pchegst.c File Reference

#include "common.h"

Include dependency graph for pchegst.c:

Go to the source code of this file.

Macros
#define	A(m, n)   BLKADDR(A, PLASMA_Complex32_t, m, n)
#define	B(m, n)   BLKADDR(B, PLASMA_Complex32_t, m, n)

Functions
void	plasma_pchegst_quark (PLASMA_enum itype, PLASMA_enum uplo, PLASMA_desc A, PLASMA_desc B, PLASMA_sequence *sequence, PLASMA_request *request)

---

Detailed Description

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

Version:

2.4.5

Author:

Hatem Ltaief

Date:

2010-11-15 c Tue Nov 22 14:35:43 2011

Definition in file pchegst.c.

---

Macro Definition Documentation

#define A	(		m,
			n
	)		BLKADDR(A, PLASMA_Complex32_t, m, n)

Definition at line 17 of file pchegst.c.

#define B	(		m,
			n
	)		BLKADDR(B, PLASMA_Complex32_t, m, n)

Definition at line 18 of file pchegst.c.

---

Function Documentation

void plasma_pchegst_quark	(	PLASMA_enum	itype,
		PLASMA_enum	uplo,
		PLASMA_desc	A,
		PLASMA_desc	B,
		PLASMA_sequence *	sequence,
		PLASMA_request *	request
	)

Parallel Transformation to standard eigenvalue problem - dynamic scheduler

Definition at line 22 of file pchegst.c.

References B, BLKLDD, plasma_desc_t::n, plasma_desc_t::nb, plasma_desc_t::nt, plasma_context_self(), plasma_desc_submatrix(), plasma_pchemm_quark(), plasma_pcher2k_quark(), plasma_pctrmm_quark(), plasma_pctrsm_quark(), PLASMA_SUCCESS, PlasmaConjTrans, PlasmaLeft, PlasmaLower, PlasmaNonUnit, PlasmaNoTrans, PlasmaRight, plasma_context_struct::quark, QUARK_CORE_chegst(), plasma_sequence_t::quark_sequence, QUARK_Task_Flag_Set(), Quark_Task_Flags_Initializer, plasma_sequence_t::status, and TASK_SEQUENCE.

{
    plasma_context_t *plasma;
    Quark_Task_Flags task_flags = Quark_Task_Flags_Initializer;

    int k;
    int ldak, ldbk;
    int tempkn;
    static float done = 1.0;
    static PLASMA_Complex32_t zone = 1.0;
    static PLASMA_Complex32_t mzone = -1.0;
    static PLASMA_Complex32_t zhalf = 0.5;
    static PLASMA_Complex32_t mzhalf = -0.5;

    plasma = plasma_context_self();
    if (sequence->status != PLASMA_SUCCESS)
        return;

    QUARK_Task_Flag_Set(&task_flags, TASK_SEQUENCE, (intptr_t)sequence->quark_sequence);

    if (itype == 1) {
       if (uplo == PlasmaLower) {
          for (k = 0; k < A.nt; k++){
              tempkn = k == A.nt-1 ? A.n-k*A.nb : A.nb;
              ldak = BLKLDD(A, k);
              ldbk = BLKLDD(B, k);

              QUARK_CORE_chegst(
                  plasma->quark, &task_flags,
                  itype, uplo, tempkn, 
                  A(k, k), ldak,
                  B(k, k), ldbk, 
                  sequence, request, A.nb*k);

              if (k*A.nb+tempkn < A.n) {
                  plasma_pctrsm_quark(
                      PlasmaRight, uplo, 
                      PlasmaConjTrans, PlasmaNonUnit, 
                      zone,
                      plasma_desc_submatrix(B, k*B.nb, k*B.nb, tempkn, tempkn),
                      plasma_desc_submatrix(A, k*A.nb+tempkn, k*A.nb, A.n-k*A.nb-tempkn, tempkn),
                      sequence, request);
                 
                  plasma_pchemm_quark(
                      PlasmaRight, uplo, mzhalf,
                      plasma_desc_submatrix(A, k*A.nb, k*A.nb, tempkn, tempkn),
                      plasma_desc_submatrix(B, k*B.nb+tempkn, k*B.nb, B.n-k*B.nb-tempkn, tempkn),
                      zone,
                      plasma_desc_submatrix(A, k*A.nb+tempkn, k*A.nb, A.n-k*A.nb-tempkn, tempkn),
                      sequence, request);

                  plasma_pcher2k_quark(
                      uplo, PlasmaNoTrans,
                      mzone,
                      plasma_desc_submatrix(A, k*A.nb+tempkn, k*A.nb, A.n-k*A.nb-tempkn, tempkn),
                      plasma_desc_submatrix(B, k*B.nb+tempkn, k*B.nb, B.n-k*B.nb-tempkn, tempkn),
                      done,
                      plasma_desc_submatrix(A, k*A.nb+tempkn, k*A.nb+tempkn, A.n-k*A.nb-tempkn,  A.n-k*A.nb-tempkn),
                      sequence, request);

