core_slaswp.c File Reference

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

Include dependency graph for core_slaswp.c:

Go to the source code of this file.

Macros
#define	A(m, n)   BLKADDR(descA, float, m, n)

Functions
void	CORE_slaswp (int N, float *A, int LDA, int I1, int I2, int *IPIV, int INC)
void	QUARK_CORE_slaswp (Quark *quark, Quark_Task_Flags *task_flags, int n, float *A, int lda, int i1, int i2, int *ipiv, int inc)
void	CORE_slaswp_quark (Quark *quark)
void	QUARK_CORE_slaswp_f2 (Quark *quark, Quark_Task_Flags *task_flags, int n, float *A, int lda, int i1, int i2, int *ipiv, int inc, float *fake1, int szefake1, int flag1, float *fake2, int szefake2, int flag2)
void	CORE_slaswp_f2_quark (Quark *quark)
int	CORE_slaswp_ontile (PLASMA_desc descA, int i1, int i2, int *ipiv, int inc)
void	QUARK_CORE_slaswp_ontile (Quark *quark, Quark_Task_Flags *task_flags, PLASMA_desc descA, float *Aij, int i1, int i2, int *ipiv, int inc, float *fakepanel)
void	CORE_slaswp_ontile_quark (Quark *quark)
void	QUARK_CORE_slaswp_ontile_f2 (Quark *quark, Quark_Task_Flags *task_flags, PLASMA_desc descA, float *Aij, int i1, int i2, int *ipiv, int inc, float *fake1, int szefake1, int flag1, float *fake2, int szefake2, int flag2)
void	CORE_slaswp_ontile_f2_quark (Quark *quark)
int	CORE_sswptr_ontile (PLASMA_desc descA, int i1, int i2, int *ipiv, int inc, float *Akk, int ldak)
void	QUARK_CORE_sswptr_ontile (Quark *quark, Quark_Task_Flags *task_flags, PLASMA_desc descA, float *Aij, int i1, int i2, int *ipiv, int inc, float *Akk, int ldak)
void	CORE_sswptr_ontile_quark (Quark *quark)
int	CORE_slaswpc_ontile (PLASMA_desc descA, int i1, int i2, int *ipiv, int inc)
void	QUARK_CORE_slaswpc_ontile (Quark *quark, Quark_Task_Flags *task_flags, PLASMA_desc descA, float *Aij, int i1, int i2, int *ipiv, int inc, float *fakepanel)
void	CORE_slaswpc_ontile_quark (Quark *quark)

---

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:

Mathieu Faverge

Date:

2010-11-15 s Tue Nov 22 14:35:21 2011

Definition in file core_slaswp.c.

---

Macro Definition Documentation

#define A	(		m,
			n
	)		BLKADDR(descA, float, m, n)

Definition at line 18 of file core_slaswp.c.

---

Function Documentation

void CORE_slaswp	(	int	N,
		float *	A,
		int	LDA,
		int	I1,
		int	I2,
		int *	IPIV,
		int	INC
	)

Definition at line 29 of file core_slaswp.c.

{
    LAPACKE_slaswp_work( LAPACK_COL_MAJOR, N, A, LDA, I1, I2, IPIV, INC );
}

void CORE_slaswp_f2_quark

(

Quark *

quark

)

Definition at line 102 of file core_slaswp.c.

References A, and quark_unpack_args_9.

{
    int n, lda, i1, i2, inc;
    int *ipiv;
    float *A;
    void *fake1, *fake2;
    
    quark_unpack_args_9(quark, n, A, lda, i1, i2, ipiv, inc, fake1, fake2);
    LAPACKE_slaswp_work(LAPACK_COL_MAJOR, n, A, lda, i1, i2, ipiv, inc );
}

Here is the caller graph for this function:

int CORE_slaswp_ontile	(	PLASMA_desc	descA,
		int	i1,
		int	i2,
		int *	ipiv,
		int	inc
	)

CORE_slaswp_ontile apply the slaswp function on a matrix stored in tile layout

Parameters:

[in,out]	A	The descriptor of the matrix A to permute.
[in]	i1	The first element of IPIV for which a row interchange will be done.
[in]	i2	The last element of IPIV for which a row interchange will be done.
[in]	ipiv	The pivot indices; Only the element in position i1 to i2 are accessed. The pivot are offset by A.i.
[in]	inc	The increment between successive values of IPIV. If IPIV is negative, the pivots are applied in reverse order.

