Esempio n. 1
0
void blx_gemm_int
     (
       obj_t*  a,
       obj_t*  b,
       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;
	gemm_var_oft f;

	// Alias A, B, and C in case we need to update attached scalars.
	bli_obj_alias_to( a, &a_local );
	bli_obj_alias_to( b, &b_local );
	bli_obj_alias_to( c, &c_local );

	// 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
	);
}
Esempio n. 2
0
void bli_gemm_int
     (
       obj_t*  alpha,
       obj_t*  a,
       obj_t*  b,
       obj_t*  beta,
       obj_t*  c,
       cntx_t* cntx,
       cntl_t* cntl,
       thrinfo_t* thread
     )
{
	obj_t     a_local;
	obj_t     b_local;
	obj_t     c_local;
	gemm_voft 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;
	}

	// 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 ) )
	{
		// This should never execute.
		bli_abort();

        if ( bli_thread_am_ochief( thread ) )
		    bli_scalm( beta, c );
        bli_thread_obarrier( thread );
		return;
	}

	// Alias A, B, and C in case we need to update attached scalars.
	bli_obj_alias_to( *a, a_local );
	bli_obj_alias_to( *b, b_local );
	bli_obj_alias_to( *c, 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 );
	}

	// Create the next node in the thrinfo_t structure.
	bli_thrinfo_grow( cntx, cntl, thread );

	// Extract the function pointer from the current control tree node.
	f = bli_cntl_var_func( cntl );

	// Somewhat hackish support for 3m3, 3m2, and 4m1b method implementations.
	{
		ind_t im = bli_cntx_get_ind_method( cntx );

		if ( im != BLIS_NAT )
		{
			if      ( im == BLIS_3M3  && f == bli_gemm_packa    ) f = bli_gemm3m3_packa;
			else if ( im == BLIS_3M2  && f == bli_gemm_ker_var2 ) f = bli_gemm3m2_ker_var2;
			else if ( im == BLIS_4M1B && f == bli_gemm_ker_var2 ) f = bli_gemm4mb_ker_var2;
		}
	}

	// Invoke the variant.
	f
	(
	  &a_local,
	  &b_local,
	  &c_local,
	  cntx,
	  cntl,
      thread
	);
}
Esempio n. 3
0
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
	);
}