void bli_trmm_int( obj_t* alpha,
obj_t* a,
obj_t* b,
obj_t* beta,
obj_t* c,
trmm_t* cntl )
{
obj_t a_local;
obj_t b_local;
obj_t c_local;
bool_t side, uplo;
varnum_t n;
impl_t i;
FUNCPTR_T f;
// Check parameters.
if ( bli_error_checking_is_enabled() )
bli_trmm_int_check( alpha, a, b, beta, c, cntl );
// If C has a zero dimension, return early.
if ( bli_obj_has_zero_dim( *c ) ) return;
// If A or B has a zero dimension, scale C by beta and return early.
if ( bli_obj_has_zero_dim( *a ) ||
bli_obj_has_zero_dim( *b ) )
{
bli_scalm( beta, c );
return;
}
// Alias A and B in case we need to update attached scalars.
bli_obj_alias_to( *a, a_local );
bli_obj_alias_to( *b, b_local );
// Alias C in case we need to induce a transposition.
bli_obj_alias_to( *c, c_local );
// If we are about to call a leaf-level implementation, and matrix C
// still needs a transposition, then we must induce one by swapping the
// strides and dimensions. Note that this transposition would normally
// be handled explicitly in the packing of C, but if C is not being
// packed, this is our last chance to handle the transposition.
if ( cntl_is_leaf( cntl ) && bli_obj_has_trans( *c ) )
{
bli_obj_induce_trans( c_local );
bli_obj_set_onlytrans( BLIS_NO_TRANSPOSE, c_local );
}
// If alpha is non-unit, typecast and apply it to the scalar attached
// to B.
if ( !bli_obj_equals( alpha, &BLIS_ONE ) )
{
bli_obj_scalar_apply_scalar( alpha, &b_local );
}
// If beta is non-unit, typecast and apply it to the scalar attached
// to C.
if ( !bli_obj_equals( beta, &BLIS_ONE ) )
{
bli_obj_scalar_apply_scalar( beta, &c_local );
}
// Set two bools: one based on the implied side parameter (the structure
// of the root object) and one based on the uplo field of the triangular
// matrix's root object (whether that is matrix A or matrix B).
if ( bli_obj_root_is_triangular( *a ) )
{
side = 0;
if ( bli_obj_root_is_lower( *a ) ) uplo = 0;
else uplo = 1;
}
else // if ( bli_obj_root_is_triangular( *b ) )
{
side = 1;
// Set a bool based on the uplo field of A's root object.
if ( bli_obj_root_is_lower( *b ) ) uplo = 0;
else uplo = 1;
}
// Extract the variant number and implementation type.
n = cntl_var_num( cntl );
i = cntl_impl_type( cntl );
// Index into the variant array to extract the correct function pointer.
f = vars[side][uplo][n][i];
// Invoke the variant.
f( &a_local,
&b_local,
&c_local,
cntl );
}
void bli_trmm3_front( side_t side,
obj_t* alpha,
obj_t* a,
obj_t* b,
obj_t* beta,
obj_t* c,
trmm_t* l_cntl,
trmm_t* r_cntl )
{
trmm_t* cntl;
obj_t a_local;
obj_t b_local;
obj_t c_local;
// Check parameters.
if ( bli_error_checking_is_enabled() )
bli_trmm3_check( side, alpha, a, b, beta, c );
// If alpha is zero, scale by beta and return.
if ( bli_obj_equals( alpha, &BLIS_ZERO ) )
{
bli_scalm( beta, c );
return;
}
// Alias A, B, and C so we can tweak the objects if necessary.
bli_obj_alias_to( *a, a_local );
bli_obj_alias_to( *b, b_local );
bli_obj_alias_to( *c, c_local );
// 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 induce a transposition. This
// allows us to only explicitly implement the no-transpose cases. Once
// the transposition is induced, the correct algorithm will be called,
// since, for example, an algorithm over a transposed lower triangular
// matrix A moves in the same direction (forwards) as a non-transposed
// upper triangular matrix. And with the transposition induced, the
// matrix now appears to be upper triangular, so the upper triangular
// algorithm will grab the correct partitions, as if it were upper
// triangular (with no transpose) all along.
if ( bli_obj_has_trans( a_local ) )
{
bli_obj_induce_trans( a_local );
bli_obj_set_onlytrans( BLIS_NO_TRANSPOSE, a_local );
}
#if 0
if ( bli_is_right( side ) )
{
bli_obj_induce_trans( a_local );
bli_obj_induce_trans( b_local );
bli_obj_induce_trans( c_local );
bli_toggle_side( side );
}
#endif
#if 1
// If A is being multiplied from the right, swap A and B so that
// the matrix will actually be on the right.
if ( bli_is_right( side ) )
{
bli_obj_swap( a_local, b_local );
}
// An optimization: If C is row-stored, transpose the entire operation
// so as to allow the macro-kernel more favorable access patterns
// through C. (The effect of the transposition of A and B is negligible
// because those operands are always packed to contiguous memory.)
if ( bli_obj_is_row_stored( c_local ) )
{
bli_obj_swap( a_local, b_local );
bli_obj_induce_trans( a_local );
bli_obj_induce_trans( b_local );
bli_obj_induce_trans( c_local );
bli_toggle_side( side );
}
#endif
// Set each alias as the root object.
// NOTE: We MUST wait until we are done potentially swapping the objects
// before setting the root fields!
bli_obj_set_as_root( a_local );
bli_obj_set_as_root( b_local );
bli_obj_set_as_root( c_local );
// Choose the control tree.
if ( bli_is_left( side ) ) cntl = l_cntl;
else cntl = r_cntl;
trmm_thrinfo_t** infos = bli_create_trmm_thrinfo_paths( FALSE );
dim_t n_threads = thread_num_threads( infos[0] );
// Invoke the internal back-end.
bli_level3_thread_decorator( n_threads,
(level3_int_t) bli_trmm_int,
alpha,
&a_local,
&b_local,
beta,
&c_local,
(void*) cntl,
(void**) infos );
bli_trmm_thrinfo_free_paths( infos, n_threads );
}
void bli_trmm3_front( side_t side,
obj_t* alpha,
obj_t* a,
obj_t* b,
obj_t* beta,
obj_t* c,
gemm_t* cntl )
{
obj_t a_local;
obj_t b_local;
obj_t c_local;
// Check parameters.
if ( bli_error_checking_is_enabled() )
bli_trmm3_check( side, alpha, a, b, beta, c );
// If alpha is zero, scale by beta and return.
if ( bli_obj_equals( alpha, &BLIS_ZERO ) )
{
bli_scalm( beta, c );
return;
}
// Alias A, B, and C so we can tweak the objects if necessary.
bli_obj_alias_to( *a, a_local );
bli_obj_alias_to( *b, b_local );
bli_obj_alias_to( *c, c_local );
// 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 induce a transposition. This
// allows us to only explicitly implement the no-transpose cases. Once
// the transposition is induced, the correct algorithm will be called,
// since, for example, an algorithm over a transposed lower triangular
// matrix A moves in the same direction (forwards) as a non-transposed
// upper triangular matrix. And with the transposition induced, the
// matrix now appears to be upper triangular, so the upper triangular
// algorithm will grab the correct partitions, as if it were upper
// triangular (with no transpose) all along.
if ( bli_obj_has_trans( a_local ) )
{
bli_obj_induce_trans( a_local );
bli_obj_set_onlytrans( BLIS_NO_TRANSPOSE, a_local );
}
#if 0
// If A is being multiplied from the right, transpose all operands
// so that we can perform the computation as if A were being multiplied
// from the left.
if ( bli_is_right( side ) )
{
bli_toggle_side( side );
bli_obj_induce_trans( a_local );
bli_obj_induce_trans( b_local );
bli_obj_induce_trans( c_local );
}
#else
// An optimization: If C is stored by rows and the micro-kernel prefers
// contiguous columns, or if C is stored by columns and the micro-kernel
// prefers contiguous rows, transpose the entire operation to allow the
// micro-kernel to access elements of C in its preferred manner.
if (
( bli_obj_is_row_stored( c_local ) &&
bli_func_prefers_contig_cols( bli_obj_datatype( c_local ),
bli_gemm_cntl_ukrs( cntl ) ) ) ||
( bli_obj_is_col_stored( c_local ) &&
bli_func_prefers_contig_rows( bli_obj_datatype( c_local ),
bli_gemm_cntl_ukrs( cntl ) ) )
)
{
bli_toggle_side( side );
bli_obj_induce_trans( a_local );
bli_obj_induce_trans( b_local );
bli_obj_induce_trans( c_local );
}
// If A is being multiplied from the right, swap A and B so that
// the matrix will actually be on the right.
if ( bli_is_right( side ) )
{
bli_obj_swap( a_local, b_local );
}
#endif
// Set each alias as the root object.
// NOTE: We MUST wait until we are done potentially swapping the objects
// before setting the root fields!
bli_obj_set_as_root( a_local );
bli_obj_set_as_root( b_local );
bli_obj_set_as_root( c_local );
// Notice that, unlike trmm_r, there is no dependency in the jc loop
// for trmm3_r, so we can pass in FALSE for jc_dependency.
trmm_thrinfo_t** infos = bli_create_trmm_thrinfo_paths( FALSE );
dim_t n_threads = thread_num_threads( infos[0] );
// Invoke the internal back-end.
bli_level3_thread_decorator( n_threads,
(level3_int_t) bli_trmm_int,
alpha,
&a_local,
&b_local,
beta,
&c_local,
(void*) cntl,
(void**) infos );
bli_trmm_thrinfo_free_paths( infos, n_threads );
}
void bli_syrk_front( obj_t* alpha,
obj_t* a,
obj_t* beta,
obj_t* c,
gemm_t* cntl )
{
obj_t a_local;
obj_t at_local;
obj_t c_local;
// Check parameters.
if ( bli_error_checking_is_enabled() )
bli_syrk_check( alpha, a, beta, c );
// If alpha is zero, scale by beta and return.
if ( bli_obj_equals( alpha, &BLIS_ZERO ) )
{
bli_scalm( beta, c );
return;
}
// Alias A and C in case we need to apply transformations.
bli_obj_alias_to( *a, a_local );
bli_obj_alias_to( *c, c_local );
bli_obj_set_as_root( c_local );
// For syrk, the right-hand "B" operand is simply A^T.
bli_obj_alias_to( *a, at_local );
bli_obj_induce_trans( at_local );
// An optimization: If C is stored by rows and the micro-kernel prefers
// contiguous columns, or if C is stored by columns and the micro-kernel
// prefers contiguous rows, transpose the entire operation to allow the
// micro-kernel to access elements of C in its preferred manner.
if (
( bli_obj_is_row_stored( c_local ) &&
bli_func_prefers_contig_cols( bli_obj_datatype( c_local ),
bli_gemm_cntl_ukrs( cntl ) ) ) ||
( bli_obj_is_col_stored( c_local ) &&
bli_func_prefers_contig_rows( bli_obj_datatype( c_local ),
bli_gemm_cntl_ukrs( cntl ) ) )
)
{
bli_obj_induce_trans( c_local );
}
herk_thrinfo_t** infos = bli_create_herk_thrinfo_paths();
dim_t n_threads = thread_num_threads( infos[0] );
// Invoke the internal back-end.
bli_level3_thread_decorator( n_threads,
(level3_int_t) bli_herk_int,
alpha,
&a_local,
&at_local,
beta,
&c_local,
(void*) cntl,
(void**) infos );
bli_herk_thrinfo_free_paths( infos, n_threads );
}
void bli_symm_front
(
side_t side,
obj_t* alpha,
obj_t* a,
obj_t* b,
obj_t* beta,
obj_t* c,
cntx_t* cntx,
cntl_t* cntl
)
{
obj_t a_local;
obj_t b_local;
obj_t c_local;
// Check parameters.
if ( bli_error_checking_is_enabled() )
bli_symm_check( side, alpha, a, b, beta, c, cntx );
// If alpha is zero, scale by beta and return.
if ( bli_obj_equals( alpha, &BLIS_ZERO ) )
{
bli_scalm( beta, c );
return;
}
// Reinitialize the memory allocator to accommodate the blocksizes
// in the current context.
bli_memsys_reinit( cntx );
// Alias A, B, and C in case we need to apply transformations.
bli_obj_alias_to( *a, a_local );
bli_obj_alias_to( *b, b_local );
bli_obj_alias_to( *c, c_local );
// An optimization: If C is stored by rows and the micro-kernel prefers
// contiguous columns, or if C is stored by columns and the micro-kernel
// prefers contiguous rows, transpose the entire operation to allow the
// micro-kernel to access elements of C in its preferred manner.
if ( bli_cntx_l3_ukr_dislikes_storage_of( &c_local, BLIS_GEMM_UKR, cntx ) )
{
bli_toggle_side( side );
bli_obj_induce_trans( b_local );
bli_obj_induce_trans( c_local );
}
// Swap A and B if multiplying A from the right so that "B" contains
// the symmetric matrix.
if ( bli_is_right( side ) )
{
bli_obj_swap( a_local, b_local );
}
// Set the operation family id in the context.
bli_cntx_set_family( BLIS_GEMM, cntx );
// Record the threading for each level within the context.
bli_cntx_set_thrloop_from_env( BLIS_SYMM, BLIS_LEFT, cntx );
// Invoke the internal back-end.
bli_l3_thread_decorator
(
bli_gemm_int,
alpha,
&a_local,
&b_local,
beta,
&c_local,
cntx,
cntl
);
}
void bli_gemm_front
(
obj_t* alpha,
obj_t* a,
obj_t* b,
obj_t* beta,
obj_t* c,
cntx_t* cntx,
cntl_t* cntl
)
{
#ifdef BLIS_SMALL_MATRIX_ENABLE
#ifndef BLIS_ENABLE_MULTITHREADING
gint_t status = bli_gemm_small_matrix(alpha, a, b, beta, c, cntx, cntl);
if(BLIS_SUCCESS != status)
#endif
#endif
{
obj_t a_local;
obj_t b_local;
obj_t c_local;
// Check parameters.
if ( bli_error_checking_is_enabled() )
bli_gemm_check( alpha, a, b, beta, c, cntx );
// If alpha is zero, scale by beta and return.
if ( bli_obj_equals( alpha, &BLIS_ZERO ) )
{
bli_scalm( beta, c );
return;
}
// Reinitialize the memory allocator to accommodate the blocksizes
// in the current context.
bli_memsys_reinit( cntx );
// Alias A, B, and C in case we need to apply transformations.
bli_obj_alias_to( *a, a_local );
bli_obj_alias_to( *b, b_local );
bli_obj_alias_to( *c, c_local );
// An optimization: If C is stored by rows and the micro-kernel prefers
// contiguous columns, or if C is stored by columns and the micro-kernel
// prefers contiguous rows, transpose the entire operation to allow the
// micro-kernel to access elements of C in its preferred manner.
if ( bli_cntx_l3_ukr_dislikes_storage_of( &c_local, BLIS_GEMM_UKR, cntx ) )
{
bli_obj_swap( a_local, b_local );
bli_obj_induce_trans( a_local );
bli_obj_induce_trans( b_local );
bli_obj_induce_trans( c_local );
}
// Set the operation family id in the context.
bli_cntx_set_family( BLIS_GEMM, cntx );
// Record the threading for each level within the context.
bli_cntx_set_thrloop_from_env( BLIS_GEMM, BLIS_LEFT, cntx,
bli_obj_length( c_local ),
bli_obj_width( c_local ),
bli_obj_width( a_local ) );
// Invoke the internal back-end via the thread handler.
bli_l3_thread_decorator
(
bli_gemm_int,
alpha,
&a_local,
&b_local,
beta,
&c_local,
cntx,
cntl
);
}
}
void bli_herk_front( obj_t* alpha,
obj_t* a,
obj_t* beta,
obj_t* c,
cntx_t* cntx,
gemm_t* cntl )
{
obj_t a_local;
obj_t ah_local;
obj_t c_local;
// Check parameters.
if ( bli_error_checking_is_enabled() )
bli_herk_check( alpha, a, beta, c, cntx );
// If alpha is zero, scale by beta, zero the imaginary components of
// the diagonal elements, and return.
if ( bli_obj_equals( alpha, &BLIS_ZERO ) )
{
bli_scalm( beta, c );
bli_setid( &BLIS_ZERO, c );
return;
}
// Reinitialize the memory allocator to accommodate the blocksizes
// in the current context.
bli_mem_reinit( cntx );
// Alias A and C in case we need to apply transformations.
bli_obj_alias_to( *a, a_local );
bli_obj_alias_to( *c, c_local );
bli_obj_set_as_root( c_local );
// For herk, the right-hand "B" operand is simply A'.
bli_obj_alias_to( *a, ah_local );
bli_obj_induce_trans( ah_local );
bli_obj_toggle_conj( ah_local );
// An optimization: If C is stored by rows and the micro-kernel prefers
// contiguous columns, or if C is stored by columns and the micro-kernel
// prefers contiguous rows, transpose the entire operation to allow the
// micro-kernel to access elements of C in its preferred manner.
if ( bli_cntx_l3_nat_ukr_dislikes_storage_of( &c_local, BLIS_GEMM_UKR, cntx ) )
{
bli_obj_toggle_conj( a_local );
bli_obj_toggle_conj( ah_local );
bli_obj_induce_trans( c_local );
}
thrinfo_t** infos = bli_l3_thrinfo_create_paths( BLIS_HERK, BLIS_LEFT );
dim_t n_threads = bli_thread_num_threads( infos[0] );
// Invoke the internal back-end.
bli_l3_thread_decorator( n_threads,
(l3_int_t) bli_herk_int,
alpha,
&a_local,
&ah_local,
beta,
&c_local,
(void*) cntx,
(void*) cntl,
(void**) infos );
bli_l3_thrinfo_free_paths( infos, n_threads );
// The Hermitian rank-k product was computed as A*A', even for the
// diagonal elements. Mathematically, the imaginary components of
// diagonal elements of a Hermitian rank-k product should always be
// zero. However, in practice, they sometimes accumulate meaningless
// non-zero values. To prevent this, we explicitly set those values
// to zero before returning.
bli_setid( &BLIS_ZERO, &c_local );
}
void bli_syr2k_front
(
obj_t* alpha,
obj_t* a,
obj_t* b,
obj_t* beta,
obj_t* c,
cntx_t* cntx,
cntl_t* cntl
)
{
bli_init_once();
obj_t c_local;
obj_t a_local;
obj_t bt_local;
obj_t b_local;
obj_t at_local;
// Check parameters.
if ( bli_error_checking_is_enabled() )
bli_syr2k_check( alpha, a, b, beta, c, cntx );
// If alpha is zero, scale by beta and return.
if ( bli_obj_equals( alpha, &BLIS_ZERO ) )
{
bli_scalm( beta, c );
return;
}
// Alias A, B, and C in case we need to apply transformations.
bli_obj_alias_to( a, &a_local );
bli_obj_alias_to( b, &b_local );
bli_obj_alias_to( c, &c_local );
bli_obj_set_as_root( &c_local );
// For syr2k, the first and second right-hand "B" operands are simply B'
// and A'.
bli_obj_alias_to( b, &bt_local );
bli_obj_induce_trans( &bt_local );
bli_obj_alias_to( a, &at_local );
bli_obj_induce_trans( &at_local );
// An optimization: If C is stored by rows and the micro-kernel prefers
// contiguous columns, or if C is stored by columns and the micro-kernel
// prefers contiguous rows, transpose the entire operation to allow the
// micro-kernel to access elements of C in its preferred manner.
if ( bli_cntx_l3_ukr_dislikes_storage_of( &c_local, BLIS_GEMM_UKR, cntx ) )
{
bli_obj_induce_trans( &c_local );
}
// Record the threading for each level within the context.
bli_cntx_set_thrloop_from_env( BLIS_SYR2K, BLIS_LEFT, cntx,
bli_obj_length( &c_local ),
bli_obj_width( &c_local ),
bli_obj_width( &a_local ) );
// Invoke herk twice, using beta only the first time.
// Invoke the internal back-end.
bli_l3_thread_decorator
(
bli_gemm_int,
BLIS_HERK, // operation family id
alpha,
&a_local,
&bt_local,
beta,
&c_local,
cntx,
cntl
);
bli_l3_thread_decorator
(
bli_gemm_int,
BLIS_HERK, // operation family id
alpha,
&b_local,
&at_local,
&BLIS_ONE,
&c_local,
cntx,
cntl
);
}
void bli_gemm_int( obj_t* alpha,
obj_t* a,
obj_t* b,
obj_t* beta,
obj_t* c,
gemm_t* cntl,
gemm_thrinfo_t* thread )
{
obj_t a_local;
obj_t b_local;
obj_t c_local;
varnum_t n;
impl_t i;
FUNCPTR_T f;
// Check parameters.
if ( bli_error_checking_is_enabled() )
bli_gemm_int_check( alpha, a, b, beta, c, cntl );
// If C has a zero dimension, return early.
if ( bli_obj_has_zero_dim( *c ) ) return;
// If A or B has a zero dimension, scale C by beta and return early.
if ( bli_obj_has_zero_dim( *a ) ||
bli_obj_has_zero_dim( *b ) )
{
if( thread_am_ochief( thread ) )
bli_scalm( beta, c );
thread_obarrier( thread );
return;
}
// If A or B is marked as being filled with zeros, scale C by beta and
// return early.
if ( bli_obj_is_zeros( *a ) ||
bli_obj_is_zeros( *b ) )
{
if( thread_am_ochief( thread ) )
bli_scalm( beta, c );
thread_obarrier( thread );
return;
}
// Alias A and B in case we need to update attached scalars.
bli_obj_alias_to( *a, a_local );
bli_obj_alias_to( *b, b_local );
// Alias C in case we need to induce a transposition.
bli_obj_alias_to( *c, c_local );
// If we are about to call a leaf-level implementation, and matrix C
// still needs a transposition, then we must induce one by swapping the
// strides and dimensions. Note that this transposition would normally
// be handled explicitly in the packing of C, but if C is not being
// packed, this is our last chance to handle the transposition.
if ( cntl_is_leaf( cntl ) && bli_obj_has_trans( *c ) )
{
//if( thread_am_ochief( thread ) ) {
bli_obj_induce_trans( c_local );
bli_obj_set_onlytrans( BLIS_NO_TRANSPOSE, c_local );
// }
}
// If alpha is non-unit, typecast and apply it to the scalar attached
// to B.
if ( !bli_obj_equals( alpha, &BLIS_ONE ) )
{
bli_obj_scalar_apply_scalar( alpha, &b_local );
}
// If beta is non-unit, typecast and apply it to the scalar attached
// to C.
if ( !bli_obj_equals( beta, &BLIS_ONE ) )
{
bli_obj_scalar_apply_scalar( beta, &c_local );
}
// Extract the variant number and implementation type.
n = cntl_var_num( cntl );
i = cntl_impl_type( cntl );
// Index into the variant array to extract the correct function pointer.
f = vars[n][i];
// Invoke the variant.
f( &a_local,
&b_local,
&c_local,
cntl,
thread );
}
void bli_gemm_front( obj_t* alpha,
obj_t* a,
obj_t* b,
obj_t* beta,
obj_t* c,
gemm_t* cntl )
{
obj_t a_local;
obj_t b_local;
obj_t c_local;
// Check parameters.
if ( bli_error_checking_is_enabled() )
bli_gemm_check( alpha, a, b, beta, c );
// If alpha is zero, scale by beta and return.
if ( bli_obj_equals( alpha, &BLIS_ZERO ) )
{
bli_scalm( beta, c );
return;
}
// Alias A, B, and C in case we need to apply transformations.
bli_obj_alias_to( *a, a_local );
bli_obj_alias_to( *b, b_local );
bli_obj_alias_to( *c, c_local );
// An optimization: If C is stored by rows and the micro-kernel prefers
// contiguous columns, or if C is stored by columns and the micro-kernel
// prefers contiguous rows, transpose the entire operation to allow the
// micro-kernel to access elements of C in its preferred manner.
if (
( bli_obj_is_row_stored( c_local ) &&
bli_func_prefers_contig_cols( bli_obj_datatype( c_local ),
bli_gemm_cntl_ukrs( cntl ) ) ) ||
( bli_obj_is_col_stored( c_local ) &&
bli_func_prefers_contig_rows( bli_obj_datatype( c_local ),
bli_gemm_cntl_ukrs( cntl ) ) )
)
{
bli_obj_swap( a_local, b_local );
bli_obj_induce_trans( a_local );
bli_obj_induce_trans( b_local );
bli_obj_induce_trans( c_local );
}
gemm_thrinfo_t** infos = bli_create_gemm_thrinfo_paths();
dim_t n_threads = thread_num_threads( infos[0] );
// Invoke the internal back-end.
bli_level3_thread_decorator( n_threads,
(level3_int_t) bli_gemm_int,
alpha,
&a_local,
&b_local,
beta,
&c_local,
(void*) cntl,
(void**) infos );
bli_gemm_thrinfo_free_paths( infos, n_threads );
#ifdef BLIS_ENABLE_FLOP_COUNT
// Increment the global flop counter.
bli_flop_count_inc( 2.0 * bli_obj_length( *c )
* bli_obj_width( *c )
* bli_obj_width_after_trans( a_local )
* ( bli_obj_is_complex( *c ) ? 4.0 : 1.0 ) );
#endif
}
void bli_trsm_front( side_t side,
obj_t* alpha,
obj_t* a,
obj_t* b,
cntx_t* cntx,
trsm_t* l_cntl,
trsm_t* r_cntl )
{
trsm_t* cntl;
obj_t a_local;
obj_t b_local;
obj_t c_local;
// Check parameters.
if ( bli_error_checking_is_enabled() )
bli_trsm_check( side, alpha, a, b, &BLIS_ZERO, b, cntx );
// If alpha is zero, scale by beta and return.
if ( bli_obj_equals( alpha, &BLIS_ZERO ) )
{
bli_scalm( alpha, b );
return;
}
// Reinitialize the memory allocator to accommodate the blocksizes
// in the current context.
bli_mem_reinit( cntx );
// Alias A and B so we can tweak the objects if necessary.
bli_obj_alias_to( *a, a_local );
bli_obj_alias_to( *b, b_local );
bli_obj_alias_to( *b, c_local );
// 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 induce a transposition. This
// allows us to only explicitly implement the no-transpose cases. Once
// the transposition is induced, the correct algorithm will be called,
// since, for example, an algorithm over a transposed lower triangular
// matrix A moves in the same direction (forwards) as a non-transposed
// upper triangular matrix. And with the transposition induced, the
// matrix now appears to be upper triangular, so the upper triangular
// algorithm will grab the correct partitions, as if it were upper
// triangular (with no transpose) all along.
if ( bli_obj_has_trans( a_local ) )
{
bli_obj_induce_trans( a_local );
bli_obj_set_onlytrans( BLIS_NO_TRANSPOSE, a_local );
}
#if 0
// If A is being solved against from the right, transpose all operands
// so that we can perform the computation as if A were being solved
// from the left.
if ( bli_is_right( side ) )
{
bli_toggle_side( side );
bli_obj_induce_trans( a_local );
bli_obj_induce_trans( b_local );
bli_obj_induce_trans( c_local );
}
#else
// If A is being solved against from the right, swap A and B so that
// the triangular matrix will actually be on the right.
if ( bli_is_right( side ) )
{
bli_obj_swap( a_local, b_local );
}
#endif
// Set each alias as the root object.
// NOTE: We MUST wait until we are done potentially swapping the objects
// before setting the root fields!
bli_obj_set_as_root( a_local );
bli_obj_set_as_root( b_local );
bli_obj_set_as_root( c_local );
// Choose the control tree.
if ( bli_is_left( side ) ) cntl = l_cntl;
else cntl = r_cntl;
trsm_thrinfo_t** infos = bli_create_trsm_thrinfo_paths( bli_is_right( side ) );
dim_t n_threads = thread_num_threads( infos[0] );
// Invoke the internal back-end.
bli_level3_thread_decorator( n_threads,
(l3_int_t) bli_trsm_int,
alpha,
&a_local,
&b_local,
alpha,
&c_local,
(void*) cntx,
(void*) cntl,
(void**) infos );
bli_trsm_thrinfo_free_paths( infos, n_threads );
}
void bli_trsm_int
(
obj_t* alpha,
obj_t* a,
obj_t* b,
obj_t* beta,
obj_t* c,
cntx_t* cntx,
rntm_t* rntm,
cntl_t* cntl,
thrinfo_t* thread
)
{
obj_t a_local;
obj_t b_local;
obj_t c_local;
trsm_var_oft f;
// Check parameters.
if ( bli_error_checking_is_enabled() )
bli_gemm_basic_check( alpha, a, b, beta, c, cntx );
// If C has a zero dimension, return early.
if ( bli_obj_has_zero_dim( c ) ) return;
// If A or B has a zero dimension, scale C by beta and return early.
if ( bli_obj_has_zero_dim( a ) ||
bli_obj_has_zero_dim( b ) )
{
if ( bli_thread_am_ochief( thread ) )
bli_scalm( beta, c );
bli_thread_obarrier( thread );
return;
}
// Alias A and B in case we need to update attached scalars.
bli_obj_alias_to( a, &a_local );
bli_obj_alias_to( b, &b_local );
// Alias C in case we need to induce a transposition.
bli_obj_alias_to( c, &c_local );
// If we are about to call a leaf-level implementation, and matrix C
// still needs a transposition, then we must induce one by swapping the
// strides and dimensions. Note that this transposition would normally
// be handled explicitly in the packing of C, but if C is not being
// packed, this is our last chance to handle the transposition.
if ( bli_cntl_is_leaf( cntl ) && bli_obj_has_trans( c ) )
{
bli_obj_induce_trans( &c_local );
bli_obj_set_onlytrans( BLIS_NO_TRANSPOSE, &c_local );
}
// If beta is non-unit, apply it to the scalar attached to C.
if ( !bli_obj_equals( beta, &BLIS_ONE ) )
{
bli_obj_scalar_apply_scalar( beta, &c_local );
}
// Set two bools: one based on the implied side parameter (the structure
// of the root object) and one based on the uplo field of the triangular
// matrix's root object (whether that is matrix A or matrix B).
if ( bli_obj_root_is_triangular( a ) )
{
// If alpha is non-unit, typecast and apply it to the scalar
// attached to B (the non-triangular matrix).
if ( !bli_obj_equals( alpha, &BLIS_ONE ) )
{
bli_obj_scalar_apply_scalar( alpha, &b_local );
}
}
else // if ( bli_obj_root_is_triangular( b ) )
{
// If alpha is non-unit, typecast and apply it to the scalar
// attached to A (the non-triangular matrix).
if ( !bli_obj_equals( alpha, &BLIS_ONE ) )
{
bli_obj_scalar_apply_scalar( alpha, &a_local );
}
}
// FGVZ->TMS: Is this barrier still needed?
bli_thread_obarrier( thread );
// Create the next node in the thrinfo_t structure.
bli_thrinfo_grow( rntm, cntl, thread );
// Extract the function pointer from the current control tree node.
f = bli_cntl_var_func( cntl );
// Invoke the variant.
f
(
&a_local,
&b_local,
&c_local,
cntx,
rntm,
cntl,
thread
);
}
void bli_her2k_front
(
obj_t* alpha,
obj_t* a,
obj_t* b,
obj_t* beta,
obj_t* c,
cntx_t* cntx,
cntl_t* cntl
)
{
bli_init_once();
obj_t alpha_conj;
obj_t c_local;
obj_t a_local;
obj_t bh_local;
obj_t b_local;
obj_t ah_local;
// Check parameters.
if ( bli_error_checking_is_enabled() )
bli_her2k_check( alpha, a, b, beta, c, cntx );
// If alpha is zero, scale by beta, zero the imaginary components of
// the diagonal elements, and return.
if ( bli_obj_equals( alpha, &BLIS_ZERO ) )
{
bli_scalm( beta, c );
bli_setid( &BLIS_ZERO, c );
return;
}
// Alias A, B, and C in case we need to apply transformations.
bli_obj_alias_to( a, &a_local );
bli_obj_alias_to( b, &b_local );
bli_obj_alias_to( c, &c_local );
bli_obj_set_as_root( &c_local );
// For her2k, the first and second right-hand "B" operands are simply B'
// and A'.
bli_obj_alias_to( b, &bh_local );
bli_obj_induce_trans( &bh_local );
bli_obj_toggle_conj( &bh_local );
bli_obj_alias_to( a, &ah_local );
bli_obj_induce_trans( &ah_local );
bli_obj_toggle_conj( &ah_local );
// Initialize a conjugated copy of alpha.
bli_obj_scalar_init_detached_copy_of( bli_obj_dt( a ),
BLIS_CONJUGATE,
alpha,
&alpha_conj );
// An optimization: If C is stored by rows and the micro-kernel prefers
// contiguous columns, or if C is stored by columns and the micro-kernel
// prefers contiguous rows, transpose the entire operation to allow the
// micro-kernel to access elements of C in its preferred manner.
if ( bli_cntx_l3_ukr_dislikes_storage_of( &c_local, BLIS_GEMM_UKR, cntx ) )
{
bli_obj_swap( &a_local, &bh_local );
bli_obj_swap( &b_local, &ah_local );
bli_obj_induce_trans( &a_local );
bli_obj_induce_trans( &bh_local );
bli_obj_induce_trans( &b_local );
bli_obj_induce_trans( &ah_local );
bli_obj_induce_trans( &c_local );
}
// Record the threading for each level within the context.
bli_cntx_set_thrloop_from_env( BLIS_HER2K, BLIS_LEFT, cntx,
bli_obj_length( &c_local ),
bli_obj_width( &c_local ),
bli_obj_width( &a_local ) );
// Invoke herk twice, using beta only the first time.
// Invoke the internal back-end.
bli_l3_thread_decorator
(
bli_gemm_int,
BLIS_HERK, // operation family id
alpha,
&a_local,
&bh_local,
beta,
&c_local,
cntx,
cntl
);
bli_l3_thread_decorator
(
bli_gemm_int,
BLIS_HERK, // operation family id
&alpha_conj,
&b_local,
&ah_local,
&BLIS_ONE,
&c_local,
cntx,
cntl
);
// The Hermitian rank-2k product was computed as A*B'+B*A', even for
// the diagonal elements. Mathematically, the imaginary components of
// diagonal elements of a Hermitian rank-2k product should always be
// zero. However, in practice, they sometimes accumulate meaningless
// non-zero values. To prevent this, we explicitly set those values
// to zero before returning.
bli_setid( &BLIS_ZERO, &c_local );
}
void bli_her2k_front( obj_t* alpha,
obj_t* a,
obj_t* b,
obj_t* beta,
obj_t* c,
gemm_t* cntl )
{
obj_t alpha_conj;
obj_t c_local;
obj_t a_local;
obj_t bh_local;
obj_t b_local;
obj_t ah_local;
// Check parameters.
if ( bli_error_checking_is_enabled() )
bli_her2k_check( alpha, a, b, beta, c );
// If alpha is zero, scale by beta, zero the imaginary components of
// the diagonal elements, and return.
if ( bli_obj_equals( alpha, &BLIS_ZERO ) )
{
bli_scalm( beta, c );
bli_setid( &BLIS_ZERO, c );
return;
}
// Alias A, B, and C in case we need to apply transformations.
bli_obj_alias_to( *a, a_local );
bli_obj_alias_to( *b, b_local );
bli_obj_alias_to( *c, c_local );
bli_obj_set_as_root( c_local );
// For her2k, the first and second right-hand "B" operands are simply B'
// and A'.
bli_obj_alias_to( *b, bh_local );
bli_obj_induce_trans( bh_local );
bli_obj_toggle_conj( bh_local );
bli_obj_alias_to( *a, ah_local );
bli_obj_induce_trans( ah_local );
bli_obj_toggle_conj( ah_local );
// Initialize a conjugated copy of alpha.
bli_obj_scalar_init_detached_copy_of( bli_obj_datatype( *a ),
BLIS_CONJUGATE,
alpha,
&alpha_conj );
// An optimization: If C is stored by rows and the micro-kernel prefers
// contiguous columns, or if C is stored by columns and the micro-kernel
// prefers contiguous rows, transpose the entire operation to allow the
// micro-kernel to access elements of C in its preferred manner.
if (
( bli_obj_is_row_stored( c_local ) &&
bli_func_prefers_contig_cols( bli_obj_datatype( c_local ),
bli_gemm_cntl_ukrs( cntl ) ) ) ||
( bli_obj_is_col_stored( c_local ) &&
bli_func_prefers_contig_rows( bli_obj_datatype( c_local ),
bli_gemm_cntl_ukrs( cntl ) ) )
)
{
bli_obj_swap( a_local, bh_local );
bli_obj_swap( b_local, ah_local );
bli_obj_induce_trans( a_local );
bli_obj_induce_trans( bh_local );
bli_obj_induce_trans( b_local );
bli_obj_induce_trans( ah_local );
bli_obj_induce_trans( c_local );
}
#if 0
// Invoke the internal back-end.
bli_her2k_int( alpha,
&a_local,
&bh_local,
&alpha_conj,
&b_local,
&ah_local,
beta,
&c_local,
cntl );
#else
// Invoke herk twice, using beta only the first time.
herk_thrinfo_t** infos = bli_create_herk_thrinfo_paths();
dim_t n_threads = thread_num_threads( infos[0] );
// Invoke the internal back-end.
bli_level3_thread_decorator( n_threads,
(level3_int_t) bli_herk_int,
alpha,
&a_local,
&bh_local,
beta,
&c_local,
(void*) cntl,
(void**) infos );
bli_level3_thread_decorator( n_threads,
(level3_int_t) bli_herk_int,
&alpha_conj,
&b_local,
&ah_local,
&BLIS_ONE,
&c_local,
(void*) cntl,
(void**) infos );
bli_herk_thrinfo_free_paths( infos, n_threads );
#endif
// The Hermitian rank-2k product was computed as A*B'+B*A', even for
// the diagonal elements. Mathematically, the imaginary components of
// diagonal elements of a Hermitian rank-2k product should always be
// zero. However, in practice, they sometimes accumulate meaningless
// non-zero values. To prevent this, we explicitly set those values
// to zero before returning.
bli_setid( &BLIS_ZERO, &c_local );
}
void bli_symm_front( side_t side,
obj_t* alpha,
obj_t* a,
obj_t* b,
obj_t* beta,
obj_t* c,
gemm_t* cntl )
{
obj_t a_local;
obj_t b_local;
obj_t c_local;
// Check parameters.
if ( bli_error_checking_is_enabled() )
bli_symm_check( side, alpha, a, b, beta, c );
// If alpha is zero, scale by beta and return.
if ( bli_obj_equals( alpha, &BLIS_ZERO ) )
{
bli_scalm( beta, c );
return;
}
// Alias A, B, and C in case we need to apply transformations.
bli_obj_alias_to( *a, a_local );
bli_obj_alias_to( *b, b_local );
bli_obj_alias_to( *c, c_local );
// An optimization: If C is stored by rows and the micro-kernel prefers
// contiguous columns, or if C is stored by columns and the micro-kernel
// prefers contiguous rows, transpose the entire operation to allow the
// micro-kernel to access elements of C in its preferred manner.
if (
( bli_obj_is_row_stored( c_local ) &&
bli_func_prefers_contig_cols( bli_obj_datatype( c_local ),
cntl_gemm_ukrs( cntl ) ) ) ||
( bli_obj_is_col_stored( c_local ) &&
bli_func_prefers_contig_rows( bli_obj_datatype( c_local ),
cntl_gemm_ukrs( cntl ) ) )
)
{
bli_toggle_side( side );
bli_obj_induce_trans( b_local );
bli_obj_induce_trans( c_local );
}
// Swap A and B if multiplying A from the right so that "B" contains
// the symmetric matrix.
if ( bli_is_right( side ) )
{
bli_obj_swap( a_local, b_local );
}
gemm_thrinfo_t** infos = bli_create_gemm_thrinfo_paths();
dim_t n_threads = thread_num_threads( infos[0] );
// Invoke the internal back-end.
bli_level3_thread_decorator( n_threads,
(level3_int_t) bli_gemm_int,
alpha,
&a_local,
&b_local,
beta,
&c_local,
(void*) cntl,
(void**) infos );
bli_gemm_thrinfo_free_paths( infos, n_threads );
}
void bli_her2k_front( obj_t* alpha,
obj_t* a,
obj_t* b,
obj_t* beta,
obj_t* c,
herk_t* cntl )
{
obj_t alpha_conj;
obj_t c_local;
obj_t a_local;
obj_t bh_local;
obj_t b_local;
obj_t ah_local;
// Check parameters.
if ( bli_error_checking_is_enabled() )
bli_her2k_check( alpha, a, b, beta, c );
// If alpha is zero, scale by beta and return.
if ( bli_obj_equals( alpha, &BLIS_ZERO ) )
{
bli_scalm( beta, c );
return;
}
// Alias A, B, and C in case we need to apply transformations.
bli_obj_alias_to( *a, a_local );
bli_obj_alias_to( *b, b_local );
bli_obj_alias_to( *c, c_local );
bli_obj_set_as_root( c_local );
// For her2k, the first and second right-hand "B" operands are simply B'
// and A'.
bli_obj_alias_to( *b, bh_local );
bli_obj_induce_trans( bh_local );
bli_obj_toggle_conj( bh_local );
bli_obj_alias_to( *a, ah_local );
bli_obj_induce_trans( ah_local );
bli_obj_toggle_conj( ah_local );
// Initialize a conjugated copy of alpha.
bli_obj_scalar_init_detached_copy_of( bli_obj_datatype( *a ),
BLIS_CONJUGATE,
alpha,
&alpha_conj );
// An optimization: If C is row-stored, transpose the entire operation
// so as to allow the macro-kernel more favorable access patterns
// through C. (The effect of the transposition of A and A' is negligible
// because those operands are always packed to contiguous memory.)
if ( bli_obj_is_row_stored( c_local ) )
{
bli_obj_swap( a_local, bh_local );
bli_obj_swap( b_local, ah_local );
bli_obj_induce_trans( a_local );
bli_obj_induce_trans( bh_local );
bli_obj_induce_trans( b_local );
bli_obj_induce_trans( ah_local );
bli_obj_induce_trans( c_local );
}
#if 0
// Invoke the internal back-end.
bli_her2k_int( alpha,
&a_local,
&bh_local,
&alpha_conj,
&b_local,
&ah_local,
beta,
&c_local,
cntl );
#else
// Invoke herk twice, using beta only the first time.
bli_herk_int( alpha,
&a_local,
&bh_local,
beta,
&c_local,
cntl );
bli_herk_int( &alpha_conj,
&b_local,
&ah_local,
&BLIS_ONE,
&c_local,
cntl );
#endif
}
void bli_ger_int( conj_t conjx,
conj_t conjy,
obj_t* alpha,
obj_t* x,
obj_t* y,
obj_t* a,
cntx_t* cntx,
ger_t* cntl )
{
varnum_t n;
impl_t i;
FUNCPTR_T f;
obj_t alpha_local;
obj_t x_local;
obj_t y_local;
obj_t a_local;
// Check parameters.
if ( bli_error_checking_is_enabled() )
bli_ger_check( alpha, x, y, a );
// If A has a zero dimension, return early.
if ( bli_obj_has_zero_dim( a ) ) return;
// If x or y has a zero dimension, return early.
if ( bli_obj_has_zero_dim( x ) ||
bli_obj_has_zero_dim( y ) ) return;
// Alias the objects, applying conjx and conjy to x and y, respectively.
bli_obj_alias_with_conj( conjx, x, &x_local );
bli_obj_alias_with_conj( conjy, y, &y_local );
bli_obj_alias_to( a, &a_local );
// If matrix A is marked for conjugation, we interpret this as a request
// to apply a conjugation to the other operands.
if ( bli_obj_has_conj( &a_local ) )
{
bli_obj_toggle_conj( &a_local );
bli_obj_toggle_conj( &x_local );
bli_obj_toggle_conj( &y_local );
bli_obj_scalar_init_detached_copy_of( bli_obj_dt( alpha ),
BLIS_CONJUGATE,
alpha,
&alpha_local );
}
else
{
bli_obj_alias_to( *alpha, alpha_local );
}
// If we are about the call a leaf-level implementation, and matrix A
// still needs a transposition, then we must induce one by swapping the
// strides and dimensions.
if ( bli_cntl_is_leaf( cntl ) && bli_obj_has_trans( &a_local ) )
{
bli_obj_induce_trans( &a_local );
bli_obj_set_onlytrans( BLIS_NO_TRANSPOSE, &a_local );
}
// 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[n][i];
// Invoke the variant.
f( &alpha_local,
&x_local,
&y_local,
&a_local,
cntx,
cntl );
}