void bli_trmm_blk_var1( obj_t* alpha,
obj_t* a,
obj_t* b,
obj_t* beta,
obj_t* c,
trmm_t* cntl )
{
obj_t a1, a1_pack;
obj_t b_pack;
obj_t c1, c1_pack;
dim_t i;
dim_t b_alg;
dim_t m_trans;
dim_t offA;
// Initialize all pack objects that are passed into packm_init().
bli_obj_init_pack( &a1_pack );
bli_obj_init_pack( &b_pack );
bli_obj_init_pack( &c1_pack );
// Set the default length of and offset to the non-zero part of A.
m_trans = bli_obj_length_after_trans( *a );
offA = 0;
// If A is lower triangular, we have to adjust where the non-zero part of
// A begins. If A is upper triangular, we have to adjust the length of
// the non-zero part. If A is general/dense, then we keep the defaults.
if ( bli_obj_is_lower( *a ) )
offA = bli_abs( bli_obj_diag_offset_after_trans( *a ) );
else if ( bli_obj_is_upper( *a ) )
m_trans = bli_abs( bli_obj_diag_offset_after_trans( *a ) ) +
bli_obj_width_after_trans( *a );
// Scale C by beta (if instructed).
bli_scalm_int( beta,
c,
cntl_sub_scalm( cntl ) );
// Initialize object for packing B.
bli_packm_init( b, &b_pack,
cntl_sub_packm_b( cntl ) );
// Pack B and scale by alpha (if instructed).
bli_packm_int( alpha,
b, &b_pack,
cntl_sub_packm_b( cntl ) );
// Partition along the m dimension.
for ( i = offA; i < m_trans; i += b_alg )
{
// Determine the current algorithmic blocksize.
b_alg = bli_determine_blocksize_f( i, m_trans, a,
cntl_blocksize( cntl ) );
// Acquire partitions for A1 and C1.
bli_acquire_mpart_t2b( BLIS_SUBPART1,
i, b_alg, a, &a1 );
bli_acquire_mpart_t2b( BLIS_SUBPART1,
i, b_alg, c, &c1 );
// Initialize objects for packing A1 and C1.
bli_packm_init( &a1, &a1_pack,
cntl_sub_packm_a( cntl ) );
bli_packm_init( &c1, &c1_pack,
cntl_sub_packm_c( cntl ) );
// Pack A1 and scale by alpha (if instructed).
bli_packm_int( alpha,
&a1, &a1_pack,
cntl_sub_packm_a( cntl ) );
// Pack C1 and scale by beta (if instructed).
bli_packm_int( beta,
&c1, &c1_pack,
cntl_sub_packm_c( cntl ) );
// Perform trmm subproblem.
bli_trmm_int( alpha,
&a1_pack,
&b_pack,
beta,
&c1_pack,
cntl_sub_trmm( cntl ) );
// Unpack C1 (if C1 was packed).
bli_unpackm_int( &c1_pack, &c1,
cntl_sub_unpackm_c( cntl ) );
}
// If any packing buffers were acquired within packm, release them back
// to the memory manager.
bli_obj_release_pack( &a1_pack );
bli_obj_release_pack( &b_pack );
bli_obj_release_pack( &c1_pack );
}
void bli_gemm_blk_var3f( obj_t* a,
obj_t* b,
obj_t* c,
gemm_t* cntl,
gemm_thrinfo_t* thread )
{
obj_t c_pack_s;
obj_t a1_pack_s, b1_pack_s;
obj_t a1, b1;
obj_t* a1_pack = NULL;
obj_t* b1_pack = NULL;
obj_t* c_pack = NULL;
dim_t i;
dim_t b_alg;
dim_t k_trans;
if( thread_am_ochief( thread ) ){
// Initialize object for packing C
bli_obj_init_pack( &c_pack_s );
bli_packm_init( c, &c_pack_s,
cntl_sub_packm_c( cntl ) );
// Scale C by beta (if instructed).
bli_scalm_int( &BLIS_ONE,
c,
cntl_sub_scalm( cntl ) );
}
c_pack = thread_obroadcast( thread, &c_pack_s );
// Initialize pack objects for A and B that are passed into packm_init().
if( thread_am_ichief( thread ) ){
bli_obj_init_pack( &a1_pack_s );
bli_obj_init_pack( &b1_pack_s );
}
a1_pack = thread_ibroadcast( thread, &a1_pack_s );
b1_pack = thread_ibroadcast( thread, &b1_pack_s );
// Pack C (if instructed).
bli_packm_int( c, c_pack,
cntl_sub_packm_c( cntl ),
gemm_thread_sub_opackm( thread ) );
// Query dimension in partitioning direction.
k_trans = bli_obj_width_after_trans( *a );
// Partition along the k dimension.
for ( i = 0; i < k_trans; i += b_alg )
{
// Determine the current algorithmic blocksize.
// NOTE: We call a gemm/hemm/symm-specific function to determine
// the kc blocksize so that we can implement the "nudging" of kc
// to be a multiple of mr or nr, as needed.
b_alg = bli_gemm_determine_kc_f( i, k_trans, a, b,
cntl_blocksize( cntl ) );
// Acquire partitions for A1 and B1.
bli_acquire_mpart_l2r( BLIS_SUBPART1,
i, b_alg, a, &a1 );
bli_acquire_mpart_t2b( BLIS_SUBPART1,
i, b_alg, b, &b1 );
// Initialize objects for packing A1 and B1.
if( thread_am_ichief( thread ) ) {
bli_packm_init( &a1, a1_pack,
cntl_sub_packm_a( cntl ) );
bli_packm_init( &b1, b1_pack,
cntl_sub_packm_b( cntl ) );
}
thread_ibarrier( thread );
// Pack A1 (if instructed).
bli_packm_int( &a1, a1_pack,
cntl_sub_packm_a( cntl ),
gemm_thread_sub_ipackm( thread ) );
// Pack B1 (if instructed).
bli_packm_int( &b1, b1_pack,
cntl_sub_packm_b( cntl ),
gemm_thread_sub_ipackm( thread ) );
// Perform gemm subproblem.
bli_gemm_int( &BLIS_ONE,
a1_pack,
b1_pack,
&BLIS_ONE,
c_pack,
cntl_sub_gemm( cntl ),
gemm_thread_sub_gemm( thread) );
// This variant executes multiple rank-k updates. Therefore, if the
// internal beta scalar on matrix C is non-zero, we must use it
// only for the first iteration (and then BLIS_ONE for all others).
// And since c_pack is a local obj_t, we can simply overwrite the
// internal beta scalar with BLIS_ONE once it has been used in the
// first iteration.
thread_ibarrier( thread );
if ( i == 0 && thread_am_ichief( thread ) ) bli_obj_scalar_reset( c_pack );
}
thread_obarrier( thread );
// Unpack C (if C was packed).
bli_unpackm_int( c_pack, c,
cntl_sub_unpackm_c( cntl ),
gemm_thread_sub_opackm( thread ) );
// If any packing buffers were acquired within packm, release them back
// to the memory manager.
if( thread_am_ochief( thread ) )
bli_packm_release( c_pack, cntl_sub_packm_c( cntl ) );
if( thread_am_ichief( thread ) ){
bli_packm_release( a1_pack, cntl_sub_packm_a( cntl ) );
bli_packm_release( b1_pack, cntl_sub_packm_b( cntl ) );
}
}
