#include <lapacke.h>
#include "common.h"

Include dependency graph for core_sgbrce.c:

Macros
#define	A(_m, _n) (float *)plasma_geteltaddr(A, ((_m)-1), ((_n)-1), eltsize)
#define	V(_m) &(V[(_m)-1])
#define	TAU(_m) &(TAU[(_m)-1])

Functions
int	CORE_sgbrce (int uplo, int N, PLASMA_desc A, float V, float *TAU, int st, int ed, int eltsize)

Detailed Description

PLASMA core_blas kernel PLASMA is a software package provided by Univ. of Tennessee, Univ. of California Berkeley and Univ. of Colorado Denver

Version:: 2.4.5

Author:: Azzam Haidar

Date:: 2011-05-15 s Tue Nov 22 14:35:23 2011

Definition in file core_sgbrce.c.

Macro Definition Documentation

#define A	(	_m,
		_n
	)	(float *)plasma_geteltaddr(A, ((_m)-1), ((_n)-1), eltsize)

CORE_sgbrce is a kernel that will operate on a region (triangle) of data bounded by st and ed. This kernel apply a right update, create a new nnz, then it eliminate it, and move to the next right update, create a new nnz, eliminate it and so on until finishing. When this is done, it take advantage that data are on cache and will apply the left on the remaining part of this region that has not been updated by the left yet.

Parameters:

[in]	uplo	PlasmaLower: PlasmaUpper:
[in]	N	The order of the matrix A.
[in,out]	A	A pointer to the descriptor of the matrix A.
[out]	V	float array, dimension (N). The scalar elementary reflectors are written in this array. So it is used as a workspace for V at each step of the bulge chasing algorithm.
[out]	TAU	float array, dimension (N). The scalar factors of the elementary reflectors are written in thisarray. So it is used as a workspace for TAU at each step of the bulge chasing algorithm.
[in]	st	A pointer to the start index where this kernel will operate.
[in]	ed	A pointer to the end index where this kernel will operate.
[in]	eltsize	PLASMA internal value which refer to the size of the precision.

Returns:

Return values:

PLASMA_SUCCESS	successful exit
<0	if -i, the i-th argument had an illegal value TYPE 1-BDL Householder add -1 because of C

Definition at line 72 of file core_sgbrce.c.

#define TAU ( _m ) &(TAU[(_m)-1])

Definition at line 74 of file core_sgbrce.c.

#define V ( _m ) &(V[(_m)-1])

Definition at line 73 of file core_sgbrce.c.

Function Documentation

int CORE_sgbrce	(	int	uplo,
		int	N,
		PLASMA_desc *	A,
		float *	V,
		float *	TAU,
		int	st,
		int	ed,
		int	eltsize
	)

Definition at line 76 of file core_sgbrce.c.

References A, CORE_slarfx2(), coreblas_error, ELTLDD, max, plasma_desc_t::mb, min, PLASMA_SUCCESS, PlasmaLeft, PlasmaLower, PlasmaRight, TAU, and V.

{
    int    NB, J1, J2, J3, KDM2, len, pt;
    int    len1, len2, t1ed, t2st;
    int    i;
    static float zzero = 0.0;
    PLASMA_desc vA=*A;
    /* Check input arguments */
    if (N < 0) {
        coreblas_error(2, "Illegal value of N");
        return -2;
    }
    if (ed <= st) {
        coreblas_error(6, "Illegal value of st and ed (internal)");
        return -6;
    }
    /* Quick return */
    if (N == 0)
        return PLASMA_SUCCESS;
    NB = A->mb;
    KDM2 = A->mb-2;
    if( uplo == PlasmaLower ){
        /* ========================
         *       LOWER CASE
         * ========================*/
        for (i = ed; i >= st+1 ; i--){
            /* apply Householder from the right. and create newnnz outside the band if J3 < N */
            J1  = ed+1;
            J2  = min((i+1+KDM2), N);
            J3  = min((J2+1), N);
            len = J3-J1+1;
            if(J3>J2)*A(J3,(i-1))=zzero;/* could be removed because A is supposed to be band.*/
            t1ed  = (J3/NB)*NB;
            t2st  = max(t1ed+1,J1);
            len1  = t1ed-J1+1; 
            len2  = J3-t2st+1;
            if(len1>0)CORE_slarfx2(PlasmaRight, len1, (*V(i)), (*TAU(i)), A(J1,  i-1), ELTLDD(vA, J1)  , A(J1  , i), ELTLDD(vA, J1)  );
            if(len2>0)CORE_slarfx2(PlasmaRight, len2, (*V(i)), (*TAU(i)), A(t2st,i-1), ELTLDD(vA, t2st), A(t2st, i), ELTLDD(vA, t2st));
            len    = J3-J2; 
            if(len>0){
                /* generate Householder to annihilate a(j+kd,j-1) within the band */
                *V(J3)         = *A(J3,(i-1));
                *A(J3,(i-1))   = 0.0;
                LAPACKE_slarfg_work( 2, A(J2,(i-1)), V(J3), 1, TAU(J3));
            }
        }
        /* APPLY LEFT ON THE REMAINING ELEMENT OF KERNEL 2 */
        for (i = ed; i >= st+1 ; i--){
            J2  = min((i+1+KDM2), N);
            J3  = min((J2+1), N);
            len    = J3-J2;
            if(len>0){
                pt    = J2;
                J1    = i;
                J2    = min(ed,N);
                t1ed  = (J2/NB)*NB;
                t2st  = max(t1ed+1,J1);
                len1  = t1ed-J1+1; 
                len2  = J2-t2st+1;
                if(len1>0)CORE_slarfx2(PlasmaLeft, len1 , *V(J3), (*TAU(J3)), A(pt, i   ), ELTLDD(vA, pt),  A((pt+1),  i  ), ELTLDD(vA, pt+1) );
                if(len2>0)CORE_slarfx2(PlasmaLeft, len2 , *V(J3), (*TAU(J3)), A(pt, t2st), ELTLDD(vA, pt),  A((pt+1), t2st), ELTLDD(vA, pt+1) );
            }
        }
    } else {
        /* ========================
         *       UPPER CASE
         * ========================*/
        for (i = ed; i >= st+1 ; i--){
            /* apply Householder from the right. and create newnnz outside the band if J3 < N */
            J1  = ed+1;
            J2  = min((i+1+KDM2), N);
            J3  = min((J2+1), N);
            len = J3-J1+1;
            if(J3>J2)*A((i-1), J3)=zzero;
            t1ed  = (J3/NB)*NB;
            t2st  = max(t1ed+1,J1);
            len1  = t1ed-J1+1; 
            len2  = J3-t2st+1;
            if(len1>0)CORE_slarfx2(PlasmaLeft, len1 , *V(i), (*TAU(i)), A(i-1, J1  ), ELTLDD(vA, i-1),  A(i,  J1 ), ELTLDD(vA, i) );
            if(len2>0)CORE_slarfx2(PlasmaLeft, len2 , *V(i), (*TAU(i)), A(i-1, t2st), ELTLDD(vA, i-1),  A(i, t2st), ELTLDD(vA, i) );
            /* if nonzero element a(j+kd,j-1) has been created outside the band (if index < N) then eliminate it. */
            len    = J3-J2; 
            if(len>0){
                /* generate Householder to annihilate a(j+kd,j-1) within the band */
                *V(J3)         = *A(i-1, J3);
                *A(i-1, J3)  = 0.0;
                LAPACKE_slarfg_work( 2, A(i-1, J2), V(J3), 1, TAU(J3));
            }
        }
        /* APPLY RIGHT ON THE REMAINING ELEMENT OF KERNEL 2 */
        for (i = ed; i >= st+1 ; i--){
            /* find if there was a nnz created. if yes apply right else nothing to be done. */
            J2  = min((i+1+KDM2), N);
            J3  = min((J2+1), N);
            len    = J3-J2;
            if(len>0){
                pt    = J2;
                J1    = i;
                J2    = min(ed,N);
                t1ed  = (J2/NB)*NB;
                t2st  = max(t1ed+1,J1);
                len1  = t1ed-J1+1; 
                len2  = J2-t2st+1;
                if(len1>0)CORE_slarfx2(PlasmaRight, len1 , (*V(J3)), (*TAU(J3)), A(i   , pt), ELTLDD(vA, i),     A(i,    pt+1), ELTLDD(vA, i) );
                if(len2>0)CORE_slarfx2(PlasmaRight, len2 , (*V(J3)), (*TAU(J3)), A(t2st, pt), ELTLDD(vA, t2st),  A(t2st, pt+1), ELTLDD(vA, t2st) );
            }
        }
    } /* end of else for the upper case */
    return PLASMA_SUCCESS;
}

Here is the call graph for this function:

Here is the caller graph for this function:

Macros

Functions

Detailed Description

Macro Definition Documentation

Function Documentation