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 );
}
err_t bli_check_upper_or_lower_object( obj_t* a )
{
err_t e_val = BLIS_SUCCESS;
if ( !bli_obj_is_lower( *a ) &&
!bli_obj_is_upper( *a ) )
e_val = BLIS_EXPECTED_UPPER_OR_LOWER_OBJECT;
return e_val;
}
void bli_gemmtrsm_ukr_make_subparts( dim_t k,
obj_t* a,
obj_t* b,
obj_t* a1x,
obj_t* a11,
obj_t* bx1,
obj_t* b11 )
{
dim_t mr = bli_obj_length( *a );
dim_t nr = bli_obj_width( *b );
dim_t off_a1x, off_a11;
dim_t off_bx1, off_b11;
if ( bli_obj_is_lower( *a ) )
{
off_a1x = 0;
off_a11 = k;
off_bx1 = 0;
off_b11 = k;
}
else
{
off_a1x = mr;
off_a11 = 0;
off_bx1 = mr;
off_b11 = 0;
}
bli_obj_init_subpart_from( *a, *a1x );
bli_obj_set_dims( mr, k, *a1x );
bli_obj_inc_offs( 0, off_a1x, *a1x );
bli_obj_init_subpart_from( *a, *a11 );
bli_obj_set_dims( mr, mr, *a11 );
bli_obj_inc_offs( 0, off_a11, *a11 );
bli_obj_init_subpart_from( *b, *bx1 );
bli_obj_set_dims( k, nr, *bx1 );
bli_obj_inc_offs( off_bx1, 0, *bx1 );
bli_obj_init_subpart_from( *b, *b11 );
bli_obj_set_dims( mr, nr, *b11 );
bli_obj_inc_offs( off_b11, 0, *b11 );
// Mark a1x as having general structure (which overwrites the triangular
// property it inherited from a).
bli_obj_set_struc( BLIS_GENERAL, *a1x );
// Set the diagonal offset of a11 to 0 (which overwrites the diagonal
// offset value it inherited from a).
bli_obj_set_diag_offset( 0, *a11 );
}
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 libblis_test_gemmtrsm_ukr_experiment( test_params_t* params,
test_op_t* op,
mt_impl_t impl,
num_t datatype,
char* pc_str,
char* sc_str,
unsigned int p_cur,
double* perf,
double* resid )
{
unsigned int n_repeats = params->n_repeats;
unsigned int i;
double time_min = 1e9;
double time;
dim_t m, n, k;
char sc_a = 'c';
char sc_b = 'r';
side_t side = BLIS_LEFT;
uplo_t uploa;
obj_t kappa;
obj_t alpha;
obj_t a_big, a, b;
obj_t b11, c11;
obj_t ap, bp;
obj_t a1xp, a11p, bx1p, b11p;
obj_t c11_save;
// Map the dimension specifier to actual dimensions.
k = libblis_test_get_dim_from_prob_size( op->dim_spec[0], p_cur );
// Fix m and n to MR and NR, respectively.
m = bli_blksz_for_type( datatype, gemm_mr );
n = bli_blksz_for_type( datatype, gemm_nr );
// Store the register blocksizes so that the driver can retrieve the
// values later when printing results.
op->dim_aux[0] = m;
op->dim_aux[1] = n;
// Map parameter characters to BLIS constants.
bli_param_map_char_to_blis_uplo( pc_str[0], &uploa );
// Create test scalars.
bli_obj_scalar_init_detached( datatype, &kappa );
bli_obj_scalar_init_detached( datatype, &alpha );
// Create test operands (vectors and/or matrices).
libblis_test_mobj_create( params, datatype, BLIS_NO_TRANSPOSE,
sc_a, k+m, k+m, &a_big );
libblis_test_mobj_create( params, datatype, BLIS_NO_TRANSPOSE,
sc_b, k+m, n, &b );
libblis_test_mobj_create( params, datatype, BLIS_NO_TRANSPOSE,
sc_str[0], m, n, &c11 );
libblis_test_mobj_create( params, datatype, BLIS_NO_TRANSPOSE,
sc_str[0], m, n, &c11_save );
// Set alpha.
if ( bli_obj_is_real( b ) )
{
bli_setsc( 2.0, 0.0, &alpha );
}
else
{
bli_setsc( 2.0, 0.0, &alpha );
}
// Set the structure, uplo, and diagonal offset properties of A.
bli_obj_set_struc( BLIS_TRIANGULAR, a_big );
bli_obj_set_uplo( uploa, a_big );
// Randomize A and make it densely triangular.
bli_randm( &a_big );
// Normalize B and save.
bli_randm( &b );
bli_setsc( 1.0/( double )m, 0.0, &kappa );
bli_scalm( &kappa, &b );
// Use the last m rows of A_big as A.
bli_acquire_mpart_t2b( BLIS_SUBPART1, k, m, &a_big, &a );
// Locate the B11 block of B, copy to C11, and save.
if ( bli_obj_is_lower( a ) )
bli_acquire_mpart_t2b( BLIS_SUBPART1, k, m, &b, &b11 );
else
bli_acquire_mpart_t2b( BLIS_SUBPART1, 0, m, &b, &b11 );
bli_copym( &b11, &c11 );
bli_copym( &c11, &c11_save );
// Initialize pack objects.
bli_obj_init_pack( &ap );
bli_obj_init_pack( &bp );
// Create pack objects for a and b.
libblis_test_pobj_create( gemm_mr,
gemm_mr,
BLIS_INVERT_DIAG,
BLIS_PACKED_ROW_PANELS,
BLIS_BUFFER_FOR_A_BLOCK,
&a, &ap );
libblis_test_pobj_create( gemm_mr,
gemm_nr,
BLIS_NO_INVERT_DIAG,
BLIS_PACKED_COL_PANELS,
BLIS_BUFFER_FOR_B_PANEL,
&b, &bp );
// Pack the contents of a to ap.
bli_packm_blk_var3( &a, &ap );
// Pack the contents of b to bp.
bli_packm_blk_var2( &b, &bp );
// Create subpartitions from the a and b panels.
bli_gemmtrsm_ukr_make_subparts( k, &ap, &bp,
&a1xp, &a11p, &bx1p, &b11p );
// Repeat the experiment n_repeats times and record results.
for ( i = 0; i < n_repeats; ++i )
{
bli_copym( &c11_save, &c11 );
// Re-pack the contents of b to bp.
bli_packm_blk_var2( &b, &bp );
time = bli_clock();
libblis_test_gemmtrsm_ukr_impl( impl, side, &alpha,
&a1xp, &a11p, &bx1p, &b11p, &c11 );
time_min = bli_clock_min_diff( time_min, time );
}
// Estimate the performance of the best experiment repeat.
*perf = ( 2.0 * m * n * k + 1.0 * m * m * n ) / time_min / FLOPS_PER_UNIT_PERF;
if ( bli_obj_is_complex( b ) ) *perf *= 4.0;
// Perform checks.
libblis_test_gemmtrsm_ukr_check( side, &alpha,
&a1xp, &a11p, &bx1p, &b11p, &c11, &c11_save, resid );
// Zero out performance and residual if output matrix is empty.
//libblis_test_check_empty_problem( &c11, perf, resid );
// Release packing buffers within pack objects.
bli_obj_release_pack( &ap );
bli_obj_release_pack( &bp );
// Free the test objects.
bli_obj_free( &a_big );
bli_obj_free( &b );
bli_obj_free( &c11 );
bli_obj_free( &c11_save );
}
void bli_syr2_front
(
obj_t* alpha,
obj_t* x,
obj_t* y,
obj_t* c,
cntx_t* cntx
)
{
her2_t* her2_cntl;
num_t dt_targ_x;
num_t dt_targ_y;
//num_t dt_targ_c;
bool_t x_has_unit_inc;
bool_t y_has_unit_inc;
bool_t c_has_unit_inc;
obj_t alpha_local;
num_t dt_alpha;
// Check parameters.
if ( bli_error_checking_is_enabled() )
bli_syr2_check( alpha, x, y, c );
// Query the target datatypes of each object.
dt_targ_x = bli_obj_target_dt( x );
dt_targ_y = bli_obj_target_dt( y );
//dt_targ_c = bli_obj_target_dt( c );
// Determine whether each operand with unit stride.
x_has_unit_inc = ( bli_obj_vector_inc( x ) == 1 );
y_has_unit_inc = ( bli_obj_vector_inc( y ) == 1 );
c_has_unit_inc = ( bli_obj_is_row_stored( c ) ||
bli_obj_is_col_stored( c ) );
// Create an object to hold a copy-cast of alpha. Notice that we use
// the type union of the datatypes of x and y.
dt_alpha = bli_dt_union( dt_targ_x, dt_targ_y );
bli_obj_scalar_init_detached_copy_of( dt_alpha,
BLIS_NO_CONJUGATE,
alpha,
&alpha_local );
// If all operands have unit stride, we choose a control tree for calling
// the unblocked implementation directly without any blocking.
if ( x_has_unit_inc &&
y_has_unit_inc &&
c_has_unit_inc )
{
// We use two control trees to handle the four cases corresponding to
// combinations of upper/lower triangular storage and row/column-storage.
// The row-stored lower triangular and column-stored upper triangular
// trees are identical. Same for the remaining two trees.
if ( bli_obj_is_lower( c ) )
{
if ( bli_obj_is_row_stored( c ) ) her2_cntl = her2_cntl_bs_ke_lrow_ucol;
else her2_cntl = her2_cntl_bs_ke_lcol_urow;
}
else // if ( bli_obj_is_upper( c ) )
{
if ( bli_obj_is_row_stored( c ) ) her2_cntl = her2_cntl_bs_ke_lcol_urow;
else her2_cntl = her2_cntl_bs_ke_lrow_ucol;
}
}
else
{
// Mark objects with unit stride as already being packed. This prevents
// unnecessary packing from happening within the blocked algorithm.
if ( x_has_unit_inc ) bli_obj_set_pack_schema( BLIS_PACKED_VECTOR, x );
if ( y_has_unit_inc ) bli_obj_set_pack_schema( BLIS_PACKED_VECTOR, y );
if ( c_has_unit_inc ) bli_obj_set_pack_schema( BLIS_PACKED_UNSPEC, c );
// Here, we make a similar choice as above, except that (1) we look
// at storage tilt, and (2) we choose a tree that performs blocking.
if ( bli_obj_is_lower( c ) )
{
if ( bli_obj_is_row_stored( c ) ) her2_cntl = her2_cntl_ge_lrow_ucol;
else her2_cntl = her2_cntl_ge_lcol_urow;
}
else // if ( bli_obj_is_upper( c ) )
{
if ( bli_obj_is_row_stored( c ) ) her2_cntl = her2_cntl_ge_lcol_urow;
else her2_cntl = her2_cntl_ge_lrow_ucol;
}
}
// Invoke the internal back-end with the copy-cast scalar and the
// chosen control tree. Set conjh to BLIS_NO_CONJUGATE to invoke the
// symmetric (and not Hermitian) algorithms.
bli_her2_int( BLIS_NO_CONJUGATE,
&alpha_local,
&alpha_local,
x,
y,
c,
cntx,
her2_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_hemv( obj_t* alpha,
obj_t* a,
obj_t* x,
obj_t* beta,
obj_t* y )
{
hemv_t* hemv_cntl;
num_t dt_targ_a;
num_t dt_targ_x;
num_t dt_targ_y;
bool_t a_has_unit_inc;
bool_t x_has_unit_inc;
bool_t y_has_unit_inc;
obj_t alpha_local;
obj_t beta_local;
num_t dt_alpha;
num_t dt_beta;
// Check parameters.
if ( bli_error_checking_is_enabled() )
bli_hemv_check( alpha, a, x, beta, y );
// Query the target datatypes of each object.
dt_targ_a = bli_obj_target_datatype( *a );
dt_targ_x = bli_obj_target_datatype( *x );
dt_targ_y = bli_obj_target_datatype( *y );
// Determine whether each operand with unit stride.
a_has_unit_inc = ( bli_obj_is_row_stored( *a ) ||
bli_obj_is_col_stored( *a ) );
x_has_unit_inc = ( bli_obj_vector_inc( *x ) == 1 );
y_has_unit_inc = ( bli_obj_vector_inc( *y ) == 1 );
// Create an object to hold a copy-cast of alpha. Notice that we use
// the type union of the target datatypes of a and x to prevent any
// unnecessary loss of information during the computation.
dt_alpha = bli_datatype_union( dt_targ_a, dt_targ_x );
bli_obj_scalar_init_detached_copy_of( dt_alpha,
BLIS_NO_CONJUGATE,
alpha,
&alpha_local );
// Create an object to hold a copy-cast of beta. Notice that we use
// the datatype of y. Here's why: If y is real and beta is complex,
// there is no reason to keep beta_local in the complex domain since
// the complex part of beta*y will not be stored. If y is complex and
// beta is real then beta is harmlessly promoted to complex.
dt_beta = dt_targ_y;
bli_obj_scalar_init_detached_copy_of( dt_beta,
BLIS_NO_CONJUGATE,
beta,
&beta_local );
// If all operands have unit stride, we choose a control tree for calling
// the unblocked implementation directly without any blocking.
if ( a_has_unit_inc &&
x_has_unit_inc &&
y_has_unit_inc )
{
// We use two control trees to handle the four cases corresponding to
// combinations of upper/lower triangular storage and row/column-storage.
// The row-stored lower triangular and column-stored upper triangular
// trees are identical. Same for the remaining two trees.
if ( bli_obj_is_lower( *a ) )
{
if ( bli_obj_is_row_stored( *a ) ) hemv_cntl = hemv_cntl_bs_ke_lrow_ucol;
else hemv_cntl = hemv_cntl_bs_ke_lcol_urow;
}
else // if ( bli_obj_is_upper( *a ) )
{
if ( bli_obj_is_row_stored( *a ) ) hemv_cntl = hemv_cntl_bs_ke_lcol_urow;
else hemv_cntl = hemv_cntl_bs_ke_lrow_ucol;
}
}
else
{
// Mark objects with unit stride as already being packed. This prevents
// unnecessary packing from happening within the blocked algorithm.
if ( a_has_unit_inc ) bli_obj_set_pack_schema( BLIS_PACKED_UNSPEC, *a );
if ( x_has_unit_inc ) bli_obj_set_pack_schema( BLIS_PACKED_VECTOR, *x );
if ( y_has_unit_inc ) bli_obj_set_pack_schema( BLIS_PACKED_VECTOR, *y );
// Here, we make a similar choice as above, except that (1) we look
// at storage tilt, and (2) we choose a tree that performs blocking.
if ( bli_obj_is_lower( *a ) )
{
if ( bli_obj_is_row_tilted( *a ) ) hemv_cntl = hemv_cntl_ge_lrow_ucol;
else hemv_cntl = hemv_cntl_ge_lcol_urow;
}
else // if ( bli_obj_is_upper( *a ) )
{
if ( bli_obj_is_row_tilted( *a ) ) hemv_cntl = hemv_cntl_ge_lcol_urow;
else hemv_cntl = hemv_cntl_ge_lrow_ucol;
}
}
// Invoke the internal back-end with the copy-casts of scalars and the
// chosen control tree. Set conjh to BLIS_CONJUGATE to invoke the
// Hermitian (and not symmetric) algorithms.
bli_hemv_int( BLIS_CONJUGATE,
&alpha_local,
a,
x,
&beta_local,
y,
hemv_cntl );
}
void bli_trsv_int( obj_t* alpha,
obj_t* a,
obj_t* x,
cntx_t* cntx,
trsv_t* cntl )
{
varnum_t n;
impl_t i;
bool_t uplo;
FUNCPTR_T f;
obj_t a_local;
// Check parameters.
if ( bli_error_checking_is_enabled() )
bli_trsv_check( alpha, a, x );
// If A or x has a zero dimension, return early.
if ( bli_obj_has_zero_dim( a ) ) return;
if ( bli_obj_has_zero_dim( x ) ) return;
// Alias A in case we need to induce a transformation (ie: transposition).
bli_obj_alias_to( a, &a_local );
// NOTE: to support cases where B is complex and A is real, we will
// need to have the default side case be BLIS_RIGHT and then express
// the left case in terms of it, rather than the other way around.
// Determine uplo (for indexing to the correct function pointer).
if ( bli_obj_is_lower( &a_local ) ) uplo = 0;
else uplo = 1;
// We do not explicitly implement the cases where A is transposed.
// However, we can still handle them. Specifically, if A is marked as
// needing a transposition, we simply toggle the uplo value to cause the
// correct algorithm to be induced. When that algorithm partitions into
// A, it will grab the correct subpartitions, which will inherit A's
// transposition bit and thus downstream subproblems will do the right
// thing. Alternatively, we could accomplish the same end goal by
// inducing a transposition, via bli_obj_induce_trans(), in the code
// block below. That macro function swaps dimensions, strides, and
// offsets. As an example, given a lower triangular, column-major matrix
// that needs a transpose, we would induce that transposition by recasting
// the object as an upper triangular, row-major matrix (with no transpose
// needed). Note that how we choose to handle transposition here does NOT
// affect the optimal choice of kernel (ie: a column-major column panel
// matrix with transpose times a vector would use the same kernel as a
// row-major row panel matrix with no transpose times a vector).
if ( bli_obj_has_trans( &a_local ) )
{
//bli_obj_induce_trans( &a_local );
//bli_obj_set_onlytrans( BLIS_NO_TRANSPOSE, &a_local );
if ( uplo == 1 ) uplo = 0;
else uplo = 1;
}
// Extract the variant number and implementation type.
n = bli_cntl_var_num( cntl );
i = bli_cntl_impl_type( cntl );
// Index into the variant array to extract the correct function pointer.
f = vars[uplo][n][i];
// Invoke the variant.
f( alpha,
&a_local,
x,
cntx,
cntl );
}
void bli_trsm_blk_var1f( obj_t* a,
obj_t* b,
obj_t* c,
trsm_t* cntl )
{
obj_t a1, a1_pack;
obj_t b_pack;
obj_t c1;
dim_t i;
dim_t b_alg;
dim_t m_trans;
dim_t offA;
// Initialize all pack objects that are passed into packm_init().
bli_obj_init_pack( &a1_pack );
bli_obj_init_pack( &b_pack );
// 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 ) );
// Initialize object for packing B.
bli_packm_init( b, &b_pack,
cntl_sub_packm_b( cntl ) );
// Pack B1 (if instructed).
bli_packm_int( b, &b_pack,
cntl_sub_packm_b( cntl ) );
// Partition along the remaining portion of the m dimension.
for ( i = offA; i < m_trans; i += b_alg )
{
// Determine the current algorithmic blocksize.
b_alg = bli_determine_blocksize_f( i, m_trans, a,
cntl_blocksize( cntl ) );
// Acquire partitions for A1 and C1.
bli_acquire_mpart_t2b( BLIS_SUBPART1,
i, b_alg, a, &a1 );
bli_acquire_mpart_t2b( BLIS_SUBPART1,
i, b_alg, c, &c1 );
// Initialize object for packing A1.
bli_packm_init( &a1, &a1_pack,
cntl_sub_packm_a( cntl ) );
// Pack A1 (if instructed).
bli_packm_int( &a1, &a1_pack,
cntl_sub_packm_a( cntl ) );
// Perform trsm subproblem.
bli_trsm_int( &BLIS_ONE,
&a1_pack,
&b_pack,
&BLIS_ONE,
&c1,
cntl_sub_trsm( cntl ) );
}
// If any packing buffers were acquired within packm, release them back
// to the memory manager.
bli_obj_release_pack( &a1_pack );
bli_obj_release_pack( &b_pack );
}
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_trmm_blk_var1( obj_t* alpha,
obj_t* a,
obj_t* b,
obj_t* beta,
obj_t* c,
trmm_t* cntl )
{
obj_t a1, a1_pack;
obj_t b_pack;
obj_t c1, c1_pack;
dim_t i;
dim_t b_alg;
dim_t m_trans;
dim_t offA;
// Initialize all pack objects that are passed into packm_init().
bli_obj_init_pack( &a1_pack );
bli_obj_init_pack( &b_pack );
bli_obj_init_pack( &c1_pack );
// Set the default length of and offset to the non-zero part of A.
m_trans = bli_obj_length_after_trans( *a );
offA = 0;
// If A is lower triangular, we have to adjust where the non-zero part of
// A begins. If A is upper triangular, we have to adjust the length of
// the non-zero part. If A is general/dense, then we keep the defaults.
if ( bli_obj_is_lower( *a ) )
offA = bli_abs( bli_obj_diag_offset_after_trans( *a ) );
else if ( bli_obj_is_upper( *a ) )
m_trans = bli_abs( bli_obj_diag_offset_after_trans( *a ) ) +
bli_obj_width_after_trans( *a );
// Scale C by beta (if instructed).
bli_scalm_int( beta,
c,
cntl_sub_scalm( cntl ) );
// Initialize object for packing B.
bli_packm_init( b, &b_pack,
cntl_sub_packm_b( cntl ) );
// Pack B and scale by alpha (if instructed).
bli_packm_int( alpha,
b, &b_pack,
cntl_sub_packm_b( cntl ) );
// Partition along the m dimension.
for ( i = offA; i < m_trans; i += b_alg )
{
// Determine the current algorithmic blocksize.
b_alg = bli_determine_blocksize_f( i, m_trans, a,
cntl_blocksize( cntl ) );
// Acquire partitions for A1 and C1.
bli_acquire_mpart_t2b( BLIS_SUBPART1,
i, b_alg, a, &a1 );
bli_acquire_mpart_t2b( BLIS_SUBPART1,
i, b_alg, c, &c1 );
// Initialize objects for packing A1 and C1.
bli_packm_init( &a1, &a1_pack,
cntl_sub_packm_a( cntl ) );
bli_packm_init( &c1, &c1_pack,
cntl_sub_packm_c( cntl ) );
// Pack A1 and scale by alpha (if instructed).
bli_packm_int( alpha,
&a1, &a1_pack,
cntl_sub_packm_a( cntl ) );
// Pack C1 and scale by beta (if instructed).
bli_packm_int( beta,
&c1, &c1_pack,
cntl_sub_packm_c( cntl ) );
// Perform trmm subproblem.
bli_trmm_int( alpha,
&a1_pack,
&b_pack,
beta,
&c1_pack,
cntl_sub_trmm( cntl ) );
// Unpack C1 (if C1 was packed).
bli_unpackm_int( &c1_pack, &c1,
cntl_sub_unpackm_c( cntl ) );
}
// If any packing buffers were acquired within packm, release them back
// to the memory manager.
bli_obj_release_pack( &a1_pack );
bli_obj_release_pack( &b_pack );
bli_obj_release_pack( &c1_pack );
}