core_scasum.c File Reference

#include <cblas.h>
#include <math.h>
#include "common.h"

Include dependency graph for core_scasum.c:

Go to the source code of this file.

Functions
void	CORE_scasum (int storev, int uplo, int M, int N, PLASMA_Complex32_t *A, int lda, float *work)
void	QUARK_CORE_scasum (Quark *quark, Quark_Task_Flags *task_flags, PLASMA_enum storev, PLASMA_enum uplo, int M, int N, PLASMA_Complex32_t *A, int lda, int szeA, float *work, int szeW)
void	CORE_scasum_quark (Quark *quark)
void	QUARK_CORE_scasum_f1 (Quark *quark, Quark_Task_Flags *task_flags, PLASMA_enum storev, PLASMA_enum uplo, int M, int N, PLASMA_Complex32_t *A, int lda, int szeA, float *work, int szeW, float *fake, int szeF)
void	CORE_scasum_f1_quark (Quark *quark)

---

Function Documentation

void CORE_scasum	(	int	storev,
		int	uplo,
		int	M,
		int	N,
		PLASMA_Complex32_t *	A,
		int	lda,
		float *	work
	)

Definition at line 28 of file core_scasum.c.

References cabsf(), PlasmaColumnwise, PlasmaLower, PlasmaUpper, PlasmaUpperLower, and sum().

{
    PLASMA_Complex32_t *tmpA;
    float *tmpW, sum, abs;
    int i,j;

    switch (uplo) {
    case PlasmaUpper:
        for (j = 0; j < N; j++) {
            tmpA = A+(j*lda);
            sum = 0.0;
            for (i = 0; i < j; i++) {
                abs      = cabsf(*tmpA);
                sum     += abs;
                work[i] += abs;
                tmpA++;
            }
            work[j] += sum + cabsf(*tmpA);
        }
        break;
    case PlasmaLower:
        for (j = 0; j < N; j++) {
            tmpA = A+(j*lda)+j;

            sum = 0.0;
            work[j] += cabsf(*tmpA);

            tmpA++;
            for (i = j+1; i < M; i++) {
                abs      = cabsf(*tmpA);
                sum     += abs;
                work[i] += abs;
                tmpA++;
            }
            work[j] += sum;
        }
        break;
    case PlasmaUpperLower:
    default:
        if (storev == PlasmaColumnwise) {
            for (j = 0; j < N; j++) {
                /* work[j] += cblas_scasum(M, &(A[j*lda]), 1); */
                tmpA = A+(j*lda);
                for (i = 0; i < M; i++) {
                    work[j] +=  cabsf(*tmpA);
                    tmpA++;
                }
            }
        }
        else {
            for (j = 0; j < N; j++) {
                tmpA = A+(j*lda);
                tmpW = work;
                for (i = 0; i < M; i++) {
                    /* work[i] += cabsf( A[j*lda+i] );*/
                    *tmpW += cabsf( *tmpA );
                    tmpA++; tmpW++;
                }
            }
        }
    }
}

Here is the call graph for this function:

Here is the caller graph for this function:

void CORE_scasum_f1_quark

(

Quark *

quark

)

Definition at line 162 of file core_scasum.c.

References A, CORE_scasum(), quark_unpack_args_8, storev, and uplo.

{
    int storev;
    int uplo;
    int M;
    int N;
    PLASMA_Complex32_t *A;
    int lda;
    float *work;
    float *fake;

    quark_unpack_args_8(quark, storev, uplo, M, N, A, lda, work, fake);
    CORE_scasum(storev, uplo, M, N, A, lda, work);
}

Here is the call graph for this function:

Here is the caller graph for this function:

void CORE_scasum_quark

(

Quark *

quark

)

---

Declarations of QUARK wrappers (called by QUARK) - alphabetical order

Definition at line 119 of file core_scasum.c.

References A, CORE_scasum(), quark_unpack_args_7, storev, and uplo.

{
    int storev;
    int uplo;
    int M;
    int N;
    PLASMA_Complex32_t *A;
    int lda;
    float *work;

    quark_unpack_args_7(quark, storev, uplo, M, N, A, lda, work);
    CORE_scasum(storev, uplo, M, N, A, lda, work);
}

Here is the call graph for this function:

Here is the caller graph for this function:

void QUARK_CORE_scasum	(	Quark *	quark,
		Quark_Task_Flags *	task_flags,
		PLASMA_enum	storev,
		PLASMA_enum	uplo,
		int	m,
		int	n,
		PLASMA_Complex32_t *	A,
		int	lda,
		int	szeA,
		float *	work,
		int	szeW
	)

---

Declarations of QUARK wrappers (called by PLASMA) - alphabetical order

Definition at line 95 of file core_scasum.c.

References CORE_scasum_quark(), INOUT, INPUT, QUARK_Insert_Task(), and VALUE.

{
    QUARK_Insert_Task(
        quark, CORE_scasum_quark, task_flags,
        sizeof(PLASMA_enum),                &storev,    VALUE,
        sizeof(PLASMA_enum),                &uplo,      VALUE,
        sizeof(int),                        &M,         VALUE,
        sizeof(int),                        &N,         VALUE,
        sizeof(PLASMA_Complex32_t)*szeA,     A,                 INPUT,
        sizeof(int),                        &lda,       VALUE,
        sizeof(float)*szeW,                 work,              INOUT,
        0);
}

Here is the call graph for this function:

void QUARK_CORE_scasum_f1	(	Quark *	quark,
		Quark_Task_Flags *	task_flags,
		PLASMA_enum	storev,
		PLASMA_enum	uplo,
		int	M,
		int	N,
		PLASMA_Complex32_t *	A,
		int	lda,
		int	szeA,
		float *	work,
		int	szeW,
		float *	fake,
		int	szeF
	)

Definition at line 136 of file core_scasum.c.

References CORE_scasum_f1_quark(), DAG_CORE_ASUM, GATHERV, INOUT, INPUT, OUTPUT, QUARK_Insert_Task(), and VALUE.

{
    DAG_CORE_ASUM;
    QUARK_Insert_Task(
        quark, CORE_scasum_f1_quark, task_flags,
        sizeof(PLASMA_enum),                &storev,    VALUE,
        sizeof(PLASMA_enum),                &uplo,      VALUE,
        sizeof(int),                        &M,         VALUE,
        sizeof(int),                        &N,         VALUE,
        sizeof(PLASMA_Complex32_t)*szeA,     A,                 INPUT,
        sizeof(int),                        &lda,       VALUE,
        sizeof(float)*szeW,                 work,              INOUT,
        sizeof(float)*szeF,                 fake,              OUTPUT | GATHERV,
        0);
}

Here is the call graph for this function:

Here is the caller graph for this function: