void bli_unpackm_blk_var2( obj_t* p,
obj_t* c,
unpackm_t* cntl )
{
num_t dt_cp = bli_obj_datatype( *c );
// Normally we take the parameters from the source argument. But here,
// the packm/unpackm framework is not yet solidified enough for us to
// assume that at this point struc(P) == struc(C), (ie: since
// densification may have marked P's structure as dense when the root
// is upper or lower). So, we take the struc field from C, not P.
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 );
// Again, normally the trans argument is on the source matrix. But we
// know that the packed matrix is not transposed. If there is to be a
// transposition, it is because C was originally transposed when packed.
// Thus, we query C for the trans status, not P. Also, we only query
// the trans status (not the conjugation status), since we probably
// don't want to un-conjugate if the original matrix was conjugated
// when packed.
trans_t transc = bli_obj_onlytrans_status( *c );
dim_t m_c = bli_obj_length( *c );
dim_t n_c = bli_obj_width( *c );
dim_t m_panel = bli_obj_panel_length( *c );
dim_t n_panel = bli_obj_panel_width( *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 );
dim_t pd_p = bli_obj_panel_dim( *p );
inc_t ps_p = bli_obj_panel_stride( *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 );
FUNCPTR_T f;
// 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,
m_c,
n_c,
m_panel,
n_panel,
buf_p, rs_p, cs_p,
pd_p, ps_p,
buf_c, rs_c, cs_c );
}
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 );
}
}
dim_t bli_packm_offset_to_panel_for( dim_t offmn, obj_t* p )
{
dim_t panel_off;
if ( bli_obj_pack_schema( *p ) == BLIS_PACKED_ROWS )
{
// For the "packed rows" schema, a single row is effectively one
// row panel, and so we use the row offset as the panel offset.
// Then we multiply this offset by the effective panel stride
// (ie: the row stride) to arrive at the desired offset.
panel_off = offmn * bli_obj_row_stride( *p );
}
else if ( bli_obj_pack_schema( *p ) == BLIS_PACKED_COLUMNS )
{
// For the "packed columns" schema, a single column is effectively one
// column panel, and so we use the column offset as the panel offset.
// Then we multiply this offset by the effective panel stride
// (ie: the column stride) to arrive at the desired offset.
panel_off = offmn * bli_obj_col_stride( *p );
}
else if ( bli_obj_pack_schema( *p ) == BLIS_PACKED_ROW_PANELS )
{
// For the "packed row panels" schema, the column stride is equal to
// the panel dimension (length). So we can divide it into offmn
// (interpreted as a row offset) to arrive at a panel offset. Then
// we multiply this offset by the panel stride to arrive at the total
// offset to the panel (in units of elements).
panel_off = offmn / bli_obj_col_stride( *p );
panel_off = panel_off * bli_obj_panel_stride( *p );
// Sanity check.
if ( offmn % bli_obj_col_stride( *p ) > 0 ) bli_abort();
}
else if ( bli_obj_pack_schema( *p ) == BLIS_PACKED_COL_PANELS )
{
// For the "packed column panels" schema, the row stride is equal to
// the panel dimension (width). So we can divide it into offmn
// (interpreted as a column offset) to arrive at a panel offset. Then
// we multiply this offset by the panel stride to arrive at the total
// offset to the panel (in units of elements).
panel_off = offmn / bli_obj_row_stride( *p );
panel_off = panel_off * bli_obj_panel_stride( *p );
// Sanity check.
if ( offmn % bli_obj_row_stride( *p ) > 0 ) bli_abort();
}
else
{
panel_off = 0;
bli_check_error_code( BLIS_NOT_YET_IMPLEMENTED );
}
return panel_off;
}
void bli_trsm_rl_ker_var2( obj_t* alpha,
obj_t* a,
obj_t* b,
obj_t* beta,
obj_t* c,
trsm_t* cntl )
{
num_t dt_exec = bli_obj_execution_datatype( *c );
doff_t diagoffb = bli_obj_diag_offset( *b );
dim_t m = bli_obj_length( *c );
dim_t n = bli_obj_width( *c );
dim_t k = bli_obj_width( *a );
void* buf_a = bli_obj_buffer_at_off( *a );
inc_t rs_a = bli_obj_row_stride( *a );
inc_t cs_a = bli_obj_col_stride( *a );
inc_t ps_a = bli_obj_panel_stride( *a );
void* buf_b = bli_obj_buffer_at_off( *b );
inc_t rs_b = bli_obj_row_stride( *b );
inc_t cs_b = bli_obj_col_stride( *b );
inc_t ps_b = bli_obj_panel_stride( *b );
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 );
num_t dt_alpha;
void* buf_alpha;
FUNCPTR_T f;
// If alpha is a scalar constant, use dt_exec to extract the address of the
// corresponding constant value; otherwise, use the datatype encoded
// within the alpha object and extract the buffer at the alpha offset.
bli_set_scalar_dt_buffer( alpha, dt_exec, dt_alpha, buf_alpha );
// Index into the type combination array to extract the correct
// function pointer.
f = ftypes[dt_exec];
// Invoke the function.
f( diagoffb,
m,
n,
k,
buf_alpha,
buf_a, rs_a, cs_a, ps_a,
buf_b, rs_b, cs_b, ps_b,
buf_c, rs_c, cs_c );
}
void bli_axpym_unb_var1( obj_t* alpha,
obj_t* x,
obj_t* y,
cntx_t* cntx )
{
num_t dt_x = bli_obj_datatype( *x );
num_t dt_y = bli_obj_datatype( *y );
doff_t diagoffx = bli_obj_diag_offset( *x );
diag_t diagx = bli_obj_diag( *x );
uplo_t uplox = bli_obj_uplo( *x );
trans_t transx = bli_obj_conjtrans_status( *x );
dim_t m = bli_obj_length( *y );
dim_t n = bli_obj_width( *y );
inc_t rs_x = bli_obj_row_stride( *x );
inc_t cs_x = bli_obj_col_stride( *x );
void* buf_x = bli_obj_buffer_at_off( *x );
inc_t rs_y = bli_obj_row_stride( *y );
inc_t cs_y = bli_obj_col_stride( *y );
void* buf_y = bli_obj_buffer_at_off( *y );
num_t dt_alpha;
void* buf_alpha;
FUNCPTR_T f;
// If alpha is a scalar constant, use dt_x to extract the address of the
// corresponding constant value; otherwise, use the datatype encoded
// within the alpha object and extract the buffer at the alpha offset.
bli_set_scalar_dt_buffer( alpha, dt_x, dt_alpha, buf_alpha );
// Index into the type combination array to extract the correct
// function pointer.
f = ftypes[dt_alpha][dt_x][dt_y];
// Invoke the function.
f( diagoffx,
diagx,
uplox,
transx,
m,
n,
buf_alpha,
buf_x, rs_x, cs_x,
buf_y, rs_y, cs_y );
}
void bli_setid_unb_var1( obj_t* beta,
obj_t* x )
{
num_t dt_xr = bli_obj_datatype_proj_to_real( *x );
num_t dt_x = bli_obj_datatype( *x );
doff_t diagoffx = bli_obj_diag_offset( *x );
dim_t m = bli_obj_length( *x );
dim_t n = bli_obj_width( *x );
void* buf_beta = bli_obj_buffer_for_1x1( dt_xr, *beta );
void* buf_x = bli_obj_buffer_at_off( *x );
inc_t rs_x = bli_obj_row_stride( *x );
inc_t cs_x = bli_obj_col_stride( *x );
FUNCPTR_T f;
// Index into the type combination array to extract the correct
// function pointer.
f = ftypes[dt_x];
// Invoke the function.
f( diagoffx,
m,
n,
buf_beta,
buf_x, rs_x, cs_x );
}
void bli_norm1m_unb_var1( obj_t* x,
obj_t* norm )
{
num_t dt_x = bli_obj_datatype( *x );
doff_t diagoffx = bli_obj_diag_offset( *x );
uplo_t diagx = bli_obj_diag( *x );
uplo_t uplox = bli_obj_uplo( *x );
dim_t m = bli_obj_length( *x );
dim_t n = bli_obj_width( *x );
void* buf_x = bli_obj_buffer_at_off( *x );
inc_t rs_x = bli_obj_row_stride( *x );
inc_t cs_x = bli_obj_col_stride( *x );
void* buf_norm = bli_obj_buffer_at_off( *norm );
FUNCPTR_T f;
// Index into the type combination array to extract the correct
// function pointer.
f = ftypes[dt_x];
// Invoke the function.
f( diagoffx,
diagx,
uplox,
m,
n,
buf_x, rs_x, cs_x,
buf_norm );
}
void bli_gemv_unb_var2( obj_t* alpha,
obj_t* a,
obj_t* x,
obj_t* beta,
obj_t* y,
gemv_t* cntl )
{
num_t dt_a = bli_obj_datatype( *a );
num_t dt_x = bli_obj_datatype( *x );
num_t dt_y = bli_obj_datatype( *y );
conj_t transa = bli_obj_conjtrans_status( *a );
conj_t conjx = bli_obj_conj_status( *x );
dim_t m = bli_obj_length( *a );
dim_t n = bli_obj_width( *a );
void* buf_a = bli_obj_buffer_at_off( *a );
inc_t rs_a = bli_obj_row_stride( *a );
inc_t cs_a = bli_obj_col_stride( *a );
void* buf_x = bli_obj_buffer_at_off( *x );
inc_t incx = bli_obj_vector_inc( *x );
void* buf_y = bli_obj_buffer_at_off( *y );
inc_t incy = bli_obj_vector_inc( *y );
num_t dt_alpha;
void* buf_alpha;
num_t dt_beta;
void* buf_beta;
FUNCPTR_T f;
// The datatype of alpha MUST be the type union of a and x. This is to
// prevent any unnecessary loss of information during computation.
dt_alpha = bli_datatype_union( dt_a, dt_x );
buf_alpha = bli_obj_scalar_buffer( dt_alpha, *alpha );
// The datatype of beta MUST be the same as the datatype of y.
dt_beta = dt_y;
buf_beta = bli_obj_scalar_buffer( dt_beta, *beta );
// Index into the type combination array to extract the correct
// function pointer.
f = ftypes[dt_a][dt_x][dt_y];
// Invoke the function.
f( transa,
conjx,
m,
n,
buf_alpha,
buf_a, rs_a, cs_a,
buf_x, incx,
buf_beta,
buf_y, incy );
}
void bli_gemmtrsm_ukr( obj_t* alpha,
obj_t* a1x,
obj_t* a11,
obj_t* bx1,
obj_t* b11,
obj_t* c11 )
{
dim_t k = bli_obj_width( *a1x );
num_t dt = bli_obj_datatype( *c11 );
void* buf_a1x = bli_obj_buffer_at_off( *a1x );
void* buf_a11 = bli_obj_buffer_at_off( *a11 );
void* buf_bx1 = bli_obj_buffer_at_off( *bx1 );
void* buf_b11 = bli_obj_buffer_at_off( *b11 );
void* buf_c11 = bli_obj_buffer_at_off( *c11 );
inc_t rs_c = bli_obj_row_stride( *c11 );
inc_t cs_c = bli_obj_col_stride( *c11 );
void* buf_alpha = bli_obj_buffer_for_1x1( dt, *alpha );
inc_t ps_a = bli_obj_panel_stride( *a1x );
inc_t ps_b = bli_obj_panel_stride( *bx1 );
FUNCPTR_T f;
auxinfo_t data;
// Fill the auxinfo_t struct in case the micro-kernel uses it.
if ( bli_obj_is_lower( *a11 ) )
{ bli_auxinfo_set_next_a( buf_a1x, data ); }
else
{ bli_auxinfo_set_next_a( buf_a11, data ); }
bli_auxinfo_set_next_b( buf_bx1, data );
bli_auxinfo_set_ps_a( ps_a, data );
bli_auxinfo_set_ps_b( ps_b, data );
// Index into the type combination array to extract the correct
// function pointer.
if ( bli_obj_is_lower( *a11 ) ) f = ftypes_l[dt];
else f = ftypes_u[dt];
// Invoke the function.
f( k,
buf_alpha,
buf_a1x,
buf_a11,
buf_bx1,
buf_b11,
buf_c11, rs_c, cs_c,
&data );
}
void bli_dotxf_kernel( obj_t* alpha,
obj_t* a,
obj_t* x,
obj_t* beta,
obj_t* y )
{
num_t dt_a = bli_obj_datatype( *a );
num_t dt_x = bli_obj_datatype( *x );
num_t dt_y = bli_obj_datatype( *y );
conj_t conjat = bli_obj_conj_status( *a );
conj_t conjx = bli_obj_conj_status( *x );
dim_t m = bli_obj_vector_dim( *x );
dim_t b_n = bli_obj_vector_dim( *y );
void* buf_a = bli_obj_buffer_at_off( *a );
inc_t rs_a = bli_obj_row_stride( *a );
inc_t cs_a = bli_obj_col_stride( *a );
inc_t inc_x = bli_obj_vector_inc( *x );
void* buf_x = bli_obj_buffer_at_off( *x );
inc_t inc_y = bli_obj_vector_inc( *y );
void* buf_y = bli_obj_buffer_at_off( *y );
num_t dt_alpha;
void* buf_alpha;
num_t dt_beta;
void* buf_beta;
FUNCPTR_T f;
// The datatype of alpha MUST be the type union of a and x. This is to
// prevent any unnecessary loss of information during computation.
dt_alpha = bli_datatype_union( dt_a, dt_x );
buf_alpha = bli_obj_buffer_for_1x1( dt_alpha, *alpha );
// The datatype of beta MUST be the same as the datatype of y.
dt_beta = dt_y;
buf_beta = bli_obj_buffer_for_1x1( dt_beta, *beta );
// Index into the type combination array to extract the correct
// function pointer.
f = ftypes[dt_a][dt_x][dt_y];
// Invoke the function.
f( conjat,
conjx,
m,
b_n,
buf_alpha,
buf_a, rs_a, cs_a,
buf_x, inc_x,
buf_beta,
buf_y, inc_y );
}
void bli_unpackm_unb_var1
(
obj_t* p,
obj_t* c,
cntx_t* cntx,
cntl_t* cntl,
thrinfo_t* thread
)
{
num_t dt_pc = bli_obj_datatype( *p );
doff_t diagoffp = bli_obj_diag_offset( *p );
uplo_t uplop = bli_obj_uplo( *p );
trans_t transc = bli_obj_onlytrans_status( *c );
dim_t m_c = bli_obj_length( *c );
dim_t n_c = bli_obj_width( *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_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 );
FUNCPTR_T f;
// Index into the type combination array to extract the correct
// function pointer.
f = ftypes[dt_pc];
// Invoke the function.
f( diagoffp,
uplop,
transc,
m_c,
n_c,
buf_p, rs_p, cs_p,
buf_c, rs_c, cs_c,
cntx
);
}
void bli_scalm_unb_var1( obj_t* alpha,
obj_t* x,
cntx_t* cntx )
{
num_t dt_x = bli_obj_datatype( *x );
doff_t diagoffx = bli_obj_diag_offset( *x );
uplo_t diagx = bli_obj_diag( *x );
uplo_t uplox = bli_obj_uplo( *x );
dim_t m = bli_obj_length( *x );
dim_t n = bli_obj_width( *x );
void* buf_x = bli_obj_buffer_at_off( *x );
inc_t rs_x = bli_obj_row_stride( *x );
inc_t cs_x = bli_obj_col_stride( *x );
void* buf_alpha;
obj_t x_local;
FUNCPTR_T f;
// Alias x to x_local so we can apply alpha if it is non-unit.
bli_obj_alias_to( *x, x_local );
// If alpha is non-unit, apply it to the scalar attached to x.
if ( !bli_obj_equals( alpha, &BLIS_ONE ) )
{
bli_obj_scalar_apply_scalar( alpha, &x_local );
}
// Grab the address of the internal scalar buffer for the scalar
// attached to x.
buf_alpha_x = bli_obj_internal_scalar_buffer( *x );
// Index into the type combination array to extract the correct
// function pointer.
// NOTE: We use dt_x for both alpha and x because alpha was obtained
// from the attached scalar of x, which is guaranteed to be of the
// same datatype as x.
f = ftypes[dt_x][dt_x];
// Invoke the function.
// NOTE: We unconditionally pass in BLIS_NO_CONJUGATE for alpha
// because it would have already been conjugated by the front-end.
f( BLIS_NO_CONJUGATE,
diagoffx,
diagx,
uplox,
m,
n,
buf_alpha,
buf_x, rs_x, cs_x );
}
void bli_ger_unb_var2( obj_t* alpha,
obj_t* x,
obj_t* y,
obj_t* a,
cntx_t* cntx,
ger_t* cntl )
{
num_t dt_x = bli_obj_datatype( *x );
num_t dt_y = bli_obj_datatype( *y );
num_t dt_a = bli_obj_datatype( *a );
conj_t conjx = bli_obj_conj_status( *x );
conj_t conjy = bli_obj_conj_status( *y );
dim_t m = bli_obj_length( *a );
dim_t n = bli_obj_width( *a );
void* buf_x = bli_obj_buffer_at_off( *x );
inc_t incx = bli_obj_vector_inc( *x );
void* buf_y = bli_obj_buffer_at_off( *y );
inc_t incy = bli_obj_vector_inc( *y );
void* buf_a = bli_obj_buffer_at_off( *a );
inc_t rs_a = bli_obj_row_stride( *a );
inc_t cs_a = bli_obj_col_stride( *a );
num_t dt_alpha;
void* buf_alpha;
FUNCPTR_T f;
// The datatype of alpha MUST be the type union of x and y. This is to
// prevent any unnecessary loss of information during computation.
dt_alpha = bli_datatype_union( dt_x, dt_y );
buf_alpha = bli_obj_buffer_for_1x1( dt_alpha, *alpha );
// Index into the type combination array to extract the correct
// function pointer.
f = ftypes[dt_a];
// Invoke the function.
f( conjx,
conjy,
m,
n,
buf_alpha,
buf_x, incx,
buf_y, incy,
buf_a, rs_a, cs_a,
cntx );
}
void bli_her2_unb_var1( conj_t conjh,
obj_t* alpha,
obj_t* alpha_conj,
obj_t* x,
obj_t* y,
obj_t* c,
her2_t* cntl )
{
num_t dt_x = bli_obj_datatype( *x );
num_t dt_y = bli_obj_datatype( *y );
num_t dt_c = bli_obj_datatype( *c );
uplo_t uplo = bli_obj_uplo( *c );
conj_t conjx = bli_obj_conj_status( *x );
conj_t conjy = bli_obj_conj_status( *y );
dim_t m = bli_obj_length( *c );
void* buf_x = bli_obj_buffer_at_off( *x );
inc_t incx = bli_obj_vector_inc( *x );
void* buf_y = bli_obj_buffer_at_off( *y );
inc_t incy = bli_obj_vector_inc( *y );
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 );
num_t dt_alpha;
void* buf_alpha;
FUNCPTR_T f;
// The datatype of alpha MUST be the type union of the datatypes of x and y.
dt_alpha = bli_datatype_union( dt_x, dt_y );
buf_alpha = bli_obj_buffer_for_1x1( dt_alpha, *alpha );
// Index into the type combination array to extract the correct
// function pointer.
f = ftypes[dt_x][dt_y][dt_c];
// Invoke the function.
f( uplo,
conjx,
conjy,
conjh,
m,
buf_alpha,
buf_x, incx,
buf_y, incy,
buf_c, rs_c, cs_c );
}
void bli_addm_unb_var1( obj_t* x,
obj_t* y,
cntx_t* cntx )
{
num_t dt_x = bli_obj_datatype( *x );
num_t dt_y = bli_obj_datatype( *y );
doff_t diagoffx = bli_obj_diag_offset( *x );
diag_t diagx = bli_obj_diag( *x );
uplo_t uplox = bli_obj_uplo( *x );
trans_t transx = bli_obj_conjtrans_status( *x );
dim_t m = bli_obj_length( *y );
dim_t n = bli_obj_width( *y );
inc_t rs_x = bli_obj_row_stride( *x );
inc_t cs_x = bli_obj_col_stride( *x );
void* buf_x = bli_obj_buffer_at_off( *x );
inc_t rs_y = bli_obj_row_stride( *y );
inc_t cs_y = bli_obj_col_stride( *y );
void* buf_y = bli_obj_buffer_at_off( *y );
FUNCPTR_T f;
// Index into the type combination array to extract the correct
// function pointer.
f = ftypes[dt_x][dt_y];
// Invoke the function.
f( diagoffx,
diagx,
uplox,
transx,
m,
n,
buf_x, rs_x, cs_x,
buf_y, rs_y, cs_y );
}
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" );
}
void bli_fprintm( FILE* file, char* s1, obj_t* x, char* format, char* s2 )
{
num_t dt_x = bli_obj_datatype( *x );
dim_t m = bli_obj_length( *x );
dim_t n = bli_obj_width( *x );
inc_t rs_x = bli_obj_row_stride( *x );
inc_t cs_x = bli_obj_col_stride( *x );
void* buf_x = bli_obj_buffer_at_off( *x );
FUNCPTR_T f;
if ( bli_error_checking_is_enabled() )
bli_fprintm_check( file, s1, x, format, s2 );
// Handle constants up front.
if ( dt_x == BLIS_CONSTANT )
{
float* sp = bli_obj_buffer_for_const( BLIS_FLOAT, *x );
double* dp = bli_obj_buffer_for_const( BLIS_DOUBLE, *x );
scomplex* cp = bli_obj_buffer_for_const( BLIS_SCOMPLEX, *x );
dcomplex* zp = bli_obj_buffer_for_const( BLIS_DCOMPLEX, *x );
gint_t* ip = bli_obj_buffer_for_const( BLIS_INT, *x );
fprintf( file, "%s\n", s1 );
fprintf( file, " float: %9.2e\n", bli_sreal( *sp ) );
fprintf( file, " double: %9.2e\n", bli_dreal( *dp ) );
fprintf( file, " scomplex: %9.2e + %9.2e\n", bli_creal( *cp ), bli_cimag( *cp ) );
fprintf( file, " dcomplex: %9.2e + %9.2e\n", bli_zreal( *zp ), bli_zimag( *zp ) );
fprintf( file, " int: %ld\n", *ip );
fprintf( file, "\n" );
return;
}
// Index into the type combination array to extract the correct
// function pointer.
f = ftypes[dt_x];
// Invoke the function.
f( file,
s1,
m,
n,
buf_x, rs_x, cs_x,
format,
s2 );
}
void bli_trmv_unf_var2( obj_t* alpha,
obj_t* a,
obj_t* x,
cntx_t* cntx,
trmv_t* cntl )
{
num_t dt_a = bli_obj_datatype( *a );
num_t dt_x = bli_obj_datatype( *x );
uplo_t uplo = bli_obj_uplo( *a );
trans_t trans = bli_obj_conjtrans_status( *a );
diag_t diag = bli_obj_diag( *a );
dim_t m = bli_obj_length( *a );
void* buf_a = bli_obj_buffer_at_off( *a );
inc_t rs_a = bli_obj_row_stride( *a );
inc_t cs_a = bli_obj_col_stride( *a );
void* buf_x = bli_obj_buffer_at_off( *x );
inc_t incx = bli_obj_vector_inc( *x );
num_t dt_alpha;
void* buf_alpha;
FUNCPTR_T f;
// The datatype of alpha MUST be the type union of a and x. This is to
// prevent any unnecessary loss of information during computation.
dt_alpha = bli_datatype_union( dt_a, dt_x );
buf_alpha = bli_obj_buffer_for_1x1( dt_alpha, *alpha );
// Index into the type combination array to extract the correct
// function pointer.
f = ftypes[dt_a][dt_x];
// Invoke the function.
f( uplo,
trans,
diag,
m,
buf_alpha,
buf_a, rs_a, cs_a,
buf_x, incx );
}
void bli_scald_unb_var1( obj_t* beta,
obj_t* x )
{
num_t dt_x = bli_obj_datatype( *x );
conj_t conjbeta = bli_obj_conj_status( *beta );
doff_t diagoffx = bli_obj_diag_offset( *x );
dim_t m = bli_obj_length( *x );
dim_t n = bli_obj_width( *x );
void* buf_x = bli_obj_buffer_at_off( *x );
inc_t rs_x = bli_obj_row_stride( *x );
inc_t cs_x = bli_obj_col_stride( *x );
void* buf_beta;
num_t dt_beta;
FUNCPTR_T f;
// If beta is a scalar constant, use dt_x to extract the address of the
// corresponding constant value; otherwise, use the datatype encoded
// within the beta object and extract the buffer at the beta offset.
bli_set_scalar_dt_buffer( beta, dt_x, dt_beta, buf_beta );
// Index into the type combination array to extract the correct
// function pointer.
f = ftypes[dt_beta][dt_x];
// Invoke the function.
f( conjbeta,
diagoffx,
m,
n,
buf_beta,
buf_x, rs_x, cs_x );
}
void bli_mksymm_unb_var1( obj_t* a )
{
num_t dt_a = bli_obj_datatype( *a );
uplo_t uploa = bli_obj_uplo( *a );
dim_t m = bli_obj_length( *a );
void* buf_a = bli_obj_buffer_at_off( *a );
inc_t rs_a = bli_obj_row_stride( *a );
inc_t cs_a = bli_obj_col_stride( *a );
FUNCPTR_T f;
// Index into the type combination array to extract the correct
// function pointer.
f = ftypes[dt_a];
// Invoke the function.
f( uploa,
m,
buf_a, rs_a, cs_a );
}
void bli_herk_l_ker_var2( obj_t* a,
obj_t* b,
obj_t* c,
gemm_t* cntl,
herk_thrinfo_t* thread )
{
num_t dt_exec = bli_obj_execution_datatype( *c );
doff_t diagoffc = bli_obj_diag_offset( *c );
pack_t schema_a = bli_obj_pack_schema( *a );
pack_t schema_b = bli_obj_pack_schema( *b );
dim_t m = bli_obj_length( *c );
dim_t n = bli_obj_width( *c );
dim_t k = bli_obj_width( *a );
void* buf_a = bli_obj_buffer_at_off( *a );
inc_t cs_a = bli_obj_col_stride( *a );
inc_t pd_a = bli_obj_panel_dim( *a );
inc_t ps_a = bli_obj_panel_stride( *a );
void* buf_b = bli_obj_buffer_at_off( *b );
inc_t rs_b = bli_obj_row_stride( *b );
inc_t pd_b = bli_obj_panel_dim( *b );
inc_t ps_b = bli_obj_panel_stride( *b );
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 );
obj_t scalar_a;
obj_t scalar_b;
void* buf_alpha;
void* buf_beta;
FUNCPTR_T f;
func_t* gemm_ukrs;
void* gemm_ukr;
// Detach and multiply the scalars attached to A and B.
bli_obj_scalar_detach( a, &scalar_a );
bli_obj_scalar_detach( b, &scalar_b );
bli_mulsc( &scalar_a, &scalar_b );
// Grab the addresses of the internal scalar buffers for the scalar
// merged above and the scalar attached to C.
buf_alpha = bli_obj_internal_scalar_buffer( scalar_b );
buf_beta = bli_obj_internal_scalar_buffer( *c );
// Index into the type combination array to extract the correct
// function pointer.
f = ftypes[dt_exec];
// Extract from the control tree node the func_t object containing
// the gemm micro-kernel function addresses, and then query the
// function address corresponding to the current datatype.
gemm_ukrs = cntl_gemm_ukrs( cntl );
gemm_ukr = bli_func_obj_query( dt_exec, gemm_ukrs );
// Invoke the function.
f( diagoffc,
schema_a,
schema_b,
m,
n,
k,
buf_alpha,
buf_a, cs_a, pd_a, ps_a,
buf_b, rs_b, pd_b, ps_b,
buf_beta,
buf_c, rs_c, cs_c,
gemm_ukr,
thread );
}
void bli_herk_u_ker_var2
(
obj_t* a,
obj_t* b,
obj_t* c,
cntx_t* cntx,
cntl_t* cntl,
thrinfo_t* thread
)
{
num_t dt_exec = bli_obj_exec_dt( c );
doff_t diagoffc = bli_obj_diag_offset( c );
pack_t schema_a = bli_obj_pack_schema( a );
pack_t schema_b = bli_obj_pack_schema( b );
dim_t m = bli_obj_length( c );
dim_t n = bli_obj_width( c );
dim_t k = bli_obj_width( a );
void* buf_a = bli_obj_buffer_at_off( a );
inc_t cs_a = bli_obj_col_stride( a );
inc_t is_a = bli_obj_imag_stride( a );
dim_t pd_a = bli_obj_panel_dim( a );
inc_t ps_a = bli_obj_panel_stride( a );
void* buf_b = bli_obj_buffer_at_off( b );
inc_t rs_b = bli_obj_row_stride( b );
inc_t is_b = bli_obj_imag_stride( b );
dim_t pd_b = bli_obj_panel_dim( b );
inc_t ps_b = bli_obj_panel_stride( b );
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 );
obj_t scalar_a;
obj_t scalar_b;
void* buf_alpha;
void* buf_beta;
FUNCPTR_T f;
// Detach and multiply the scalars attached to A and B.
bli_obj_scalar_detach( a, &scalar_a );
bli_obj_scalar_detach( b, &scalar_b );
bli_mulsc( &scalar_a, &scalar_b );
// Grab the addresses of the internal scalar buffers for the scalar
// merged above and the scalar attached to C.
buf_alpha = bli_obj_internal_scalar_buffer( &scalar_b );
buf_beta = bli_obj_internal_scalar_buffer( c );
// Index into the type combination array to extract the correct
// function pointer.
f = ftypes[dt_exec];
// Invoke the function.
f( diagoffc,
schema_a,
schema_b,
m,
n,
k,
buf_alpha,
buf_a, cs_a, is_a,
pd_a, ps_a,
buf_b, rs_b, is_b,
pd_b, ps_b,
buf_beta,
buf_c, rs_c, cs_c,
cntx,
thread );
}
void bli_trsm_rl_ker_var2
(
obj_t* a,
obj_t* b,
obj_t* c,
cntx_t* cntx,
cntl_t* cntl,
thrinfo_t* thread
)
{
num_t dt_exec = bli_obj_execution_datatype( *c );
doff_t diagoffb = bli_obj_diag_offset( *b );
pack_t schema_a = bli_obj_pack_schema( *a );
pack_t schema_b = bli_obj_pack_schema( *b );
dim_t m = bli_obj_length( *c );
dim_t n = bli_obj_width( *c );
dim_t k = bli_obj_width( *a );
void* buf_a = bli_obj_buffer_at_off( *a );
inc_t cs_a = bli_obj_col_stride( *a );
dim_t pd_a = bli_obj_panel_dim( *a );
inc_t ps_a = bli_obj_panel_stride( *a );
void* buf_b = bli_obj_buffer_at_off( *b );
inc_t rs_b = bli_obj_row_stride( *b );
dim_t pd_b = bli_obj_panel_dim( *b );
inc_t ps_b = bli_obj_panel_stride( *b );
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_alpha1;
void* buf_alpha2;
FUNCPTR_T f;
// Grab the address of the internal scalar buffer for the scalar
// attached to A (the non-triangular matrix). This will be the alpha
// scalar used in the gemmtrsm subproblems (ie: the scalar that would
// be applied to the packed copy of A prior to it being updated by
// the trsm subproblem). This scalar may be unit, if for example it
// was applied during packing.
buf_alpha1 = bli_obj_internal_scalar_buffer( *a );
// Grab the address of the internal scalar buffer for the scalar
// attached to C. This will be the "beta" scalar used in the gemm-only
// subproblems that correspond to micro-panels that do not intersect
// the diagonal. We need this separate scalar because it's possible
// that the alpha attached to B was reset, if it was applied during
// packing.
buf_alpha2 = bli_obj_internal_scalar_buffer( *c );
// Index into the type combination array to extract the correct
// function pointer.
f = ftypes[dt_exec];
// Invoke the function.
f( diagoffb,
schema_a,
schema_b,
m,
n,
k,
buf_alpha1,
buf_a, cs_a, pd_a, ps_a,
buf_b, rs_b, pd_b, ps_b,
buf_alpha2,
buf_c, rs_c, cs_c,
cntx,
thread );
}
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_hemv_unf_var3( conj_t conjh,
obj_t* alpha,
obj_t* a,
obj_t* x,
obj_t* beta,
obj_t* y,
hemv_t* cntl )
{
num_t dt_a = bli_obj_datatype( *a );
num_t dt_x = bli_obj_datatype( *x );
num_t dt_y = bli_obj_datatype( *y );
uplo_t uplo = bli_obj_uplo( *a );
conj_t conja = bli_obj_conj_status( *a );
conj_t conjx = bli_obj_conj_status( *x );
dim_t m = bli_obj_length( *a );
void* buf_a = bli_obj_buffer_at_off( *a );
inc_t rs_a = bli_obj_row_stride( *a );
inc_t cs_a = bli_obj_col_stride( *a );
void* buf_x = bli_obj_buffer_at_off( *x );
inc_t incx = bli_obj_vector_inc( *x );
void* buf_y = bli_obj_buffer_at_off( *y );
inc_t incy = bli_obj_vector_inc( *y );
num_t dt_alpha;
void* buf_alpha;
num_t dt_beta;
void* buf_beta;
FUNCPTR_T f;
// The datatype of alpha MUST be the type union of a and x. This is to
// prevent any unnecessary loss of information during computation.
dt_alpha = bli_datatype_union( dt_a, dt_x );
buf_alpha = bli_obj_buffer_for_1x1( dt_alpha, *alpha );
// The datatype of beta MUST be the same as the datatype of y.
dt_beta = dt_y;
buf_beta = bli_obj_buffer_for_1x1( dt_beta, *beta );
#if 0
obj_t x_copy, y_copy;
bli_obj_create( dt_x, m, 1, 0, 0, &x_copy );
bli_obj_create( dt_y, m, 1, 0, 0, &y_copy );
bli_copyv( x, &x_copy );
bli_copyv( y, &y_copy );
buf_x = bli_obj_buffer_at_off( x_copy );
buf_y = bli_obj_buffer_at_off( y_copy );
incx = 1;
incy = 1;
#endif
// Index into the type combination array to extract the correct
// function pointer.
f = ftypes[dt_a][dt_x][dt_y];
// Invoke the function.
f( uplo,
conja,
conjx,
conjh,
m,
buf_alpha,
buf_a, rs_a, cs_a,
buf_x, incx,
buf_beta,
buf_y, incy );
#if 0
bli_copyv( &y_copy, y );
bli_obj_free( &x_copy );
bli_obj_free( &y_copy );
#endif
}
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 );
}
void libblis_test_randm_check( obj_t* x,
double* resid )
{
doff_t diagoffx = bli_obj_diag_offset( *x );
uplo_t uplox = bli_obj_uplo( *x );
dim_t m_x = bli_obj_length( *x );
dim_t n_x = bli_obj_width( *x );
inc_t rs_x = bli_obj_row_stride( *x );
inc_t cs_x = bli_obj_col_stride( *x );
void* buf_x = bli_obj_buffer_at_off( *x );
*resid = 0.0;
//
// The two most likely ways that randm would fail is if all elements
// were zero, or if all elements were greater than or equal to one.
// We check both of these conditions by computing the sum of the
// absolute values of the elements of x.
//
if ( bli_obj_is_float( *x ) )
{
float sum_x;
bli_sabsumm( diagoffx,
uplox,
m_x,
n_x,
buf_x, rs_x, cs_x,
&sum_x );
if ( sum_x == *bli_s0 ) *resid = 1.0;
else if ( sum_x >= 1.0 * m_x * n_x ) *resid = 2.0;
}
else if ( bli_obj_is_double( *x ) )
{
double sum_x;
bli_dabsumm( diagoffx,
uplox,
m_x,
n_x,
buf_x, rs_x, cs_x,
&sum_x );
if ( sum_x == *bli_d0 ) *resid = 1.0;
else if ( sum_x >= 1.0 * m_x * n_x ) *resid = 2.0;
}
else if ( bli_obj_is_scomplex( *x ) )
{
float sum_x;
bli_cabsumm( diagoffx,
uplox,
m_x,
n_x,
buf_x, rs_x, cs_x,
&sum_x );
if ( sum_x == *bli_s0 ) *resid = 1.0;
else if ( sum_x >= 2.0 * m_x * n_x ) *resid = 2.0;
}
else // if ( bli_obj_is_dcomplex( *x ) )
{
double sum_x;
bli_zabsumm( diagoffx,
uplox,
m_x,
n_x,
buf_x, rs_x, cs_x,
&sum_x );
if ( sum_x == *bli_d0 ) *resid = 1.0;
else if ( sum_x >= 2.0 * m_x * n_x ) *resid = 2.0;
}
}
void bli_trsm_ru_ker_var2( obj_t* a,
obj_t* b,
obj_t* c,
trsm_t* cntl )
{
num_t dt_exec = bli_obj_execution_datatype( *c );
doff_t diagoffb = bli_obj_diag_offset( *b );
dim_t m = bli_obj_length( *c );
dim_t n = bli_obj_width( *c );
dim_t k = bli_obj_width( *a );
void* buf_a = bli_obj_buffer_at_off( *a );
inc_t cs_a = bli_obj_col_stride( *a );
inc_t pd_a = bli_obj_panel_dim( *a );
inc_t ps_a = bli_obj_panel_stride( *a );
void* buf_b = bli_obj_buffer_at_off( *b );
inc_t rs_b = bli_obj_row_stride( *b );
inc_t pd_b = bli_obj_panel_dim( *b );
inc_t ps_b = bli_obj_panel_stride( *b );
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_alpha1;
void* buf_alpha2;
FUNCPTR_T f;
func_t* gemmtrsm_ukrs;
func_t* gemm_ukrs;
void* gemmtrsm_ukr;
void* gemm_ukr;
// Grab the address of the internal scalar buffer for the scalar
// attached to A. This will be the alpha scalar used in the gemmtrsm
// subproblems (ie: the scalar that would be applied to the packed
// copy of A prior to it being updated by the trsm subproblem). This
// scalar may be unit, if for example it was applied during packing.
buf_alpha1 = bli_obj_internal_scalar_buffer( *a );
// Grab the address of the internal scalar buffer for the scalar
// attached to C. This will be the "beta" scalar used in the gemm-only
// subproblems that correspond to micro-panels that do not intersect
// the diagonal. We need this separate scalar because it's possible
// that the alpha attached to B was reset, if it was applied during
// packing.
buf_alpha2 = bli_obj_internal_scalar_buffer( *c );
// Index into the type combination array to extract the correct
// function pointer.
f = ftypes[dt_exec];
// Adjust cs_a and rs_b if A and B were packed for 4m or 3m. This
// is needed because cs_a and rs_b are used to index into the
// micro-panels of A and B, respectively, and since the pointer
// types in the macro-kernel (scomplex or dcomplex) will result
// in pointer arithmetic that moves twice as far as it should,
// given the datatypes actually stored (float or double), we must
// halve the strides to compensate.
if ( bli_obj_is_panel_packed_4m( *a ) ||
bli_obj_is_panel_packed_3m( *a ) ) { cs_a /= 2; rs_b /= 2; }
// Extract from the control tree node the func_t objects containing
// the gemmtrsm and gemm micro-kernel function addresses, and then
// query the function addresses corresponding to the current datatype.
gemmtrsm_ukrs = cntl_gemmtrsm_l_ukrs( cntl );
gemm_ukrs = cntl_gemm_ukrs( cntl );
gemmtrsm_ukr = bli_func_obj_query( dt_exec, gemmtrsm_ukrs );
gemm_ukr = bli_func_obj_query( dt_exec, gemm_ukrs );
// Invoke the function.
f( diagoffb,
m,
n,
k,
buf_alpha1,
buf_a, cs_a, pd_a, ps_a,
buf_b, rs_b, pd_b, ps_b,
buf_alpha2,
buf_c, rs_c, cs_c,
gemmtrsm_ukr,
gemm_ukr );
}
void bli_trsm_rl_ker_var2( obj_t* a,
obj_t* b,
obj_t* c,
trsm_t* cntl,
trsm_thrinfo_t* thread )
{
num_t dt_exec = bli_obj_execution_datatype( *c );
doff_t diagoffb = bli_obj_diag_offset( *b );
pack_t schema_a = bli_obj_pack_schema( *a );
pack_t schema_b = bli_obj_pack_schema( *b );
dim_t m = bli_obj_length( *c );
dim_t n = bli_obj_width( *c );
dim_t k = bli_obj_width( *a );
void* buf_a = bli_obj_buffer_at_off( *a );
inc_t cs_a = bli_obj_col_stride( *a );
dim_t pd_a = bli_obj_panel_dim( *a );
inc_t ps_a = bli_obj_panel_stride( *a );
void* buf_b = bli_obj_buffer_at_off( *b );
inc_t rs_b = bli_obj_row_stride( *b );
dim_t pd_b = bli_obj_panel_dim( *b );
inc_t ps_b = bli_obj_panel_stride( *b );
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_alpha1;
void* buf_alpha2;
FUNCPTR_T f;
func_t* gemmtrsm_ukrs;
func_t* gemm_ukrs;
void* gemmtrsm_ukr;
void* gemm_ukr;
// Grab the address of the internal scalar buffer for the scalar
// attached to A. This will be the alpha scalar used in the gemmtrsm
// subproblems (ie: the scalar that would be applied to the packed
// copy of A prior to it being updated by the trsm subproblem). This
// scalar may be unit, if for example it was applied during packing.
buf_alpha1 = bli_obj_internal_scalar_buffer( *a );
// Grab the address of the internal scalar buffer for the scalar
// attached to C. This will be the "beta" scalar used in the gemm-only
// subproblems that correspond to micro-panels that do not intersect
// the diagonal. We need this separate scalar because it's possible
// that the alpha attached to B was reset, if it was applied during
// packing.
buf_alpha2 = bli_obj_internal_scalar_buffer( *c );
// Index into the type combination array to extract the correct
// function pointer.
f = ftypes[dt_exec];
// Extract from the control tree node the func_t objects containing
// the gemmtrsm and gemm micro-kernel function addresses, and then
// query the function addresses corresponding to the current datatype.
gemmtrsm_ukrs = cntl_gemmtrsm_u_ukrs( cntl );
gemm_ukrs = cntl_gemm_ukrs( cntl );
gemmtrsm_ukr = bli_func_obj_query( dt_exec, gemmtrsm_ukrs );
gemm_ukr = bli_func_obj_query( dt_exec, gemm_ukrs );
// Invoke the function.
f( diagoffb,
schema_a,
schema_b,
m,
n,
k,
buf_alpha1,
buf_a, cs_a, pd_a, ps_a,
buf_b, rs_b, pd_b, ps_b,
buf_alpha2,
buf_c, rs_c, cs_c,
gemmtrsm_ukr,
gemm_ukr,
thread );
}
void bli_trmm_rl_ker_var2( obj_t* a,
obj_t* b,
obj_t* c,
trmm_t* cntl )
{
num_t dt_exec = bli_obj_execution_datatype( *c );
doff_t diagoffb = bli_obj_diag_offset( *b );
dim_t m = bli_obj_length( *c );
dim_t n = bli_obj_width( *c );
dim_t k = bli_obj_width( *a );
void* buf_a = bli_obj_buffer_at_off( *a );
inc_t rs_a = bli_obj_row_stride( *a );
inc_t cs_a = bli_obj_col_stride( *a );
inc_t ps_a = bli_obj_panel_stride( *a );
void* buf_b = bli_obj_buffer_at_off( *b );
inc_t rs_b = bli_obj_row_stride( *b );
inc_t cs_b = bli_obj_col_stride( *b );
inc_t ps_b = bli_obj_panel_stride( *b );
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 );
obj_t scalar_a;
obj_t scalar_b;
void* buf_alpha;
void* buf_beta;
FUNCPTR_T f;
// Detach and multiply the scalars attached to A and B.
bli_obj_scalar_detach( a, &scalar_a );
bli_obj_scalar_detach( b, &scalar_b );
bli_mulsc( &scalar_a, &scalar_b );
// Grab the addresses of the internal scalar buffers for the scalar
// merged above and the scalar attached to C.
buf_alpha = bli_obj_internal_scalar_buffer( scalar_b );
buf_beta = bli_obj_internal_scalar_buffer( *c );
// Index into the type combination array to extract the correct
// function pointer.
f = ftypes[dt_exec];
// Invoke the function.
f( diagoffb,
m,
n,
k,
buf_alpha,
buf_a, rs_a, cs_a, ps_a,
buf_b, rs_b, cs_b, ps_b,
buf_beta,
buf_c, rs_c, cs_c );
}
