C C Parallel Sparse BLAS version 2.2 C (C) Copyright 2006/2007/2008 C Salvatore Filippone University of Rome Tor Vergata C Alfredo Buttari University of Rome Tor Vergata C C Redistribution and use in source and binary forms, with or without C modification, are permitted provided that the following conditions C are met: C 1. Redistributions of source code must retain the above copyright C notice, this list of conditions and the following disclaimer. C 2. Redistributions in binary form must reproduce the above copyright C notice, this list of conditions, and the following disclaimer in the C documentation and/or other materials provided with the distribution. C 3. The name of the PSBLAS group or the names of its contributors may C not be used to endorse or promote products derived from this C software without specific written permission. C C THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS C ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED C TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR C PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE PSBLAS GROUP OR ITS CONTRIBUTORS C BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR C CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF C SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS C INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN C CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) C ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE C POSSIBILITY OF SUCH DAMAGE. C C C SUBROUTINE SCSMM(TRANS,M,N,K,ALPHA,PL,FIDA,DESCRA,A,IA1,IA2, C INFOA,PR,B,LDB,BETA,C,LDC,WORK,LWORK,IERROR) C Purpose C ======= C C Computing matrix-matrix product C C <-- ALPHA PL A PR B + BETA C or C C <-- ALPHA PL At PR B + BETA C C C Parameters C ========== C C TRANS - CHARACTER*1 C On entry TRANS specifies if the routine operates with matrix A C or with the transpose of A as follows: C TRANS = 'N' -> use matrix A C TRANS = 'T' or 'C' -> use A' (transpose of matrix A) C Unchanged on exit. C C M - INTEGER C On entry: number of rows of matrix A (A') and C number of rows of matrix C C Unchanged on exit. C C N - INTEGER C On entry: number of columns of matrix B C and number of columns of matrix C. C Unchanged on exit. C C K - INTEGER C On entry: number of columns of matrix A (A') and C number of rows of matrix B C Unchanged on exit. C C ALPHA - DOUBLE PRECISION C On entry: multiplicative constant. C Unchanged on exit. C C PL - INTEGER array of dimension (M) C On entry PL specifies the row permutation of matrix A C (PL(1) == 0 if no permutation). C Unchanged on exit. C C FIDA - CHARACTER*5 C On entry FIDA defines the format of the input sparse matrix. C Unchanged on exit. C C DESCRA - CHARACTER*1 array of DIMENSION (9) C On entry DESCRA describes the characteristics of the input C sparse matrix. C Unchanged on exit. C C A - DOUBLE PRECISION array of DIMENSION (*) C On entry A specifies the values of the input sparse C matrix. C Unchanged on exit. C C IA1 - INTEGER array of dimension (*) C On entry IA1 holds integer information on input sparse C matrix. Actual information will depend on data format used. C Unchanged on exit. C C IA2 - INTEGER array