 3.0, 3.0,
BLIS_MC, 1.0, 1.0,
BLIS_NR, 1.0, 1.0,
BLIS_MR, 1.0, 1.0,
BLIS_KR, 1.0, 1.0,
cntx
);
}
else if ( method == BLIS_4MH )
{
bli_cntx_set_ind_blkszs
(
method, 6,
BLIS_NC, 1.0, 1.0,
BLIS_KC, 1.0, 1.0,
BLIS_MC, 1.0, 1.0,
BLIS_NR, 1.0, 1.0,
BLIS_MR, 1.0, 1.0,
BLIS_KR, 1.0, 1.0,
cntx
);
}
else if ( method == BLIS_4M1B )
{
bli_cntx_set_ind_blkszs
(
method, 6,
BLIS_NC, 2.0, 2.0,
BLIS_KC, 1.0, 1.0,
BLIS_MC, 2.0, 2.0,
BLIS_NR, 1.0, 1.0,
BLIS_MR, 1.0, 1.0,
BLIS_KR, 1.0, 1.0,
cntx
);
}
else if ( method == BLIS_4M1A )
{
bli_cntx_set_ind_blkszs
(
method, 6,
BLIS_NC, 1.0, 1.0,
BLIS_KC, 2.0, 2.0,
BLIS_MC, 1.0, 1.0,
BLIS_NR, 1.0, 1.0,
BLIS_MR, 1.0, 1.0,
BLIS_KR, 1.0, 1.0,
cntx
);
}
else if ( method == BLIS_1M )
{
const bool_t is_pb = FALSE;
// We MUST set the induced method in the context prior to calling
// bli_cntx_l3_ukr_prefers_cols_dt() because that function queries
// the induced method. It needs the induced method value in order
// to determine whether to evaluate the "prefers column storage"
// predicate using the storage preference of the kernel for dt, or
// the storage preference of the kernel for the real projection of
// dt. Failing to set the induced method here can lead to strange
// undefined behavior at runtime if the native complex kernel's
// storage preference happens to not equal that of the native real
// kernel.
bli_cntx_set_method( method, cntx );
// Initialize the blocksizes according to the micro-kernel preference as
// well as the algorithm.
if ( bli_cntx_l3_ukr_prefers_cols_dt( dt, BLIS_GEMM_UKR, cntx ) )
{
// This branch is used for algorithms 1m_c_bp, 1m_r_pb.
// Set the pack_t schemas for the c_bp or r_pb algorithms.
if ( !is_pb )
{
bli_cntx_set_schema_a_block( BLIS_PACKED_ROW_PANELS_1E, cntx );
bli_cntx_set_schema_b_panel( BLIS_PACKED_COL_PANELS_1R, cntx );
}
else // if ( is_pb )
{
bli_cntx_set_schema_b_panel( BLIS_PACKED_ROW_PANELS_1R, cntx );
bli_cntx_set_schema_a_block( BLIS_PACKED_COL_PANELS_1E, cntx );
}
bli_cntx_set_ind_blkszs
(
method, 6,
BLIS_NC, 1.0, 1.0,
BLIS_KC, 2.0, 2.0, // halve kc...
BLIS_MC, 2.0, 2.0, // halve mc...
BLIS_NR, 1.0, 1.0,
BLIS_MR, 2.0, 1.0, // ...and mr (but NOT packmr)
BLIS_KR, 1.0, 1.0,
cntx
);
}
else // if ( bli_cntx_l3_ukr_prefers_rows_dt( dt, BLIS_GEMM_UKR, cntx ) )
{
// This branch is used for algorithms 1m_r_bp, 1m_c_pb.
// Set the pack_t schemas for the r_bp or c_pb algorithms.
if ( !is_pb )
{
bli_cntx_set_schema_a_block( BLIS_PACKED_ROW_PANELS_1R, cntx );
bli_cntx_set_schema_b_panel( BLIS_PACKED_COL_PANELS_1E, cntx );
}
else // if ( is_pb )
{
bli_cntx_set_schema_b_panel( BLIS_PACKED_ROW_PANELS_1E, cntx );
bli_cntx_set_schema_a_block( BLIS_PACKED_COL_PANELS_1R, cntx );
}
bli_cntx_set_ind_blkszs
(
method, 6,
BLIS_NC, 2.0, 2.0, // halve nc...
BLIS_KC, 2.0, 2.0, // halve kc...
BLIS_MC, 1.0, 1.0,
BLIS_NR, 2.0, 1.0, // ...and nr (but NOT packnr)
BLIS_MR, 1.0, 1.0,
BLIS_KR, 1.0, 1.0,
cntx
);
}
}
else // if ( method == BLIS_NAT )
{
// No change in blocksizes needed for native execution.
}
// -- Set misc. other fields -----------------------------------------------
if ( method == BLIS_3MH )
{
// Schemas vary with _stage().
}
else if ( method == BLIS_3M1 )
{
bli_cntx_set_schema_a_block( BLIS_PACKED_ROW_PANELS_3MI, cntx );
bli_cntx_set_schema_b_panel( BLIS_PACKED_COL_PANELS_3MI, cntx );
}
else if ( method == BLIS_4MH )
{
// Schemas vary with _stage().
}
else if ( method == BLIS_4M1A || method == BLIS_4M1B )
{
bli_cntx_set_schema_a_block( BLIS_PACKED_ROW_PANELS_4MI, cntx );
bli_cntx_set_schema_b_panel( BLIS_PACKED_COL_PANELS_4MI, cntx );
}
else if ( method == BLIS_1M )
{
const bool_t is_pb = FALSE;
// Set the anti-preference field to TRUE when executing a panel-block
// algorithm, and FALSE otherwise. This will cause higher-level generic
// code to establish (if needed) disagreement between the storage of C and
// the micro-kernel output preference so that the two will come back into
// agreement in the panel-block macro-kernel (which implemented in terms
// of the block-panel macro-kernel with some induced transpositions).
bli_cntx_set_anti_pref( is_pb, cntx );
}
else // if ( method == BLIS_NAT )
{
}
}
void GENBARNAME(cntx_init)
(
cntx_t* cntx
)
{
blksz_t blkszs[ BLIS_NUM_BLKSZS ];
func_t* funcs;
mbool_t* mbools;
dim_t i;
// -- Clear the context ----------------------------------------------------
bli_cntx_clear( cntx );
// -- Set blocksizes -------------------------------------------------------
// s d c z
bli_blksz_init_easy( &blkszs[ BLIS_KR ], 1, 1, 1, 1 );
bli_blksz_init_easy( &blkszs[ BLIS_MR ], 8, 4, 4, 2 );
bli_blksz_init_easy( &blkszs[ BLIS_NR ], 4, 4, 2, 2 );
bli_blksz_init_easy( &blkszs[ BLIS_MC ], 512, 256, 256, 128 );
bli_blksz_init_easy( &blkszs[ BLIS_KC ], 256, 256, 256, 256 );
bli_blksz_init_easy( &blkszs[ BLIS_NC ], 4096, 4096, 4096, 4096 );
bli_blksz_init_easy( &blkszs[ BLIS_M2 ], 1000, 1000, 1000, 1000 );
bli_blksz_init_easy( &blkszs[ BLIS_N2 ], 1000, 1000, 1000, 1000 );
bli_blksz_init_easy( &blkszs[ BLIS_AF ], 8, 4, 4, 2 );
bli_blksz_init_easy( &blkszs[ BLIS_DF ], 8, 4, 4, 2 );
bli_blksz_init_easy( &blkszs[ BLIS_XF ], 8, 4, 4, 2 );
// Initialize the context with the default blocksize objects and their
// multiples.
bli_cntx_set_blkszs
(
BLIS_NAT, 11,
BLIS_NC, &blkszs[ BLIS_NC ], BLIS_NR,
BLIS_KC, &blkszs[ BLIS_KC ], BLIS_KR,
BLIS_MC, &blkszs[ BLIS_MC ], BLIS_MR,
BLIS_NR, &blkszs[ BLIS_NR ], BLIS_NR,
BLIS_MR, &blkszs[ BLIS_MR ], BLIS_MR,
BLIS_KR, &blkszs[ BLIS_KR ], BLIS_KR,
BLIS_M2, &blkszs[ BLIS_M2 ], BLIS_M2,
BLIS_N2, &blkszs[ BLIS_N2 ], BLIS_N2,
BLIS_AF, &blkszs[ BLIS_AF ], BLIS_AF,
BLIS_DF, &blkszs[ BLIS_DF ], BLIS_DF,
BLIS_XF, &blkszs[ BLIS_XF ], BLIS_XF,
cntx
);
// -- Set level-3 virtual micro-kernels ------------------------------------
funcs = bli_cntx_l3_vir_ukrs_buf( cntx );
gen_func_init_co( &funcs[ BLIS_GEMM_UKR ], gemm1m_ukr_name );
gen_func_init_co( &funcs[ BLIS_GEMMTRSM_L_UKR ], gemmtrsm1m_l_ukr_name );
gen_func_init_co( &funcs[ BLIS_GEMMTRSM_U_UKR ], gemmtrsm1m_u_ukr_name );
gen_func_init_co( &funcs[ BLIS_TRSM_L_UKR ], trsm1m_l_ukr_name );
gen_func_init_co( &funcs[ BLIS_TRSM_U_UKR ], trsm1m_u_ukr_name );
// -- Set level-3 native micro-kernels and preferences ---------------------
funcs = bli_cntx_l3_nat_ukrs_buf( cntx );
mbools = bli_cntx_l3_nat_ukrs_prefs_buf( cntx );
gen_func_init( &funcs[ BLIS_GEMM_UKR ], gemm_ukr_name );
gen_func_init( &funcs[ BLIS_GEMMTRSM_L_UKR ], gemmtrsm_l_ukr_name );
gen_func_init( &funcs[ BLIS_GEMMTRSM_U_UKR ], gemmtrsm_u_ukr_name );
gen_func_init( &funcs[ BLIS_TRSM_L_UKR ], trsm_l_ukr_name );
gen_func_init( &funcs[ BLIS_TRSM_U_UKR ], trsm_u_ukr_name );
bli_mbool_init( &mbools[ BLIS_GEMM_UKR ], FALSE, FALSE, FALSE, FALSE );
bli_mbool_init( &mbools[ BLIS_GEMMTRSM_L_UKR ], FALSE, FALSE, FALSE, FALSE );
bli_mbool_init( &mbools[ BLIS_GEMMTRSM_U_UKR ], FALSE, FALSE, FALSE, FALSE );
bli_mbool_init( &mbools[ BLIS_TRSM_L_UKR ], FALSE, FALSE, FALSE, FALSE );
bli_mbool_init( &mbools[ BLIS_TRSM_U_UKR ], FALSE, FALSE, FALSE, FALSE );
// -- Set level-1f kernels -------------------------------------------------
funcs = bli_cntx_l1f_kers_buf( cntx );
gen_func_init( &funcs[ BLIS_AXPY2V_KER ], axpy2v_ker_name );
gen_func_init( &funcs[ BLIS_DOTAXPYV_KER ], dotaxpyv_ker_name );
gen_func_init( &funcs[ BLIS_AXPYF_KER ], axpyf_ker_name );
gen_func_init( &funcs[ BLIS_DOTXF_KER ], dotxf_ker_name );
gen_func_init( &funcs[ BLIS_DOTXAXPYF_KER ], dotxaxpyf_ker_name );
// -- Set level-1v kernels -------------------------------------------------
funcs = bli_cntx_l1v_kers_buf( cntx );
gen_func_init( &funcs[ BLIS_ADDV_KER ], addv_ker_name );
gen_func_init( &funcs[ BLIS_AMAXV_KER ], amaxv_ker_name );
gen_func_init( &funcs[ BLIS_AXPBYV_KER ], axpbyv_ker_name );
gen_func_init( &funcs[ BLIS_AXPYV_KER ], axpyv_ker_name );
gen_func_init( &funcs[ BLIS_COPYV_KER ], copyv_ker_name );
gen_func_init( &funcs[ BLIS_DOTV_KER ], dotv_ker_name );
gen_func_init( &funcs[ BLIS_DOTXV_KER ], dotxv_ker_name );
gen_func_init( &funcs[ BLIS_INVERTV_KER ], invertv_ker_name );
gen_func_init( &funcs[ BLIS_SCALV_KER ], scalv_ker_name );
gen_func_init( &funcs[ BLIS_SCAL2V_KER ], scal2v_ker_name );
gen_func_init( &funcs[ BLIS_SETV_KER ], setv_ker_name );
gen_func_init( &funcs[ BLIS_SUBV_KER ], subv_ker_name );
gen_func_init( &funcs[ BLIS_SWAPV_KER ], swapv_ker_name );
gen_func_init( &funcs[ BLIS_XPBYV_KER ], xpbyv_ker_name );
// -- Set level-1m (packm/unpackm) kernels ---------------------------------
funcs = bli_cntx_packm_kers_buf( cntx );
// Initialize all packm kernel func_t entries to NULL.
for ( i = BLIS_PACKM_0XK_KER; i <= BLIS_PACKM_31XK_KER; ++i )
{
bli_func_init_null( &funcs[ i ] );
}
gen_func_init( &funcs[ BLIS_PACKM_2XK_KER ], packm_2xk_ker_name );
gen_func_init( &funcs[ BLIS_PACKM_3XK_KER ], packm_3xk_ker_name );
gen_func_init( &funcs[ BLIS_PACKM_4XK_KER ], packm_4xk_ker_name );
gen_func_init( &funcs[ BLIS_PACKM_6XK_KER ], packm_6xk_ker_name );
gen_func_init( &funcs[ BLIS_PACKM_8XK_KER ], packm_8xk_ker_name );
gen_func_init( &funcs[ BLIS_PACKM_10XK_KER ], packm_10xk_ker_name );
gen_func_init( &funcs[ BLIS_PACKM_12XK_KER ], packm_12xk_ker_name );
gen_func_init( &funcs[ BLIS_PACKM_14XK_KER ], packm_14xk_ker_name );
gen_func_init( &funcs[ BLIS_PACKM_16XK_KER ], packm_16xk_ker_name );
gen_func_init( &funcs[ BLIS_PACKM_24XK_KER ], packm_24xk_ker_name );
gen_func_init( &funcs[ BLIS_PACKM_30XK_KER ], packm_30xk_ker_name );
funcs = bli_cntx_unpackm_kers_buf( cntx );
// Initialize all packm kernel func_t entries to NULL.
for ( i = BLIS_UNPACKM_0XK_KER; i <= BLIS_UNPACKM_31XK_KER; ++i )
{
bli_func_init_null( &funcs[ i ] );
}
gen_func_init( &funcs[ BLIS_UNPACKM_2XK_KER ], unpackm_2xk_ker_name );
gen_func_init( &funcs[ BLIS_UNPACKM_4XK_KER ], unpackm_4xk_ker_name );
gen_func_init( &funcs[ BLIS_UNPACKM_6XK_KER ], unpackm_6xk_ker_name );
gen_func_init( &funcs[ BLIS_UNPACKM_8XK_KER ], unpackm_8xk_ker_name );
gen_func_init( &funcs[ BLIS_UNPACKM_10XK_KER ], unpackm_10xk_ker_name );
gen_func_init( &funcs[ BLIS_UNPACKM_12XK_KER ], unpackm_12xk_ker_name );
gen_func_init( &funcs[ BLIS_UNPACKM_14XK_KER ], unpackm_14xk_ker_name );
gen_func_init( &funcs[ BLIS_UNPACKM_16XK_KER ], unpackm_16xk_ker_name );
// -- Set miscellaneous fields ---------------------------------------------
bli_cntx_set_method( BLIS_NAT, cntx );
bli_cntx_set_schema_a_block( BLIS_PACKED_ROW_PANELS, cntx );
bli_cntx_set_schema_b_panel( BLIS_PACKED_COL_PANELS, cntx );
bli_cntx_set_schema_c_panel( BLIS_NOT_PACKED, cntx );
bli_cntx_set_anti_pref( FALSE, cntx );
bli_cntx_set_thrloop( 1, 1, 1, 1, 1, cntx );
bli_cntx_set_membrk( bli_memsys_global_membrk(), cntx );
}
