Exemple #1
0
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 );
}
Exemple #2
0
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 );

}
Exemple #3
0
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 );
}
Exemple #4
0
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
	);
}
Exemple #5
0
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 );

}