of dimension (*) C On entry IA2 holds integer information on input sparse C matrix. Actual information will depend on data format used. C Unchanged on exit. C C INFOA - INTEGER array of length 10. C On entry can hold auxiliary information on input matrices C formats or environment of subsequent calls. C Might be changed on exit. C C PR - INTEGER array of dimension (K) C On entry PR specifies the column permutation of matrix A C (PR(1) == 0 if no permutation). C Unchanged on exit. C C B - DOUBLE PRECISION matrix of dimension (LDB,*) C On entry: dense matrix. C Unchanged on exit. C C LDB - INTEGER C On entry: leading dimension of B C Unchanged on exit. C C BETA - DOUBLE PRECISION C On entry: multiplicative constant. C Unchanged on exit. C C C - DOUBLE PRECISION matrix of dimension (LDC,*) C On entry: dense matrix. C On exit is updated with the matrix-matrix product. C C LDC - INTEGER C On entry: leading dimension of C C Unchanged on exit. C C WORK - DOUBLE PRECISION array of dimension (LWORK) C On entry: work area. C On exit INT(WORK(1)) contains the minimum value C for LWORK satisfying SCSMM memory requirements. C C LWORK - INTEGER C On entry LWORK specifies the dimension of WORK C LWORK should be set as follows: C LWORK = (LWORK for DxxxMM) + Pr*K*N + Pl*M*N C where Pr (Pl) = 1 if right (left) permutation has to C be performed, 0 otherwise. C Unchanged on exit. C C IERROR - INTEGER C On exit IERROR contains the value of error flag as follows: C IERROR = 0 no error C IERROR > 0 warning C IERROR < 0 fatal error C C Local Variables C =============== C C LWORKM - INTEGER C Minimum work area dimension for SCSMM C C LWORKB - INTEGER C Work area dimension for matrix B in subroutine DLPUPD C C LWORKC - INTEGER C Work area dimension for matrix C in subroutine DLPUPD C C LWORKS - INTEGER C Work area dimension for subroutine DSWMM C C P - INTEGER C Pointer to work area C C LP - LOGICAL C LP is true if left permutation is required C C RP - LOGICAL C RP is true if right permutation is required C C Notes C ===== C Some tests have shown that it is more efficient to divide the C sparse matrix-dense matrix multiplication step and the dense C matrix permutation step, and it is more efficient to put C together the left permutation and update (C <- xxx + BETA C) C steps. So, the sequence of operations is: C Right permutation DLPUPD C Matrix-Matrix product DSWMM C Left permutation and update DLPUPD C In order to avoid useless memory transfer, the above scheme is C simplified according to whether right and left permutation have to C be performed. If left permutation is not required, the update step C is performed in the sparse matrix-dense matrix multiplication kernel. C C It is not possible to call this subroutine with LWORK=0 to get # C the minimal value for LWORK. This functionality needs a better # C connection with DxxxMM # C C SUBROUTINE SCSMM(TRANS,M,N,K,ALPHA,PL,FIDA,DESCRA,A,IA1,IA2, & INFOA,PR,B,LDB,BETA,C,LDC,WORK,LWORK,IERROR) use psb_const_mod IMPLICIT NONE C .. Scalar Arguments .. INTEGER M,N,K,LDB,LDC,LWORK, IERROR CHARACTER TRANS real(psb_spk_) ALPHA,BETA C .. Array Arguments .. INTEGER IA1(*),IA2(*),INFOA(*),PL(*),PR(*) CHARACTER DESCRA*11, FIDA*5 real(psb_spk_) A(*),B(LDB,*),C(LDC,*),WORK(*) C .. Local Scalars .. INTEGER LWORKM, LWORKB, LWORKC, LWORKS, P, ERR_ACT LOGICAL LP, RP C .. Local Array.. INTEGER INT_VAL(5) CHARACTER*20 NAME CHARACTER*30 STRINGS(2) C .. External Subroutines .. EXTERNAL SSWMM, SLPUPD, SSCAL C .. Executable Statements .. C C Check for argument errors C NAME = 'SCSMM\0' IERROR = 0 CALL FCPSB_ERRACTIONSAVE(ERR_ACT) IF (M.LT.0) THEN IERROR = 10 INT_VAL(1) = 2 INT_VAL(2) = M ELSE IF (K.LT.0) THEN IERROR = 10 INT_VAL(1) = 4 INT_VAL(2) = K ELSE IF (N.LT.0) THEN IERROR = 10 INT_VAL(1) = 3 INT_VAL(2) = N ELSE IF (TRANS.NE.'T' .AND. TRANS.NE.'N' .AND. TRANS.NE.'C') THEN IERROR = 40 INT_VAL(1) = 1 STRINGS(1) = TRANS//'\0' ELSE IF (LDB.LT.K) THEN IERROR = 50 INT_VAL(1) = 15 INT_VAL(2) = 4 INT_VAL(3) = LDB INT_VAL(4) = K ELSE IF (LDC.LT.M) THEN IERROR = 50 INT_VAL(1) = 18 INT_VAL(2) = 2 INT_VAL(3) = LDC INT_VAL(4) = M ENDIF C C Error handling C IF(IERROR.NE.0) THEN CALL FCPSB_ERRPUSH(IERROR,NAME,INT_VAL) GOTO 9999 END IF C C Inizializations C LP = PL(1).NE.0 RP = PR(1).NE.0 LWORKB = K*N LWORKC = M*N LWORKM = 0 IF (RP) LWORKM = LWORKB IF (LP) LWORKM = LWORKM + LWORKC IF (LWORK.LT.LWORKM) THEN IERROR = 60 INT_VAL(1) = 20 INT_VAL(2) = LWORKM INT_VAL(3) = LWORK CALL FCPSB_ERRPUSH(IERROR,NAME,INT_VAL) GOTO 9999 ENDIF LWORKS = LWORK - LWORKM C C Check for M, N, K C IF(M.GT.0 .AND. N.GT.0 .AND. K.EQ.0) THEN C C Only C <-- BETA C required C C CALL SSCAL(M,BETA,C,IONE) ELSE IF(M.LE.0 .OR. N.LE.0 .OR. K.LE.0) THEN GOTO 9998 ENDIF C C Switching on PR and PL C IF (LP .AND. RP) THEN C C Both right and left permutation required C P=LWORKB+1 CALL SLPUPD(K,N,PR,B,LDB,DZERO,WORK,K) CALL SSWMM(TRANS,M,N,K,ALPHA,FIDA,DESCRA,A,IA1,IA2,INFOA, & WORK,K,DZERO,WORK(P),M,WORK(P+LWORKC),LWORKS,IERROR) LWORKS = INT(WORK(P+LWORKC)) IF(IERROR .NE. 0) THEN IERROR=4011 CALL FCPSB_ERRPUSH(IERROR,NAME,INT_VAL) GOTO 9999 ENDIF CALL SLPUPD(M,N,PL,WORK(P),M,BETA,C,LDC) ELSE IF(.NOT.LP .AND. RP) THEN C C Only right permutation required C P=LWORKB+1 CALL SLPUPD(K,N,PR,B,LDB,DZERO,WORK,K) CALL SSWMM(TRANS,M,N,K,ALPHA,FIDA,DESCRA,A,IA1,IA2,INFOA, & WORK,K,BETA,C,LDC,WORK(P),LWORKS,IERROR) LWORKS = INT(WORK(P)) IF(IERROR .NE. 0) THEN IERROR=4011 CALL FCPSB_ERRPUSH(IERROR,NAME,INT_VAL) GOTO 9999 ENDIF ELSE IF(.NOT.RP .AND. LP) THEN C C Only left permutation required C P=LWORKC+1 CALL SSWMM(TRANS,M,N,K,ALPHA,FIDA,DESCRA,A,IA1,IA2,INFOA, & B,LDB,DZERO,WORK,M,WORK(P),LWORKS,IERROR) LWORKS = INT(WORK(P)) IF(IERROR .NE. 0) THEN IERROR=4011 CALL FCPSB_ERRPUSH(IERROR,NAME,INT_VAL) GOTO 9999 ENDIF CALL SLPUPD(M,N,PL,WORK,M,BETA,C,LDC) ELSE IF(.NOT.RP .AND. .NOT.LP) THEN C C No permutations required C CALL SSWMM(TRANS,M,N,K,ALPHA,FIDA,DESCRA,A,IA1,IA2,INFOA, & B,LDB,BETA,C,LDC,WORK,LWORKS,IERROR) LWORKS = INT(WORK(1)) IF(IERROR .NE. 0) THEN IERROR=4011 CALL FCPSB_ERRPUSH(IERROR,NAME,INT_VAL) GOTO 9999 ENDIF ENDIF 9998 CONTINUE C C Return minimum workarea dimension C LWORKM = LWORKM + LWORKS WORK(1) = REAL(LWORKM) CALL FCPSB_ERRACTIONRESTORE(ERR_ACT) RETURN 9999 CONTINUE CALL FCPSB_ERRACTIONRESTORE(ERR_ACT) IF ( ERR_ACT .NE. 0 ) THEN CALL FCPSB_SERROR() RETURN ENDIF RETURN END