void bli_packm_unb_var1( obj_t* c,
obj_t* p,
packm_thrinfo_t* thread )
{
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 );
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 );
void* buf_kappa;
FUNCPTR_T f;
// This variant assumes that the computational 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 scale
// during packing.
buf_kappa = bli_obj_buffer_for_const( dt_cp, BLIS_ONE );
// Index into the type combination array to extract the correct
// function pointer.
f = ftypes[dt_cp];
if( thread_am_ochief( thread ) ) {
// Invoke the function.
f( strucc,
diagoffc,
diagc,
uploc,
transc,
m_p,
n_p,
m_max_p,
n_max_p,
buf_kappa,
buf_c, rs_c, cs_c,
buf_p, rs_p, cs_p );
}
}
void bli_obj_print( char* label, obj_t* obj )
{
FILE* file = stdout;
if ( bli_error_checking_is_enabled() )
bli_obj_print_check( label, obj );
fprintf( file, "\n" );
fprintf( file, "%s\n", label );
fprintf( file, "\n" );
fprintf( file, " m x n %lu x %lu\n", ( unsigned long int )bli_obj_length( *obj ),
( unsigned long int )bli_obj_width( *obj ) );
fprintf( file, "\n" );
fprintf( file, " offm, offn %lu, %lu\n", ( unsigned long int )bli_obj_row_off( *obj ),
( unsigned long int )bli_obj_col_off( *obj ) );
fprintf( file, " diagoff %ld\n", ( signed long int )bli_obj_diag_offset( *obj ) );
fprintf( file, "\n" );
fprintf( file, " buf %p\n", ( void* )bli_obj_buffer( *obj ) );
fprintf( file, " elem size %lu\n", ( unsigned long int )bli_obj_elem_size( *obj ) );
fprintf( file, " rs, cs %ld, %ld\n", ( signed long int )bli_obj_row_stride( *obj ),
( signed long int )bli_obj_col_stride( *obj ) );
fprintf( file, " is %ld\n", ( signed long int )bli_obj_imag_stride( *obj ) );
fprintf( file, " m_padded %lu\n", ( unsigned long int )bli_obj_padded_length( *obj ) );
fprintf( file, " n_padded %lu\n", ( unsigned long int )bli_obj_padded_width( *obj ) );
fprintf( file, " ps %lu\n", ( unsigned long int )bli_obj_panel_stride( *obj ) );
fprintf( file, "\n" );
fprintf( file, " info %lX\n", ( unsigned long int )(*obj).info );
fprintf( file, " - is complex %lu\n", ( unsigned long int )bli_obj_is_complex( *obj ) );
fprintf( file, " - is d. prec %lu\n", ( unsigned long int )bli_obj_is_double_precision( *obj ) );
fprintf( file, " - datatype %lu\n", ( unsigned long int )bli_obj_datatype( *obj ) );
fprintf( file, " - target dt %lu\n", ( unsigned long int )bli_obj_target_datatype( *obj ) );
fprintf( file, " - exec dt %lu\n", ( unsigned long int )bli_obj_execution_datatype( *obj ) );
fprintf( file, " - has trans %lu\n", ( unsigned long int )bli_obj_has_trans( *obj ) );
fprintf( file, " - has conj %lu\n", ( unsigned long int )bli_obj_has_conj( *obj ) );
fprintf( file, " - unit diag? %lu\n", ( unsigned long int )bli_obj_has_unit_diag( *obj ) );
fprintf( file, " - struc type %lu\n", ( unsigned long int )bli_obj_struc( *obj ) >> BLIS_STRUC_SHIFT );
fprintf( file, " - uplo type %lu\n", ( unsigned long int )bli_obj_uplo( *obj ) >> BLIS_UPLO_SHIFT );
fprintf( file, " - is upper %lu\n", ( unsigned long int )bli_obj_is_upper( *obj ) );
fprintf( file, " - is lower %lu\n", ( unsigned long int )bli_obj_is_lower( *obj ) );
fprintf( file, " - is dense %lu\n", ( unsigned long int )bli_obj_is_dense( *obj ) );
fprintf( file, " - pack schema %lu\n", ( unsigned long int )bli_obj_pack_schema( *obj ) >> BLIS_PACK_SCHEMA_SHIFT );
fprintf( file, " - packinv diag? %lu\n", ( unsigned long int )bli_obj_has_inverted_diag( *obj ) );
fprintf( file, " - pack ordifup %lu\n", ( unsigned long int )bli_obj_is_pack_rev_if_upper( *obj ) );
fprintf( file, " - pack ordiflo %lu\n", ( unsigned long int )bli_obj_is_pack_rev_if_lower( *obj ) );
fprintf( file, " - packbuf type %lu\n", ( unsigned long int )bli_obj_pack_buffer_type( *obj ) >> BLIS_PACK_BUFFER_SHIFT );
fprintf( file, "\n" );
}
siz_t bli_packv_init_pack
(
pack_t schema,
bszid_t bmult_id,
obj_t* a,
obj_t* p,
cntx_t* cntx
)
{
num_t dt = bli_obj_dt( a );
dim_t dim_a = bli_obj_vector_dim( a );
dim_t bmult = bli_cntx_get_blksz_def_dt( dt, bmult_id, cntx );
membrk_t* membrk = bli_cntx_membrk( cntx );
#if 0
mem_t* mem_p;
#endif
dim_t m_p_pad;
siz_t size_p;
inc_t rs_p, cs_p;
void* buf;
// We begin by copying the basic fields of c.
bli_obj_alias_to( a, p );
// Update the dimensions.
bli_obj_set_dims( dim_a, 1, p );
// Reset the view offsets to (0,0).
bli_obj_set_offs( 0, 0, p );
// Set the pack schema in the p object to the value in the control tree
// node.
bli_obj_set_pack_schema( schema, p );
// Compute the dimensions padded by the dimension multiples.
m_p_pad = bli_align_dim_to_mult( bli_obj_vector_dim( p ), bmult );
// Compute the size of the packed buffer.
size_p = m_p_pad * 1 * bli_obj_elem_size( p );
#if 0
// Extract the address of the mem_t object within p that will track
// properties of the packed buffer.
mem_p = bli_obj_pack_mem( *p );
if ( bli_mem_is_unalloc( mem_p ) )
{
// If the mem_t object of p has not yet been allocated, then acquire
// a memory block suitable for a vector.
bli_membrk_acquire_v( membrk,
size_p,
mem_p );
}
else
{
// If the mem_t object has already been allocated, then release and
// re-acquire the memory so there is sufficient space.
if ( bli_mem_size( mem_p ) < size_p )
{
bli_membrk_release( mem_p );
bli_membrk_acquire_v( membrk,
size_p,
mem_p );
}
}
// Grab the buffer address from the mem_t object and copy it to the
// main object buffer field. (Sometimes this buffer address will be
// copied when the value is already up-to-date, because it persists
// in the main object buffer field across loop iterations.)
buf = bli_mem_buffer( mem_p );
bli_obj_set_buffer( buf, p );
#endif
// Save the padded (packed) dimensions into the packed object.
bli_obj_set_padded_dims( m_p_pad, 1, p );
// Set the row and column strides of p based on the pack schema.
if ( schema == BLIS_PACKED_VECTOR )
{
// Set the strides to reflect a column-stored vector. Note that the
// column stride may never be used, and is only useful to determine
// how much space beyond the vector would need to be zero-padded, if
// zero-padding was needed.
rs_p = 1;
cs_p = bli_obj_padded_length( p );
bli_obj_set_strides( rs_p, cs_p, p );
}
return size_p;
}
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_packm_acquire_mpart_t2b( subpart_t requested_part,
dim_t i,
dim_t b,
obj_t* obj,
obj_t* sub_obj )
{
dim_t m, n;
// For now, we only support acquiring the middle subpartition.
if ( requested_part != BLIS_SUBPART1 )
{
bli_check_error_code( BLIS_NOT_YET_IMPLEMENTED );
}
// Partitioning top-to-bottom through packed column panels (which are
// row-stored) is not yet supported.
if ( bli_obj_is_col_packed( *obj ) )
{
bli_check_error_code( BLIS_NOT_YET_IMPLEMENTED );
}
// Query the dimensions of the parent object.
m = bli_obj_length( *obj );
n = bli_obj_width( *obj );
// Foolproofing: do not let b exceed what's left of the m dimension at
// row offset i.
if ( b > m - i ) b = m - i;
// Begin by copying the info, elem size, buffer, row stride, and column
// stride fields of the parent object. Note that this omits copying view
// information because the new partition will have its own dimensions
// and offsets.
bli_obj_init_subpart_from( *obj, *sub_obj );
// Modify offsets and dimensions of requested partition.
bli_obj_set_dims( b, n, *sub_obj );
// Tweak the padded length of the subpartition to trick the underlying
// implementation into only zero-padding for the narrow submatrix of
// interest. Usually, the value we want is b (for non-edge cases), but
// at the edges, we want the remainder of the mem_t region in the m
// dimension. Edge cases are defined as occurring when i + b is exactly
// equal to the inherited sub-object's length (which happens since the
// determine_blocksize function would have returned a smaller value of
// b for the edge iteration). In these cases, we arrive at the new
// packed length by simply subtracting off i.
{
dim_t m_pack_max = bli_obj_padded_length( *sub_obj );
dim_t m_pack_cur;
if ( i + b == m ) m_pack_cur = m_pack_max - i;
else m_pack_cur = b;
bli_obj_set_padded_length( m_pack_cur, *sub_obj );
}
// Translate the desired offsets to a panel offset and adjust the
// buffer pointer of the subpartition object.
{
char* buf_p = bli_obj_buffer( *sub_obj );
siz_t elem_size = bli_obj_elem_size( *sub_obj );
dim_t off_to_panel = bli_packm_offset_to_panel_for( i, sub_obj );
buf_p = buf_p + elem_size * off_to_panel;
bli_obj_set_buffer( ( void* )buf_p, *sub_obj );
}
}
void bli_packm_blk_var1_md
(
obj_t* c,
obj_t* p,
cntx_t* cntx,
cntl_t* cntl,
thrinfo_t* t
)
{
num_t dt_c = bli_obj_dt( c );
num_t dt_p = bli_obj_dt( p );
trans_t transc = bli_obj_conjtrans_status( c );
pack_t schema = bli_obj_pack_schema( 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;
void* buf_kappa;
FUNCPTR_T f;
// Treatment of kappa (ie: packing during scaling) depends on
// whether we are executing an induced method.
if ( bli_is_nat_packed( schema ) )
{
// This branch is 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_p, &BLIS_ONE );
}
else // if ( bli_is_ind_packed( schema ) )
{
obj_t* kappa_p;
// 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 ( 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;
}
// Acquire the buffer to the kappa chosen above.
buf_kappa = bli_obj_buffer_for_1x1( dt_p, kappa_p );
}
// Index into the type combination array to extract the correct
// function pointer.
f = ftypes[dt_c][dt_p];
// Invoke the function.
f(
transc,
schema,
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,
cntx,
t );
}