                  plasma_pchemm_quark(
                      PlasmaRight, uplo,
                      mzhalf,
                      plasma_desc_submatrix(A, k*A.nb, k*A.nb, tempkn, tempkn),
                      plasma_desc_submatrix(B, k*B.nb+tempkn, k*B.nb, B.n-k*B.nb-tempkn, tempkn),
                      zone,
                      plasma_desc_submatrix(A, k*A.nb+tempkn, k*A.nb, A.n-k*A.nb-tempkn, tempkn),
                      sequence, request);

                  plasma_pctrsm_quark(
                      PlasmaLeft, uplo,
                      PlasmaNoTrans, PlasmaNonUnit,
                      zone,
                      plasma_desc_submatrix(B, k*B.nb+tempkn, k*B.nb+tempkn, tempkn, tempkn),
                      plasma_desc_submatrix(A, k*A.nb+tempkn, k*A.nb, A.n-k*A.nb-tempkn, tempkn),
                      sequence, request);
              }
          }
       }
       else {
          for (k = 0; k < A.nt; k++){
              tempkn = k == A.nt-1 ? A.n-k*A.nb : A.nb;
              ldak = BLKLDD(A, k);
              ldbk = BLKLDD(B, k);
              QUARK_CORE_chegst(
                  plasma->quark, &task_flags,
                  itype, uplo, tempkn, 
                  A(k, k), ldak,
                  B(k, k), ldbk, 
                  sequence, request, A.nb*k);

              if (k*A.nb+tempkn < A.n) {
                  plasma_pctrsm_quark(
                      PlasmaLeft, uplo, 
                      PlasmaConjTrans, PlasmaNonUnit, 
                      zone,
                      plasma_desc_submatrix(B, k*B.nb, k*B.nb, tempkn, tempkn),
                      plasma_desc_submatrix(A, k*A.nb, k*A.nb+tempkn, tempkn, A.n-k*A.nb-tempkn),
                      sequence, request);
                 
                  plasma_pchemm_quark(
                      PlasmaLeft, uplo, mzhalf,
                      plasma_desc_submatrix(A, k*A.nb, k*A.nb, tempkn, tempkn),
                      plasma_desc_submatrix(B, k*B.nb, k*B.nb+tempkn, tempkn, B.n-k*B.nb-tempkn),
                      zone,
                      plasma_desc_submatrix(A, k*A.nb, k*A.nb+tempkn, tempkn, A.n-k*A.nb-tempkn),
                      sequence, request);
  
                  plasma_pcher2k_quark(
                      uplo, PlasmaConjTrans,
                      mzone,
                      plasma_desc_submatrix(A, k*A.nb, k*A.nb+tempkn, tempkn, A.n-k*A.nb-tempkn),
                      plasma_desc_submatrix(B, k*B.nb, k*B.nb+tempkn, tempkn, B.n-k*B.nb-tempkn),
                      done,
                      plasma_desc_submatrix(A, k*A.nb+tempkn, k*A.nb+tempkn, A.n-k*A.nb-tempkn,  A.n-k*A.nb-tempkn),
                      sequence, request);

                  plasma_pchemm_quark(
                      PlasmaLeft, uplo,
                      mzhalf,
                      plasma_desc_submatrix(A, k*A.nb, k*A.nb, tempkn, tempkn),
                      plasma_desc_submatrix(B, k*B.nb, k*B.nb+tempkn, tempkn, B.n-k*B.nb-tempkn),
                      zone,
                      plasma_desc_submatrix(A, k*A.nb, k*A.nb+tempkn, tempkn, A.n-k*A.nb-tempkn),
                      sequence, request);
  
                  plasma_pctrsm_quark(
                      PlasmaRight, uplo,
                      PlasmaNoTrans, PlasmaNonUnit,
                      zone,
                      plasma_desc_submatrix(B, k*B.nb+tempkn, k*B.nb+tempkn, tempkn, tempkn),
                      plasma_desc_submatrix(A, k*A.nb, k*A.nb+tempkn, tempkn, A.n-k*A.nb-tempkn),
                      sequence, request);
              }
          }
       }
    }
    else{
       if (uplo == PlasmaLower) {
          for (k = 0; k < A.nt; k++){
              tempkn = k == A.nt-1 ? A.n-k*A.nb : A.nb;
              ldak = BLKLDD(A, k);
              ldbk = BLKLDD(B, k);

              plasma_pctrmm_quark(
                  PlasmaRight, uplo,
                  PlasmaNoTrans, PlasmaNonUnit,
                  zone,
                  plasma_desc_submatrix(B, 0, 0, k*B.nb, k*B.nb),
                  plasma_desc_submatrix(A, k*A.nb, 0, tempkn, k*A.nb),
                  sequence, request);      

              plasma_pchemm_quark(
                  PlasmaLeft, uplo, zhalf,
                  plasma_desc_submatrix(A, k*A.nb, k*A.nb, tempkn, tempkn),
                  plasma_desc_submatrix(B, k*B.nb, 0, tempkn, k*B.nb),
                  zone,
                  plasma_desc_submatrix(A, k*A.nb, 0, tempkn, k*A.nb),
                  sequence, request);
     
              plasma_pcher2k_quark(
                  uplo, PlasmaConjTrans,
                  zone,
                  plasma_desc_submatrix(A, k*A.nb, 0, tempkn, k*A.nb),
                  plasma_desc_submatrix(B, k*B.nb, 0, tempkn, k*B.nb),
                  done,
                  plasma_desc_submatrix(A, 0, 0, k*A.nb, k*A.nb),
                  sequence, request);
      
              plasma_pchemm_quark(
                  PlasmaLeft, uplo, zhalf,
                  plasma_desc_submatrix(A, k*A.nb, k*A.nb, tempkn, tempkn),
                  plasma_desc_submatrix(B, k*B.nb, 0, tempkn, k*B.nb),
                  zone,
                  plasma_desc_submatrix(A, k*A.nb, 0, tempkn, k*A.nb),
                  sequence, request);
     
              plasma_pctrmm_quark(
                  PlasmaLeft, uplo,
                  PlasmaConjTrans, PlasmaNonUnit,
                  zone,
                  plasma_desc_submatrix(B, k*B.nb, k*B.nb, tempkn, tempkn),
                  plasma_desc_submatrix(A, k*A.nb, 0, tempkn, k*A.nb),
                  sequence, request);
    
              QUARK_CORE_chegst(
                  plasma->quark, &task_flags,
                  itype, uplo, tempkn, 
                  A(k, k), ldak,
                  B(k, k), ldbk, 
                  sequence, request, A.nb*k);
           }
       }
       else {
          for (k = 0; k < A.nt; k++){
              tempkn = k == A.nt-1 ? A.n-k*A.nb : A.nb;
              ldak = BLKLDD(A, k);
              ldbk = BLKLDD(B, k);

              plasma_pctrmm_quark(
                  PlasmaLeft, uplo,
                  PlasmaNoTrans, PlasmaNonUnit,
                  zone,
                  plasma_desc_submatrix(B, 0, 0, k*B.nb, k*B.nb),
                  plasma_desc_submatrix(A, 0, k*A.nb, k*A.nb, tempkn),
                  sequence, request);

              plasma_pchemm_quark(
                  PlasmaRight, uplo, zhalf,
                  plasma_desc_submatrix(A, k*A.nb, k*A.nb, k*A.nb, k*A.nb),
                  plasma_desc_submatrix(B, 0, k*B.nb, k*B.nb, tempkn),
                  zone,
                  plasma_desc_submatrix(A, 0, k*A.nb, k*A.nb, tempkn),
                  sequence, request);
     
              plasma_pcher2k_quark(
                  uplo, PlasmaNoTrans,
                  zone,
                  plasma_desc_submatrix(A, 0, k*A.nb, k*A.nb, tempkn),
                  plasma_desc_submatrix(B, 0, k*B.nb, k*B.nb, tempkn),
                  done,
                  plasma_desc_submatrix(A, 0, 0, k*A.nb, k*A.nb),
                  sequence, request); 

              plasma_pchemm_quark(
                  PlasmaRight, uplo, zhalf,
                  plasma_desc_submatrix(A, k*A.nb, k*A.nb, k*A.nb, k*A.nb),
                  plasma_desc_submatrix(B, 0, k*B.nb, k*B.nb, tempkn),
                  zone,
                  plasma_desc_submatrix(A, 0, k*A.nb, k*A.nb, tempkn),
                  sequence, request);

              plasma_pctrmm_quark(
                  PlasmaRight, uplo,
                  PlasmaConjTrans, PlasmaNonUnit,
                  zone,
                  plasma_desc_submatrix(B, k*B.nb, k*B.nb, tempkn, tempkn),
                  plasma_desc_submatrix(A, 0, k*A.nb, k*A.nb, tempkn),
                  sequence, request); 
    
              QUARK_CORE_chegst(
                  plasma->quark, &task_flags,
                  itype, uplo, tempkn, 
                  A(k, k), ldak,
                  B(k, k), ldbk, 
                  sequence, request, A.nb*k);
          }
       }
    }
}

Here is the call graph for this function: