void bli_trsm_blk_var3b( obj_t* a,
obj_t* b,
obj_t* c,
trsm_t* cntl,
trsm_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;
// Prune any zero region that exists along the partitioning dimension.
bli_trsm_prune_unref_mparts_k( a, b, c );
// Initialize pack objects for C that are passed into packm_init().
if( thread_am_ochief( thread ) ) {
bli_obj_init_pack( &c_pack_s );
// Initialize object for packing C.
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 );
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 ),
trsm_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 trsm-specific function to determine the kc
// blocksize so that we can implement the "nudging" of kc to be
// a multiple of mr, as needed.
b_alg = bli_trsm_determine_kc_b( i, k_trans, b,
cntl_blocksize( cntl ) );
// Acquire partitions for A1 and B1.
bli_acquire_mpart_r2l( BLIS_SUBPART1,
i, b_alg, a, &a1 );
bli_acquire_mpart_b2t( 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 ),
trsm_thread_sub_ipackm( thread ) );
// Pack B1 (if instructed).
bli_packm_int( &b1, b1_pack,
cntl_sub_packm_b( cntl ),
trsm_thread_sub_ipackm( thread ) );
// Perform trsm subproblem.
bli_trsm_int( &BLIS_ONE,
a1_pack,
b1_pack,
&BLIS_ONE,
c_pack,
cntl_sub_trsm( cntl ),
trsm_thread_sub_trsm( thread ) );
// This variant executes multiple rank-k updates. Therefore, if the
// internal alpha scalars on A/B and C are non-zero, we must ensure
// that they are only used in the first iteration.
thread_ibarrier( thread );
if ( i == 0 && thread_am_ichief( thread ) ) {
bli_obj_scalar_reset( a );
bli_obj_scalar_reset( b );
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 ),
trsm_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 ) );
}
}
void bli_gemm_blk_var1f( obj_t* a,
obj_t* b,
obj_t* c,
cntx_t* cntx,
gemm_t* cntl,
gemm_thrinfo_t* thread )
{
//The s is for "lives on the stack"
obj_t b_pack_s;
obj_t a1_pack_s, c1_pack_s;
obj_t a1, c1;
obj_t* a1_pack = NULL;
obj_t* b_pack = NULL;
obj_t* c1_pack = NULL;
dim_t i;
dim_t b_alg;
if( thread_am_ochief( thread ) ) {
// Initialize object for packing B.
bli_obj_init_pack( &b_pack_s );
bli_packm_init( b, &b_pack_s,
cntx, cntl_sub_packm_b( cntl ) );
// Scale C by beta (if instructed).
// Since scalm doesn't support multithreading yet, must be done by chief thread (ew)
bli_scalm_int( &BLIS_ONE,
c,
cntx, cntl_sub_scalm( cntl ) );
}
b_pack = thread_obroadcast( thread, &b_pack_s );
// Initialize objects passed into bli_packm_init for A and C
if( thread_am_ichief( thread ) ) {
bli_obj_init_pack( &a1_pack_s );
bli_obj_init_pack( &c1_pack_s );
}
a1_pack = thread_ibroadcast( thread, &a1_pack_s );
c1_pack = thread_ibroadcast( thread, &c1_pack_s );
// Pack B (if instructed).
bli_packm_int( b, b_pack,
cntx, cntl_sub_packm_b( cntl ),
gemm_thread_sub_opackm( thread ) );
dim_t my_start, my_end;
bli_get_range_t2b( thread, a,
bli_cntx_get_bmult( cntl_bszid( cntl ), cntx ),
&my_start, &my_end );
// Partition along the m dimension.
for ( i = my_start; i < my_end; i += b_alg )
{
// Determine the current algorithmic blocksize.
// NOTE: Use of a (for execution datatype) is intentional!
// This causes the right blocksize to be used if c and a are
// complex and b is real.
b_alg = bli_determine_blocksize_f( i, my_end, a,
cntl_bszid( cntl ), cntx );
// 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.
if( thread_am_ichief( thread ) ) {
bli_packm_init( &a1, a1_pack,
cntx, cntl_sub_packm_a( cntl ) );
bli_packm_init( &c1, c1_pack,
cntx, cntl_sub_packm_c( cntl ) );
}
thread_ibarrier( thread );
// Pack A1 (if instructed).
bli_packm_int( &a1, a1_pack,
cntx, cntl_sub_packm_a( cntl ),
gemm_thread_sub_ipackm( thread ) );
// Pack C1 (if instructed).
bli_packm_int( &c1, c1_pack,
cntx, cntl_sub_packm_c( cntl ),
gemm_thread_sub_ipackm( thread ) );
// Perform gemm subproblem.
bli_gemm_int( &BLIS_ONE,
a1_pack,
b_pack,
&BLIS_ONE,
c1_pack,
cntx,
cntl_sub_gemm( cntl ),
gemm_thread_sub_gemm( thread ) );
thread_ibarrier( thread );
// Unpack C1 (if C1 was packed).
// Currently must be done by 1 thread
bli_unpackm_int( c1_pack, &c1,
cntx, cntl_sub_unpackm_c( cntl ),
gemm_thread_sub_ipackm( thread ) );
}
// If any packing buffers were acquired within packm, release them back
// to the memory manager.
thread_obarrier( thread );
if( thread_am_ochief( thread ) )
bli_packm_release( b_pack, cntl_sub_packm_b( cntl ) );
if( thread_am_ichief( thread ) ){
bli_packm_release( a1_pack, cntl_sub_packm_a( cntl ) );
bli_packm_release( c1_pack, cntl_sub_packm_c( cntl ) );
}
}
void bli_trsm_blk_var1f( obj_t* a,
obj_t* b,
obj_t* c,
trsm_t* cntl,
trsm_thrinfo_t* thread )
{
obj_t b_pack_s;
obj_t a1_pack_s;
obj_t a1, c1;
obj_t* b_pack = NULL;
obj_t* a1_pack = NULL;
dim_t i;
dim_t b_alg;
dim_t m_trans;
dim_t offA;
// Initialize object for packing B.
if( thread_am_ochief( thread ) ) {
bli_obj_init_pack( &b_pack_s );
bli_packm_init( b, &b_pack_s,
cntl_sub_packm_b( cntl ) );
}
b_pack = thread_obroadcast( thread, &b_pack_s );
// Initialize object for packing B.
if( thread_am_ichief( thread ) ) {
bli_obj_init_pack( &a1_pack_s );
}
a1_pack = thread_ibroadcast( thread, &a1_pack_s );
// Pack B1 (if instructed).
bli_packm_int( b, b_pack,
cntl_sub_packm_b( cntl ),
trsm_thread_sub_opackm( thread ) );
// 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 ( bli_obj_is_lower( *a ) )
offA = bli_abs( bli_obj_diag_offset_after_trans( *a ) );
dim_t start, end;
num_t dt = bli_obj_execution_datatype( *a );
bli_get_range_t2b( thread, offA, m_trans,
//bli_lcm( bli_info_get_default_nr( BLIS_TRSM, dt ), bli_info_get_default_mr( BLIS_TRSM, dt ) ),
bli_info_get_default_mc( BLIS_TRSM, dt ),
&start, &end );
// Partition along the remaining portion of the m dimension.
for ( i = start; i < end; i += b_alg )
{
// Determine the current algorithmic blocksize.
b_alg = bli_determine_blocksize_f( i, end, 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 object for packing A1.
if( thread_am_ichief( thread ) ) {
bli_packm_init( &a1, a1_pack,
cntl_sub_packm_a( cntl ) );
}
thread_ibarrier( thread );
// Pack A1 (if instructed).
bli_packm_int( &a1, a1_pack,
cntl_sub_packm_a( cntl ),
trsm_thread_sub_ipackm( thread ) );
// Perform trsm subproblem.
bli_trsm_int( &BLIS_ONE,
a1_pack,
b_pack,
&BLIS_ONE,
&c1,
cntl_sub_trsm( cntl ),
trsm_thread_sub_trsm( thread ) );
thread_ibarrier( thread );
}
// If any packing buffers were acquired within packm, release them back
// to the memory manager.
thread_obarrier( thread );
if( thread_am_ochief( thread ) )
bli_packm_release( b_pack, cntl_sub_packm_b( cntl ) );
if( thread_am_ichief( thread ) )
bli_packm_release( a1_pack, cntl_sub_packm_a( cntl ) );
}
void bli_trsm_blk_var2b( obj_t* a,
obj_t* b,
obj_t* c,
trsm_t* cntl,
trsm_thrinfo_t* thread )
{
obj_t a_pack_s;
obj_t b1_pack_s, c1_pack_s;
obj_t b1, c1;
obj_t* a_pack = NULL;
obj_t* b1_pack = NULL;
obj_t* c1_pack = NULL;
dim_t i;
dim_t b_alg;
dim_t n_trans;
// Initialize pack objects for A that are passed into packm_init().
if( thread_am_ochief( thread ) ) {
bli_obj_init_pack( &a_pack_s );
// Initialize object for packing A.
bli_packm_init( a, &a_pack_s,
cntl_sub_packm_a( cntl ) );
// Scale C by beta (if instructed).
bli_scalm_int( &BLIS_ONE,
c,
cntl_sub_scalm( cntl ) );
}
a_pack = thread_obroadcast( thread, &a_pack_s );
// Initialize pack objects for B and C that are passed into packm_init().
if( thread_am_ichief( thread ) ) {
bli_obj_init_pack( &b1_pack_s );
bli_obj_init_pack( &c1_pack_s );
}
b1_pack = thread_ibroadcast( thread, &b1_pack_s );
c1_pack = thread_ibroadcast( thread, &c1_pack_s );
// Pack A (if instructed).
bli_packm_int( a, a_pack,
cntl_sub_packm_a( cntl ),
trmm_thread_sub_opackm( thread ) );
// Query dimension in partitioning direction.
n_trans = bli_obj_width_after_trans( *b );
dim_t start, end;
num_t dt = bli_obj_execution_datatype( *a );
bli_get_range_r2l( thread, 0, n_trans,
//bli_lcm( bli_info_get_default_nr( BLIS_TRSM, dt ),
// bli_info_get_default_mr( BLIS_TRSM, dt ) ),
bli_lcm( bli_blksz_get_nr( dt, cntl_blocksize( cntl ) ),
bli_blksz_get_mr( dt, cntl_blocksize( cntl ) ) ),
&start, &end );
// Partition along the n dimension.
for ( i = start; i < end; i += b_alg )
{
// Determine the current algorithmic blocksize.
b_alg = bli_determine_blocksize_b( i, end, b,
cntl_blocksize( cntl ) );
// Acquire partitions for B1 and C1.
bli_acquire_mpart_r2l( BLIS_SUBPART1,
i, b_alg, b, &b1 );
bli_acquire_mpart_r2l( BLIS_SUBPART1,
i, b_alg, c, &c1 );
// Initialize objects for packing A1 and B1.
if( thread_am_ichief( thread ) ) {
bli_packm_init( &b1, b1_pack,
cntl_sub_packm_b( cntl ) );
bli_packm_init( &c1, c1_pack,
cntl_sub_packm_c( cntl ) );
}
thread_ibarrier( thread );
// Pack B1 (if instructed).
bli_packm_int( &b1, b1_pack,
cntl_sub_packm_b( cntl ),
trsm_thread_sub_ipackm( thread ) );
// Pack C1 (if instructed).
bli_packm_int( &c1, c1_pack,
cntl_sub_packm_c( cntl ),
trsm_thread_sub_ipackm( thread ) );
// Perform trsm subproblem.
bli_trsm_int( &BLIS_ONE,
a_pack,
b1_pack,
&BLIS_ONE,
c1_pack,
cntl_sub_trsm( cntl ),
trsm_thread_sub_trsm( thread ) );
thread_ibarrier( thread );
// Unpack C1 (if C1 was packed).
bli_unpackm_int( c1_pack, &c1,
cntl_sub_unpackm_c( cntl ),
trsm_thread_sub_ipackm( thread ) );
}
// If any packing buffers were acquired within packm, release them back
// to the memory manager.
thread_obarrier( thread );
if( thread_am_ochief( thread ) )
bli_packm_release( a_pack, cntl_sub_packm_a( cntl ) );
if( thread_am_ichief( thread ) ) {
bli_packm_release( b1_pack, cntl_sub_packm_b( cntl ) );
bli_packm_release( c1_pack, cntl_sub_packm_c( cntl ) );
}
}
void bli_trsm_blk_var1b( obj_t* a,
obj_t* b,
obj_t* c,
trsm_t* cntl,
trsm_thrinfo_t* thread )
{
obj_t b_pack_s;
obj_t a1_pack_s;
obj_t a1, c1;
obj_t* b_pack = NULL;
obj_t* a1_pack = NULL;
dim_t i;
dim_t b_alg;
// Prune any zero region that exists along the partitioning dimension.
bli_trsm_prune_unref_mparts_m( a, b, c );
// Initialize object for packing B.
if( thread_am_ochief( thread ) ) {
bli_obj_init_pack( &b_pack_s );
bli_packm_init( b, &b_pack_s,
cntl_sub_packm_b( cntl ) );
}
b_pack = thread_obroadcast( thread, &b_pack_s );
// Initialize object for packing B.
if( thread_am_ichief( thread ) ) {
bli_obj_init_pack( &a1_pack_s );
}
a1_pack = thread_ibroadcast( thread, &a1_pack_s );
// Pack B1 (if instructed).
bli_packm_int( b, b_pack,
cntl_sub_packm_b( cntl ),
trsm_thread_sub_opackm( thread ) );
dim_t my_start, my_end;
num_t dt = bli_obj_execution_datatype( *a );
dim_t bf = ( bli_obj_root_is_triangular( *a ) ?
bli_info_get_default_mr( BLIS_TRSM, dt ) :
bli_info_get_default_nr( BLIS_TRSM, dt ) );
bli_get_range_b2t( thread, a, bf,
&my_start, &my_end );
// Partition along the remaining portion of the m dimension.
for ( i = my_start; i < my_end; i += b_alg )
{
// Determine the current algorithmic blocksize.
b_alg = bli_determine_blocksize_b( i, my_end, a,
cntl_blocksize( cntl ) );
// Acquire partitions for A1 and C1.
bli_acquire_mpart_b2t( BLIS_SUBPART1,
i, b_alg, a, &a1 );
bli_acquire_mpart_b2t( BLIS_SUBPART1,
i, b_alg, c, &c1 );
// Initialize object for packing A1.
if( thread_am_ichief( thread ) ) {
bli_packm_init( &a1, a1_pack,
cntl_sub_packm_a( cntl ) );
}
thread_ibarrier( thread );
// Pack A1 (if instructed).
bli_packm_int( &a1, a1_pack,
cntl_sub_packm_a( cntl ),
trsm_thread_sub_ipackm( thread ) );
// Perform trsm subproblem.
bli_trsm_int( &BLIS_ONE,
a1_pack,
b_pack,
&BLIS_ONE,
&c1,
cntl_sub_trsm( cntl ),
trsm_thread_sub_trsm( thread ) );
thread_ibarrier( thread );
}
// If any packing buffers were acquired within packm, release them back
// to the memory manager.
thread_obarrier( thread );
if( thread_am_ochief( thread ) )
bli_packm_release( b_pack, cntl_sub_packm_b( cntl ) );
if( thread_am_ichief( thread ) )
bli_packm_release( a1_pack, cntl_sub_packm_a( cntl ) );
}
void bli_herk_blk_var3f( obj_t* a,
obj_t* ah,
obj_t* c,
herk_t* cntl,
herk_thrinfo_t* thread )
{
obj_t c_pack_s;
obj_t a1_pack_s, ah1_pack_s;
obj_t a1, ah1;
obj_t* a1_pack = NULL;
obj_t* ah1_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 all pack objects that are passed into packm_init().
if( thread_am_ichief( thread ) ) {
bli_obj_init_pack( &a1_pack_s );
bli_obj_init_pack( &ah1_pack_s );
}
a1_pack = thread_ibroadcast( thread, &a1_pack_s );
ah1_pack = thread_ibroadcast( thread, &ah1_pack_s );
// Pack C (if instructed).
bli_packm_int( c, c_pack,
cntl_sub_packm_c( cntl ),
herk_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.
b_alg = bli_determine_blocksize_f( i, k_trans, a,
cntl_blocksize( cntl ) );
// Acquire partitions for A1 and A1'.
bli_acquire_mpart_l2r( BLIS_SUBPART1,
i, b_alg, a, &a1 );
bli_acquire_mpart_t2b( BLIS_SUBPART1,
i, b_alg, ah, &ah1 );
// Initialize objects for packing A1 and A1'.
if( thread_am_ichief( thread ) ) {
bli_packm_init( &a1, a1_pack,
cntl_sub_packm_a( cntl ) );
bli_packm_init( &ah1, ah1_pack,
cntl_sub_packm_b( cntl ) );
}
thread_ibarrier( thread );
// Pack A1 (if instructed).
bli_packm_int( &a1, a1_pack,
cntl_sub_packm_a( cntl ),
herk_thread_sub_ipackm( thread ) );
// Pack B1 (if instructed).
bli_packm_int( &ah1, ah1_pack,
cntl_sub_packm_b( cntl ),
herk_thread_sub_ipackm( thread ) );
// Perform herk subproblem.
bli_herk_int( &BLIS_ONE,
a1_pack,
ah1_pack,
&BLIS_ONE,
c_pack,
cntl_sub_herk( cntl ),
herk_thread_sub_herk( 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.
if ( i == 0 ) 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 ),
herk_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_obj_release_pack( c_pack );
}
if( thread_am_ichief( thread ) ) {
bli_obj_release_pack( a1_pack );
bli_obj_release_pack( ah1_pack );
}
}
void bli_herk_blk_var1f( obj_t* a,
obj_t* ah,
obj_t* c,
gemm_t* cntl,
herk_thrinfo_t* thread )
{
obj_t ah_pack_s;
obj_t a1_pack_s, c1_pack_s;
obj_t a1, c1;
obj_t* a1_pack;
obj_t* c1_pack;
obj_t* ah_pack;
dim_t i;
dim_t b_alg;
// Prune any zero region that exists along the partitioning dimension.
bli_herk_prune_unref_mparts_m( a, ah, c );
if( thread_am_ochief( thread ) ) {
// Initialize object for packing A'.
bli_obj_init_pack( &ah_pack_s );
bli_packm_init( ah, &ah_pack_s,
cntl_sub_packm_b( cntl ) );
// Scale C by beta (if instructed).
// Since scalm doesn't support multithreading yet, must be done by chief thread (ew)
bli_scalm_int( &BLIS_ONE,
c,
cntl_sub_scalm( cntl ) );
}
ah_pack = thread_obroadcast( thread, &ah_pack_s );
// Initialize pack objects that are passed into packm_init() for A and C.
if( thread_am_ichief( thread ) ) {
bli_obj_init_pack( &a1_pack_s );
bli_obj_init_pack( &c1_pack_s );
}
a1_pack = thread_ibroadcast( thread, &a1_pack_s );
c1_pack = thread_ibroadcast( thread, &c1_pack_s );
// Pack A' (if instructed).
bli_packm_int( ah, ah_pack,
cntl_sub_packm_b( cntl ),
herk_thread_sub_opackm( thread ) );
dim_t my_start, my_end;
bli_get_range_weighted_t2b( thread, c,
bli_blksz_get_mult_for_obj( a, cntl_blocksize( cntl ) ),
&my_start, &my_end );
// Partition along the m dimension.
for ( i = my_start; i < my_end; i += b_alg )
{
// Determine the current algorithmic blocksize.
b_alg = bli_determine_blocksize_f( i, my_end, 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.
if( thread_am_ichief( thread ) ) {
bli_packm_init( &a1, a1_pack,
cntl_sub_packm_a( cntl ) );
bli_packm_init( &c1, c1_pack,
cntl_sub_packm_c( cntl ) );
}
thread_ibarrier( thread );
// Pack A1 (if instructed).
bli_packm_int( &a1, a1_pack,
cntl_sub_packm_a( cntl ),
herk_thread_sub_ipackm( thread ) );
// Pack C1 (if instructed).
bli_packm_int( &c1, c1_pack,
cntl_sub_packm_c( cntl ),
herk_thread_sub_ipackm( thread ) );
// Perform herk subproblem.
bli_herk_int( &BLIS_ONE,
a1_pack,
ah_pack,
&BLIS_ONE,
c1_pack,
cntl_sub_gemm( cntl ),
herk_thread_sub_herk( thread ) );
thread_ibarrier( thread );
// Unpack C1 (if C1 was packed).
bli_unpackm_int( c1_pack, &c1,
cntl_sub_unpackm_c( cntl ),
herk_thread_sub_ipackm( thread ) );
}
// If any packing buffers were acquired within packm, release them back
// to the memory manager.
thread_obarrier( thread );
if( thread_am_ochief( thread ) )
bli_packm_release( ah_pack, cntl_sub_packm_b( cntl ) );
if( thread_am_ichief( thread ) ) {
bli_packm_release( a1_pack, cntl_sub_packm_a( cntl ) );
bli_packm_release( c1_pack, cntl_sub_packm_c( cntl ) );
}
}
void bli_trmm_blk_var1f( obj_t* a,
obj_t* b,
obj_t* c,
gemm_t* cntl,
trmm_thrinfo_t* thread )
{
obj_t b_pack_s;
obj_t a1_pack_s, c1_pack_s;
obj_t a1, c1;
obj_t* a1_pack = NULL;
obj_t* b_pack = NULL;
obj_t* c1_pack = NULL;
dim_t i;
dim_t b_alg;
// Prune any zero region that exists along the partitioning dimension.
bli_trmm_prune_unref_mparts_m( a, b, c );
if( thread_am_ochief( thread ) ) {
// Initialize object for packing B.
bli_obj_init_pack( &b_pack_s );
bli_packm_init( b, &b_pack_s,
cntl_sub_packm_b( cntl ) );
// Scale C by beta (if instructed).
// Since scalm doesn't support multithreading yet, must be done by chief thread (ew)
bli_scalm_int( &BLIS_ONE,
c,
cntl_sub_scalm( cntl ) );
}
b_pack = thread_obroadcast( thread, &b_pack_s );
// Initialize all pack objects that are passed into packm_init().
if( thread_am_ichief( thread ) ) {
bli_obj_init_pack( &a1_pack_s );
bli_obj_init_pack( &c1_pack_s );
}
a1_pack = thread_ibroadcast( thread, &a1_pack_s );
c1_pack = thread_ibroadcast( thread, &c1_pack_s );
// Pack B (if instructed).
bli_packm_int( b, b_pack,
cntl_sub_packm_b( cntl ),
trmm_thread_sub_opackm( thread ) );
// 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 );
dim_t my_start, my_end;
bli_get_range_weighted_t2b( thread, a,
bli_blksz_get_mult_for_obj( a, cntl_blocksize( cntl ) ),
&my_start, &my_end );
// Partition along the m dimension.
for ( i = my_start; i < my_end; i += b_alg )
{
// Determine the current algorithmic blocksize.
b_alg = bli_determine_blocksize_f( i, my_end, 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.
if( thread_am_ichief( thread ) ) {
bli_packm_init( &a1, a1_pack,
cntl_sub_packm_a( cntl ) );
bli_packm_init( &c1, c1_pack,
cntl_sub_packm_c( cntl ) );
}
thread_ibarrier( thread );
// Pack A1 (if instructed).
bli_packm_int( &a1, a1_pack,
cntl_sub_packm_a( cntl ),
trmm_thread_sub_ipackm( thread ) );
// Pack C1 (if instructed).
bli_packm_int( &c1, c1_pack,
cntl_sub_packm_c( cntl ),
trmm_thread_sub_ipackm( thread ) );
// Perform trmm subproblem.
bli_trmm_int( &BLIS_ONE,
a1_pack,
b_pack,
&BLIS_ONE,
c1_pack,
cntl_sub_gemm( cntl ),
trmm_thread_sub_trmm( thread ) );
thread_ibarrier( thread );
// Unpack C1 (if C1 was packed).
bli_unpackm_int( c1_pack, &c1,
cntl_sub_unpackm_c( cntl ),
trmm_thread_sub_ipackm( thread ) );
}
// If any packing buffers were acquired within packm, release them back
// to the memory manager.
thread_obarrier( thread );
if( thread_am_ochief( thread ) )
bli_packm_release( b_pack, cntl_sub_packm_b( cntl ) );
if( thread_am_ichief( thread ) ){
bli_packm_release( a1_pack, cntl_sub_packm_a( cntl ) );
bli_packm_release( c1_pack, cntl_sub_packm_c( cntl ) );
}
}
void bli_packm_blk_var1( obj_t* c,
obj_t* p,
packm_thrinfo_t* t )
{
num_t dt_cp = bli_obj_datatype( *c );
struc_t strucc = bli_obj_struc( *c );
doff_t diagoffc = bli_obj_diag_offset( *c );
diag_t diagc = bli_obj_diag( *c );
uplo_t uploc = bli_obj_uplo( *c );
trans_t transc = bli_obj_conjtrans_status( *c );
pack_t schema = bli_obj_pack_schema( *p );
bool_t invdiag = bli_obj_has_inverted_diag( *p );
bool_t revifup = bli_obj_is_pack_rev_if_upper( *p );
bool_t reviflo = bli_obj_is_pack_rev_if_lower( *p );
dim_t m_p = bli_obj_length( *p );
dim_t n_p = bli_obj_width( *p );
dim_t m_max_p = bli_obj_padded_length( *p );
dim_t n_max_p = bli_obj_padded_width( *p );
void* buf_c = bli_obj_buffer_at_off( *c );
inc_t rs_c = bli_obj_row_stride( *c );
inc_t cs_c = bli_obj_col_stride( *c );
void* buf_p = bli_obj_buffer_at_off( *p );
inc_t rs_p = bli_obj_row_stride( *p );
inc_t cs_p = bli_obj_col_stride( *p );
inc_t is_p = bli_obj_imag_stride( *p );
dim_t pd_p = bli_obj_panel_dim( *p );
inc_t ps_p = bli_obj_panel_stride( *p );
obj_t kappa;
/*---initialize pointer to stop gcc complaining 2-9-16 GH --- */
obj_t* kappa_p = {0};
void* buf_kappa;
func_t* packm_kers;
void* packm_ker;
FUNCPTR_T f;
// Treatment of kappa (ie: packing during scaling) depends on
// whether we are executing an induced method.
if ( bli_is_ind_packed( schema ) )
{
// 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 induced complex domain algorithms in terms of
// real domain micro-kernels. (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( thread_am_ochief( t ) )
{
if ( bli_obj_scalar_has_nonzero_imag( p ) )
{
// Detach the scalar.
bli_obj_scalar_detach( p, &kappa );
// Reset the attached scalar (to 1.0).
bli_obj_scalar_reset( p );
kappa_p = κ
}
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;
}
}
kappa_p = thread_obroadcast( t, kappa_p );
// Acquire the buffer to the kappa chosen above.
buf_kappa = bli_obj_buffer_for_1x1( dt_cp, *kappa_p );
}
else // if ( bli_is_nat_packed( schema ) )
{
// This branch if for native execution, where we assume that
// the micro-kernel will always apply the alpha scalar of the
// higher-level operation. Thus, we use BLIS_ONE for kappa so
// that the underlying packm implementation does not perform
// any scaling during packing.
buf_kappa = bli_obj_buffer_for_const( dt_cp, BLIS_ONE );
}
// Choose the correct func_t object based on the pack_t schema.
if ( bli_is_4mi_packed( schema ) ) packm_kers = packm_struc_cxk_4mi_kers;
else if ( bli_is_3mi_packed( schema ) ||
bli_is_3ms_packed( schema ) ) packm_kers = packm_struc_cxk_3mis_kers;
else if ( bli_is_ro_packed( schema ) ||
bli_is_io_packed( schema ) ||
bli_is_rpi_packed( schema ) ) packm_kers = packm_struc_cxk_rih_kers;
else packm_kers = packm_struc_cxk_kers;
// Query the datatype-specific function pointer from the func_t object.
packm_ker = bli_func_obj_query( dt_cp, packm_kers );
// Index into the type combination array to extract the correct
// function pointer.
f = ftypes[dt_cp];
// Invoke the function.
f( strucc,
diagoffc,
diagc,
uploc,
transc,
schema,
invdiag,
revifup,
reviflo,
m_p,
n_p,
m_max_p,
n_max_p,
buf_kappa,
buf_c, rs_c, cs_c,
buf_p, rs_p, cs_p,
is_p,
pd_p, ps_p,
packm_ker,
t );
}
void bli_trmm_blk_var2b( obj_t* a,
obj_t* b,
obj_t* c,
cntx_t* cntx,
gemm_t* cntl,
trmm_thrinfo_t* thread )
{
obj_t a_pack_s;
obj_t b1_pack_s, c1_pack_s;
obj_t b1, c1;
obj_t* a_pack = NULL;
obj_t* b1_pack = NULL;
obj_t* c1_pack = NULL;
dim_t i;
dim_t b_alg;
// Prune any zero region that exists along the partitioning dimension.
bli_trmm_prune_unref_mparts_n( a, b, c );
if( thread_am_ochief( thread ) ) {
// Initialize object for packing A
bli_obj_init_pack( &a_pack_s );
bli_packm_init( a, &a_pack_s,
cntx, cntl_sub_packm_a( cntl ) );
// Scale C by beta (if instructed).
bli_scalm_int( &BLIS_ONE,
c,
cntx, cntl_sub_scalm( cntl ) );
}
a_pack = thread_obroadcast( thread, &a_pack_s );
// Initialize pack objects for B and C that are passed into packm_init().
if( thread_am_ichief( thread ) ) {
bli_obj_init_pack( &b1_pack_s );
bli_obj_init_pack( &c1_pack_s );
}
b1_pack = thread_ibroadcast( thread, &b1_pack_s );
c1_pack = thread_ibroadcast( thread, &c1_pack_s );
// Pack A (if instructed).
bli_packm_int( a, a_pack,
cntx, cntl_sub_packm_a( cntl ),
trmm_thread_sub_opackm( thread ) );
dim_t my_start, my_end;
bli_get_range_weighted_r2l( thread, b,
bli_cntx_get_bmult( cntl_bszid( cntl ), cntx ),
&my_start, &my_end );
// Partition along the n dimension.
for ( i = my_start; i < my_end; i += b_alg )
{
// Determine the current algorithmic blocksize.
b_alg = bli_determine_blocksize_b( i, my_end, b,
cntl_bszid( cntl ), cntx );
// Acquire partitions for B1 and C1.
bli_acquire_mpart_r2l( BLIS_SUBPART1,
i, b_alg, b, &b1 );
bli_acquire_mpart_r2l( BLIS_SUBPART1,
i, b_alg, c, &c1 );
// Initialize objects for packing A1 and B1.
if( thread_am_ichief( thread ) ) {
bli_packm_init( &b1, b1_pack,
cntx, cntl_sub_packm_b( cntl ) );
bli_packm_init( &c1, c1_pack,
cntx, cntl_sub_packm_c( cntl ) );
}
thread_ibarrier( thread );
// Pack B1 (if instructed).
bli_packm_int( &b1, b1_pack,
cntx, cntl_sub_packm_b( cntl ),
trmm_thread_sub_ipackm( thread ) );
// Pack C1 (if instructed).
bli_packm_int( &c1, c1_pack,
cntx, cntl_sub_packm_c( cntl ),
trmm_thread_sub_ipackm( thread ) );
// Perform trmm subproblem.
bli_trmm_int( &BLIS_ONE,
a_pack,
b1_pack,
&BLIS_ONE,
c1_pack,
cntx,
cntl_sub_gemm( cntl ),
trmm_thread_sub_trmm( thread ) );
thread_ibarrier( thread );
// Unpack C1 (if C1 was packed).
bli_unpackm_int( c1_pack, &c1,
cntx, cntl_sub_unpackm_c( cntl ),
trmm_thread_sub_ipackm( thread ) );
}
// If any packing buffers were acquired within packm, release them back
// to the memory manager.
thread_obarrier( thread );
if( thread_am_ochief( thread ) )
bli_packm_release( a_pack, cntl_sub_packm_a( cntl ) );
if( thread_am_ichief( thread ) ) {
bli_packm_release( b1_pack, cntl_sub_packm_b( cntl ) );
bli_packm_release( c1_pack, cntl_sub_packm_c( cntl ) );
}
}
void bli_gemm_blk_var4f( obj_t* a,
obj_t* b,
obj_t* c,
gemm_t* cntl,
gemm_thrinfo_t* thread )
{
extern packm_t* gemm3mh_packa_cntl_ro;
extern packm_t* gemm3mh_packa_cntl_io;
extern packm_t* gemm3mh_packa_cntl_rpi;
packm_t* packa_cntl_ro = gemm3mh_packa_cntl_ro;
packm_t* packa_cntl_io = gemm3mh_packa_cntl_io;
packm_t* packa_cntl_rpi = gemm3mh_packa_cntl_rpi;
//The s is for "lives on the stack"
obj_t b_pack_s;
obj_t a1_pack_s, c1_pack_s;
obj_t a1, c1;
obj_t* a1_pack = NULL;
obj_t* b_pack = NULL;
obj_t* c1_pack = NULL;
dim_t i;
dim_t b_alg;
dim_t m_trans;
if( thread_am_ochief( thread ) ) {
// Initialize object for packing B.
bli_obj_init_pack( &b_pack_s );
bli_packm_init( b, &b_pack_s,
cntl_sub_packm_b( cntl ) );
// Scale C by beta (if instructed).
// Since scalm doesn't support multithreading yet, must be done by chief thread (ew)
bli_scalm_int( &BLIS_ONE,
c,
cntl_sub_scalm( cntl ) );
}
b_pack = thread_obroadcast( thread, &b_pack_s );
// Initialize objects passed into bli_packm_init for A and C
if( thread_am_ichief( thread ) ) {
bli_obj_init_pack( &a1_pack_s );
bli_obj_init_pack( &c1_pack_s );
}
a1_pack = thread_ibroadcast( thread, &a1_pack_s );
c1_pack = thread_ibroadcast( thread, &c1_pack_s );
// Pack B (if instructed).
bli_packm_int( b, b_pack,
cntl_sub_packm_b( cntl ),
gemm_thread_sub_opackm( thread ) );
// Query dimension in partitioning direction.
m_trans = bli_obj_length_after_trans( *a );
dim_t start, end;
bli_get_range_t2b( thread, 0, m_trans,
bli_blksz_get_mult_for_obj( a, cntl_blocksize( cntl ) ),
&start, &end );
// Partition along the m dimension.
for ( i = start; i < end; i += b_alg )
{
// Determine the current algorithmic blocksize.
// NOTE: Use of a (for execution datatype) is intentional!
// This causes the right blocksize to be used if c and a are
// complex and b is real.
b_alg = bli_determine_blocksize_f( i, end, 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.
if( thread_am_ichief( thread ) ) {
bli_packm_init( &a1, a1_pack,
packa_cntl_ro );
bli_packm_init( &c1, c1_pack,
cntl_sub_packm_c( cntl ) );
}
thread_ibarrier( thread );
// Pack A1 (if instructed).
bli_packm_int( &a1, a1_pack,
packa_cntl_ro,
gemm_thread_sub_ipackm( thread ) );
// Pack C1 (if instructed).
bli_packm_int( &c1, c1_pack,
cntl_sub_packm_c( cntl ),
gemm_thread_sub_ipackm( thread ) );
// Perform gemm subproblem.
bli_gemm_int( &BLIS_ONE,
a1_pack,
b_pack,
&BLIS_ONE,
c1_pack,
cntl_sub_gemm( cntl ),
gemm_thread_sub_gemm( thread ) );
thread_ibarrier( thread );
// Only apply beta within the first of three subproblems.
if ( thread_am_ichief( thread ) ) bli_obj_scalar_reset( c1_pack );
// Initialize objects for packing A1 and C1.
if( thread_am_ichief( thread ) ) {
bli_packm_init( &a1, a1_pack,
packa_cntl_io );
}
thread_ibarrier( thread );
// Pack A1 (if instructed).
bli_packm_int( &a1, a1_pack,
packa_cntl_io,
gemm_thread_sub_ipackm( thread ) );
// Perform gemm subproblem.
bli_gemm_int( &BLIS_ONE,
a1_pack,
b_pack,
&BLIS_ONE,
c1_pack,
cntl_sub_gemm( cntl ),
gemm_thread_sub_gemm( thread ) );
thread_ibarrier( thread );
// Initialize objects for packing A1 and C1.
if( thread_am_ichief( thread ) ) {
bli_packm_init( &a1, a1_pack,
packa_cntl_rpi );
}
thread_ibarrier( thread );
// Pack A1 (if instructed).
bli_packm_int( &a1, a1_pack,
packa_cntl_rpi,
gemm_thread_sub_ipackm( thread ) );
// Perform gemm subproblem.
bli_gemm_int( &BLIS_ONE,
a1_pack,
b_pack,
&BLIS_ONE,
c1_pack,
cntl_sub_gemm( cntl ),
gemm_thread_sub_gemm( thread ) );
thread_ibarrier( thread );
// Unpack C1 (if C1 was packed).
// Currently must be done by 1 thread
bli_unpackm_int( c1_pack, &c1,
cntl_sub_unpackm_c( cntl ),
gemm_thread_sub_ipackm( thread ) );
}
// If any packing buffers were acquired within packm, release them back
// to the memory manager.
thread_obarrier( thread );
if( thread_am_ochief( thread ) )
bli_packm_release( b_pack, cntl_sub_packm_b( cntl ) );
if( thread_am_ichief( thread ) ){
// It doesn't matter which packm cntl node we pass in, as long
// as it is valid, packm_release() will release the mem_t entry.
bli_packm_release( a1_pack, packa_cntl_ro );
bli_packm_release( c1_pack, cntl_sub_packm_c( cntl ) );
}
}
void bli_trmm_blk_var3b( obj_t* a,
obj_t* b,
obj_t* c,
gemm_t* cntl,
trmm_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 ),
trmm_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 trmm-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_trmm_determine_kc_b( i, k_trans, a, b,
cntl_blocksize( cntl ) );
// Acquire partitions for A1 and B1.
bli_acquire_mpart_r2l( BLIS_SUBPART1,
i, b_alg, a, &a1 );
bli_acquire_mpart_b2t( 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 ),
trmm_thread_sub_ipackm( thread ) );
// Pack B1 (if instructed).
bli_packm_int( &b1, b1_pack,
cntl_sub_packm_b( cntl ),
trmm_thread_sub_ipackm( thread ) );
// Perform trmm subproblem.
bli_trmm_int( &BLIS_ONE,
a1_pack,
b1_pack,
&BLIS_ONE,
c_pack,
cntl_sub_gemm( cntl ),
trmm_thread_sub_trmm( thread ) );
thread_ibarrier( thread );
}
thread_obarrier( thread );
// Unpack C (if C was packed).
bli_unpackm_int( c_pack, c,
cntl_sub_unpackm_c( cntl ),
trmm_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 ) );
}
}
void bli_herk_blk_var2f( obj_t* a,
obj_t* ah,
obj_t* c,
gemm_t* cntl,
herk_thrinfo_t* thread )
{
obj_t a_pack_s;
obj_t ah1_pack_s, c1S_pack_s;
obj_t ah1, c1, c1S;
obj_t aS_pack;
obj_t* a_pack;
obj_t* ah1_pack;
obj_t* c1S_pack;
dim_t i;
dim_t b_alg;
dim_t n_trans;
subpart_t stored_part;
// The upper and lower variants are identical, except for which
// merged subpartition is acquired in the loop body.
if ( bli_obj_is_lower( *c ) ) stored_part = BLIS_SUBPART1B;
else stored_part = BLIS_SUBPART1T;
if( thread_am_ochief( thread ) ) {
// Initialize object for packing A
bli_obj_init_pack( &a_pack_s );
bli_packm_init( a, &a_pack_s,
cntl_sub_packm_a( cntl ) );
// Scale C by beta (if instructed).
bli_scalm_int( &BLIS_ONE,
c,
cntl_sub_scalm( cntl ) );
}
a_pack = thread_obroadcast( thread, &a_pack_s );
// Initialize pack objects for C and A' that are passed into packm_init().
if( thread_am_ichief( thread ) ) {
bli_obj_init_pack( &ah1_pack_s );
bli_obj_init_pack( &c1S_pack_s );
}
ah1_pack = thread_ibroadcast( thread, &ah1_pack_s );
c1S_pack = thread_ibroadcast( thread, &c1S_pack_s );
// Pack A (if instructed).
bli_packm_int( a, a_pack,
cntl_sub_packm_a( cntl ),
herk_thread_sub_opackm( thread ) );
// Query dimension in partitioning direction.
n_trans = bli_obj_width_after_trans( *c );
dim_t start, end;
// Needs to be replaced with a weighted range because triangle
bli_get_range_weighted( thread, 0, n_trans,
bli_blksz_get_mult_for_obj( a, cntl_blocksize( cntl ) ),
bli_obj_is_lower( *c ), &start, &end );
// Partition along the n dimension.
for ( i = start; i < end; i += b_alg )
{
// Determine the current algorithmic blocksize.
b_alg = bli_determine_blocksize_f( i, end, a,
cntl_blocksize( cntl ) );
// Acquire partitions for A1' and C1.
bli_acquire_mpart_l2r( BLIS_SUBPART1,
i, b_alg, ah, &ah1 );
bli_acquire_mpart_l2r( BLIS_SUBPART1,
i, b_alg, c, &c1 );
// Partition off the stored region of C1 and the corresponding region
// of A_pack.
bli_acquire_mpart_t2b( stored_part,
i, b_alg, &c1, &c1S );
bli_acquire_mpart_t2b( stored_part,
i, b_alg, a_pack, &aS_pack );
// Initialize objects for packing A1' and C1.
if( thread_am_ichief( thread ) ) {
bli_packm_init( &ah1, ah1_pack,
cntl_sub_packm_b( cntl ) );
bli_packm_init( &c1S, c1S_pack,
cntl_sub_packm_c( cntl ) );
}
thread_ibarrier( thread ) ;
// Pack A1' (if instructed).
bli_packm_int( &ah1, ah1_pack,
cntl_sub_packm_b( cntl ),
herk_thread_sub_ipackm( thread ) );
// Pack C1 (if instructed).
bli_packm_int( &c1S, c1S_pack,
cntl_sub_packm_c( cntl ),
herk_thread_sub_ipackm( thread ) ) ;
// Perform herk subproblem.
bli_herk_int( &BLIS_ONE,
&aS_pack,
ah1_pack,
&BLIS_ONE,
c1S_pack,
cntl_sub_gemm( cntl ),
herk_thread_sub_herk( thread ) );
thread_ibarrier( thread );
// Unpack C1 (if C1 was packed).
bli_unpackm_int( c1S_pack, &c1S,
cntl_sub_unpackm_c( cntl ),
herk_thread_sub_ipackm( thread ) );
}
// If any packing buffers were acquired within packm, release them back
// to the memory manager.
thread_obarrier( thread );
if( thread_am_ochief( thread ) )
bli_packm_release( a_pack, cntl_sub_packm_a( cntl ) );
if( thread_am_ichief( thread ) ) {
bli_packm_release( ah1_pack, cntl_sub_packm_b( cntl ) );
bli_packm_release( c1S_pack, cntl_sub_packm_c( cntl ) );
}
}
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 ) );
}
}