Definition at line 147 of file core_slaswp.c.

References A, BLKLDD, cblas_sswap(), coreblas_error, plasma_desc_t::i, plasma_desc_t::m, plasma_desc_t::mb, plasma_desc_t::mt, plasma_desc_t::n, plasma_desc_t::nt, and PLASMA_SUCCESS.

{
    int i, j, ip, it;
    float *A1;
    int lda1, lda2;

    /* Change i1 to C notation */
    i1--;
    if ( descA.nt > 1 ) {
        coreblas_error(1, "Illegal value of descA.nt");
        return -1;
    }
    if ( i1 < 0 ) {
        coreblas_error(2, "Illegal value of i1");
        return -2;
    }
    if ( (i2 < i1) || (i2 > descA.m) ) {
        coreblas_error(3, "Illegal value of i2");
        return -3;
    }
    if ( ! ( (i2 - i1 - i1%descA.mb -1) < descA.mb ) ) {
        coreblas_error(2, "Illegal value of i1,i2. They have to be part of the same block.");
        return -3;
    }

    it = i1 / descA.mb;
    if (inc > 0) {
        A1 = A(it, 0);
        lda1 = BLKLDD(descA, 0);

        for (j = i1; j < i2; ++j, ipiv+=inc) {
            ip = (*ipiv) - descA.i - 1;
            if ( ip != j )
            {
                it = ip / descA.mb;
                i  = ip % descA.mb;
                lda2 = BLKLDD(descA, it);
                cblas_sswap(descA.n, A1       + j, lda1,
                                     A(it, 0) + i, lda2 );
            }
        }   
    }
    else
    {
        A1 = A(it, 0);
        lda1 = BLKLDD(descA, descA.mt-1);
        
        i1--;
        ipiv = &ipiv[(1-i2)*inc];
        for (j = i2-1; j > i1; --j, ipiv+=inc) {
            ip = (*ipiv) - descA.i - 1;
            if ( ip != j )
            {
                it = ip / descA.mb;
                i  = ip % descA.mb;
                lda2 = BLKLDD(descA, it);
                cblas_sswap(descA.n, A1       + j, lda1,
                                     A(it, 0) + i, lda2 );
            }
        }
    }

    return PLASMA_SUCCESS;
}

Here is the call graph for this function:

Here is the caller graph for this function:

void CORE_slaswp_ontile_f2_quark

(

Quark *

quark

)

Definition at line 279 of file core_slaswp.c.

References A, CORE_slaswp_ontile(), and quark_unpack_args_8.

{
    int i1, i2, inc;
    int *ipiv;
    float *A;
    PLASMA_desc descA;
    void *fake1, *fake2;

    quark_unpack_args_8(quark, descA, A, i1, i2, ipiv, inc, fake1, fake2);
    CORE_slaswp_ontile(descA, i1, i2, ipiv, inc);
}

Here is the call graph for this function:

Here is the caller graph for this function:

void CORE_slaswp_ontile_quark

(

Quark *

quark

)

Definition at line 238 of file core_slaswp.c.

References A, CORE_slaswp_ontile(), and quark_unpack_args_7.

{
    int i1, i2, inc;
    int *ipiv;
    float *A, *fake;
    PLASMA_desc descA;

    quark_unpack_args_7(quark, descA, A, i1, i2, ipiv, inc, fake);
    CORE_slaswp_ontile(descA, i1, i2, ipiv, inc);
}

Here is the call graph for this function:

Here is the caller graph for this function:

void CORE_slaswp_quark

(

Quark *

quark

)

Definition at line 61 of file core_slaswp.c.

References A, and quark_unpack_args_7.

{
    int n, lda, i1, i2, inc;
    int *ipiv;
    float *A;
    
    quark_unpack_args_7(quark, n, A, lda, i1, i2, ipiv, inc);
    LAPACKE_slaswp_work(LAPACK_COL_MAJOR, n, A, lda, i1, i2, ipiv, inc );
}

Here is the caller graph for this function:

int CORE_slaswpc_ontile	(	PLASMA_desc	descA,
		int	i1,
		int	i2,
		int *	ipiv,
		int	inc
	)

CORE_slaswpc_ontile apply the slaswp function on a matrix stored in tile layout

Parameters:

[in,out]	A	The descriptor of the matrix A to permute.
[in]	i1	The first element of IPIV for which a column interchange will be done.
[in]	i2	The last element of IPIV for which a column interchange will be done.
[in]	ipiv	The pivot indices; Only the element in position i1 to i2 are accessed. The pivot are offset by A.i.
[in]	inc	The increment between successive values of IPIV. If IPIV is negative, the pivots are applied in reverse order.

Definition at line 430 of file core_slaswp.c.

References A, BLKLDD, cblas_sswap(), coreblas_error, plasma_desc_t::j, plasma_desc_t::m, plasma_desc_t::mt, plasma_desc_t::n, plasma_desc_t::nb, and PLASMA_SUCCESS.

{
    int i, j, ip, it;
    float *A1;
    int lda;

    if ( descA.mt > 1 ) {
        coreblas_error(1, "Illegal value of descA.mt");
        return -1;
    }
    if ( i1 < 1 ) {
        coreblas_error(2, "Illegal value of i1");
        return -2;
    }
    if ( (i2 < i1) || (i2 > descA.n) ) {
        coreblas_error(3, "Illegal value of i2");
        return -3;
    }
    if ( ! ( (i2 - i1 - i1%descA.nb -1) < descA.nb ) ) {
        coreblas_error(2, "Illegal value of i1,i2. They have to be part of the same block.");
        return -3;
    }

    lda = BLKLDD(descA, 0);
        
    it = i1 / descA.nb;
    if (inc > 0) {
        A1 = A(0, it);

        for (j = i1-1; j < i2; ++j, ipiv+=inc) {
            ip = (*ipiv) - descA.j - 1;
            if ( ip != j )
            {
                it = ip / descA.nb;
                i  = ip % descA.nb;
                cblas_sswap(descA.m, A1       + j*lda, 1,
                                     A(0, it) + i*lda, 1 );
            }
        }   
    }
    else
    {
        A1 = A(0, it);
        i1 -= 2;
        ipiv = &ipiv[(1-i2)*inc];
        for (j = i2-1; j > i1; --j, ipiv+=inc) {
            ip = (*ipiv) - descA.j - 1;
            if ( ip != j )
            {
                it = ip / descA.nb;
                i  = ip % descA.nb;
                cblas_sswap(descA.m, A1       + j*lda, 1,
                                     A(0, it) + i*lda, 1 );
            }
        }
    }

    return PLASMA_SUCCESS;
}

Here is the call graph for this function:

Here is the caller graph for this function:

void CORE_slaswpc_ontile_quark

(

Quark *

quark

)

Definition at line 516 of file core_slaswp.c.

References A, CORE_slaswpc_ontile(), and quark_unpack_args_7.

{
    int i1, i2, inc;
    int *ipiv;
    float *A, *fake;
    PLASMA_desc descA;

    quark_unpack_args_7(quark, descA, A, i1, i2, ipiv, inc, fake);
    CORE_slaswpc_ontile(descA, i1, i2, ipiv, inc);
}

Here is the call graph for this function:

Here is the caller graph for this function:

int CORE_sswptr_ontile	(	PLASMA_desc	descA,
		int	i1,
		int	i2,
		int *	ipiv,
		int	inc,
		float *	Akk,
		int	ldak
	)

CORE_sswptr_ontile apply the slaswp function on a matrix stored in tile layout, followed by a strsm on the first tile of the panel.

Parameters:

[in,out]	A	The descriptor of the matrix A to permute.
[in]	i1	The first element of IPIV for which a row interchange will be done.
[in]	i2	The last element of IPIV for which a row interchange will be done.
[in]	ipiv	The pivot indices; Only the element in position i1 to i2 are accessed. The pivot are offset by A.i.
[in]	inc	The increment between successive values of IPIV. If IPIV is negative, the pivots are applied in reverse order.

Definition at line 325 of file core_slaswp.c.

References A, BLKLDD, cblas_strsm(), CblasColMajor, CblasLeft, CblasLower, CblasNoTrans, CblasUnit, CORE_slaswp_ontile(), coreblas_error, plasma_desc_t::m, plasma_desc_t::mb, plasma_desc_t::mt, plasma_desc_t::n, plasma_desc_t::nt, and PLASMA_SUCCESS.

{
    float zone  = 1.0;
    int lda;
    int m = descA.mt == 1 ? descA.m : descA.mb;

    if ( descA.nt > 1 ) {
        coreblas_error(1, "Illegal value of descA.nt");
        return -1;
    }
    if ( i1 < 1 ) {
        coreblas_error(2, "Illegal value of i1");
        return -2;
    }
    if ( (i2 < i1) || (i2 > m) ) {
        coreblas_error(3, "Illegal value of i2");
        return -3;
    }

    CORE_slaswp_ontile(descA, i1, i2, ipiv, inc);

    lda = BLKLDD(descA, 0);
    cblas_strsm( CblasColMajor, CblasLeft, CblasLower, 
                 CblasNoTrans, CblasUnit,
                 m, descA.n, (zone), 
                 Akk,     ldak, 
                 A(0, 0), lda );

    return PLASMA_SUCCESS;
}

Here is the call graph for this function:

Here is the caller graph for this function:

void CORE_sswptr_ontile_quark

(

Quark *

quark

)

Definition at line 385 of file core_slaswp.c.

References A, CORE_sswptr_ontile(), and quark_unpack_args_8.

{
    int i1, i2, inc, ldak;
    int *ipiv;
    float *A, *Akk;
    PLASMA_desc descA;

    quark_unpack_args_8(quark, descA, A, i1, i2, ipiv, inc, Akk, ldak);
    CORE_sswptr_ontile(descA, i1, i2, ipiv, inc, Akk, ldak);
}

Here is the call graph for this function:

Here is the caller graph for this function:

void QUARK_CORE_slaswp	(	Quark *	quark,
		Quark_Task_Flags *	task_flags,
		int	n,
		float *	A,
		int	lda,
		int	i1,
		int	i2,
		int *	ipiv,
		int	inc
	)

Definition at line 37 of file core_slaswp.c.

References CORE_slaswp_quark(), DAG_CORE_LASWP, INOUT, INPUT, LOCALITY, QUARK_Insert_Task(), and VALUE.

{
    DAG_CORE_LASWP;
    QUARK_Insert_Task(
        quark, CORE_slaswp_quark, task_flags,
        sizeof(int),                      &n,    VALUE,
        sizeof(float)*lda*n,  A,        INOUT | LOCALITY,
        sizeof(int),                      &lda,  VALUE,
        sizeof(int),                      &i1,   VALUE,
        sizeof(int),                      &i2,   VALUE,
        sizeof(int)*n,                     ipiv,     INPUT,
        sizeof(int),                      &inc,  VALUE,
        0);
}

Here is the call graph for this function:

Here is the caller graph for this function:

void QUARK_CORE_slaswp_f2	(	Quark *	quark,
		Quark_Task_Flags *	task_flags,
		int	n,
		float *	A,
		int	lda,
		int	i1,
		int	i2,
		int *	ipiv,
		int	inc,
		float *	fake1,
		int	szefake1,
		int	flag1,
		float *	fake2,
		int	szefake2,
		int	flag2
	)

Definition at line 74 of file core_slaswp.c.

References CORE_slaswp_f2_quark(), DAG_CORE_LASWP, INOUT, INPUT, LOCALITY, QUARK_Insert_Task(), and VALUE.

{
    DAG_CORE_LASWP;
    QUARK_Insert_Task(
        quark, CORE_slaswp_f2_quark, task_flags,
        sizeof(int),                        &n,     VALUE,
        sizeof(float)*lda*n,    A,         INOUT | LOCALITY,
        sizeof(int),                        &lda,   VALUE,
        sizeof(int),                        &i1,    VALUE,
        sizeof(int),                        &i2,    VALUE,
        sizeof(int)*n,                       ipiv,      INPUT,
        sizeof(int),                        &inc,   VALUE,
        sizeof(float)*szefake1, fake1,     flag1,
        sizeof(float)*szefake2, fake2,     flag2,
        0);
}

Here is the call graph for this function:

Here is the caller graph for this function:

void QUARK_CORE_slaswp_ontile	(	Quark *	quark,
		Quark_Task_Flags *	task_flags,
		PLASMA_desc	descA,
		float *	Aij,
		int	i1,
		int	i2,
		int *	ipiv,
		int	inc,
		float *	fakepanel
	)

Definition at line 214 of file core_slaswp.c.

References CORE_slaswp_ontile_quark(), DAG_CORE_LASWP, INOUT, INPUT, LOCALITY, QUARK_Insert_Task(), and VALUE.

{
    DAG_CORE_LASWP;
    QUARK_Insert_Task(
        quark, CORE_slaswp_ontile_quark, task_flags,
        sizeof(PLASMA_desc),              &descA,     VALUE,
        sizeof(float)*1,      Aij,           INOUT | LOCALITY,
        sizeof(int),                      &i1,        VALUE,
        sizeof(int),                      &i2,        VALUE,
        sizeof(int)*(i2-i1+1)*abs(inc),   ipiv,           INPUT,
        sizeof(int),                      &inc,       VALUE,
        sizeof(float)*1,      fakepanel,     INOUT,
        0);
}

Here is the call graph for this function:

Here is the caller graph for this function:

void QUARK_CORE_slaswp_ontile_f2	(	Quark *	quark,
		Quark_Task_Flags *	task_flags,
		PLASMA_desc	descA,
		float *	Aij,
		int	i1,
		int	i2,
		int *	ipiv,
		int	inc,
		float *	fake1,
		int	szefake1,
		int	flag1,
		float *	fake2,
		int	szefake2,
		int	flag2
	)

Definition at line 252 of file core_slaswp.c.

References CORE_slaswp_ontile_f2_quark(), DAG_CORE_LASWP, INOUT, INPUT, LOCALITY, QUARK_Insert_Task(), and VALUE.

{
    DAG_CORE_LASWP;
    QUARK_Insert_Task(
        quark, CORE_slaswp_ontile_f2_quark, task_flags,
        sizeof(PLASMA_desc),                &descA, VALUE,
        sizeof(float)*1,        Aij,       INOUT | LOCALITY,
        sizeof(int),                        &i1,    VALUE,
        sizeof(int),                        &i2,    VALUE,
        sizeof(int)*(i2-i1+1)*abs(inc),      ipiv,      INPUT,
        sizeof(int),                        &inc,   VALUE,
        sizeof(float)*szefake1, fake1, flag1,
        sizeof(float)*szefake2, fake2, flag2,
        0);
}

Here is the call graph for this function:

Here is the caller graph for this function:

void QUARK_CORE_slaswpc_ontile	(	Quark *	quark,
		Quark_Task_Flags *	task_flags,
		PLASMA_desc	descA,
		float *	Aij,
		int	i1,
		int	i2,
		int *	ipiv,
		int	inc,
		float *	fakepanel
	)

Definition at line 492 of file core_slaswp.c.

References CORE_slaswpc_ontile_quark(), DAG_CORE_LASWP, INOUT, INPUT, LOCALITY, QUARK_Insert_Task(), and VALUE.

{
    DAG_CORE_LASWP;
    QUARK_Insert_Task(
        quark, CORE_slaswpc_ontile_quark, task_flags,
        sizeof(PLASMA_desc),              &descA,     VALUE,
        sizeof(float)*1,      Aij,           INOUT | LOCALITY,
        sizeof(int),                      &i1,        VALUE,
        sizeof(int),                      &i2,        VALUE,
        sizeof(int)*(i2-i1+1)*abs(inc),   ipiv,           INPUT,
        sizeof(int),                      &inc,       VALUE,
        sizeof(float)*1,      fakepanel,     INOUT,
        0);
}

Here is the call graph for this function:

Here is the caller graph for this function:

void QUARK_CORE_sswptr_ontile	(	Quark *	quark,
		Quark_Task_Flags *	task_flags,
		PLASMA_desc	descA,
		float *	Aij,
		int	i1,
		int	i2,
		int *	ipiv,
		int	inc,
		float *	Akk,
		int	ldak
	)

Definition at line 359 of file core_slaswp.c.

References CORE_sswptr_ontile_quark(), DAG_CORE_TRSM, INOUT, INPUT, LOCALITY, QUARK_Insert_Task(), and VALUE.

{
    DAG_CORE_TRSM;
    QUARK_Insert_Task(
        quark, CORE_sswptr_ontile_quark, task_flags,
        sizeof(PLASMA_desc),              &descA, VALUE,
        sizeof(float)*1,      Aij,       INOUT | LOCALITY,
        sizeof(int),                      &i1,    VALUE,
        sizeof(int),                      &i2,    VALUE,
        sizeof(int)*(i2-i1+1)*abs(inc),    ipiv,      INPUT,
        sizeof(int),                      &inc,   VALUE,
        sizeof(float)*ldak,   Akk,       INPUT,
        sizeof(int),                      &ldak,  VALUE,
        0);
}

Here is the call graph for this function:

Here is the caller graph for this function: