C     This program times Fortran variants of the Level 1 BLAS 
C     operations AXPY and DOT, the Level 2 BLAS routine DGEMV,
C     and the level 3 BLAS routine DGEMM.
C
C     It is very simple-minded.
C
C     It was intended for the IBM RS/6000-530, but may
C     provide useful information on similar systems.
C
C     A DOUBLE PRECISION FUNCTION SECOND is used for timing. You
C     will need to implement this on your system.
C
C     If your cache is not 64K, change IWORDS to be the number
C     of 64bit words which can fit in cache. Also, change
C     NACTUAL so that a matrix of size NACTUAL by NACTUAL
C     comfortably fits in cache. 
C
C     Change AVMFLP and AVTIME to be typical values of the megaflop
C     rate you expect, and the number of repetitions of each
C     timing is adjusted so that AVTIME seconds should be
C     accumulated.
C
C     J. Dongarra, P. Mayes, G. Radicati
C     Program grnerates data for a paper entitled:
C     ``The IBM RS/6000 and Linear Algebra Operations'', to appear.
C
C     .. Parameters ..
      INTEGER          IWORDS, IMAX1, IMAX2, MAXN, LDA
      PARAMETER        (IWORDS=8192,IMAX1=IWORDS/8,IMAX2=IWORDS*2,
     +                 MAXN=128,LDA=MAXN+16)
C     .. Local Scalars ..
      DOUBLE PRECISION AVMFLP, AVTIME, OPS, S0, S1, S2, T1, T11, T12,
     +                 T2, T21, T22, T3, T4, T5, T6, T7, T8, TEMP,
     +                 TFLUSH, TIME
      INTEGER          I, IMAX, IREP, J, K, KOUNT, NACTUAL
C     .. Local Arrays ..
C
      DOUBLE PRECISION A(LDA,MAXN), B(LDA,MAXN), B1(16,IMAX2),
     +                 B2(16,IMAX2), C(LDA,MAXN), X(IMAX2), Y(IMAX2)
C     .. External Functions ..
      DOUBLE PRECISION SECOND
      EXTERNAL         SECOND
C     .. External Subroutines ..
      EXTERNAL         FLUSH, INIT
C     .. Intrinsic Functions ..
      INTRINSIC        MAX, DBLE, NINT
C     .. Executable Statements ..
      AVTIME = 10.0D0
      AVMFLP = 10.0D0
C
C     First compute the time for a call to FLUSH
C
      S0 = SECOND()
      S1 = SECOND()
      DO 20 I = 1, 10000
         CALL FLUSH(IWORDS)
   20 CONTINUE
      S2 = SECOND()
      TFLUSH = (S2-S1) - (S1-S0)
      TFLUSH = TFLUSH*1.0D-4
C
C     Initialize the data to values which hopefully won't
C     cause overflow problems.
C
      CALL INIT(A,LDA,MAXN)
      CALL INIT(B,LDA,MAXN)
      CALL INIT(C,LDA,MAXN)
      CALL INIT(B1,16,IMAX)
      CALL INIT(B2,16,IMAX)
      CALL INIT(X,IMAX2,1)
      CALL INIT(Y,IMAX2,1)
      NACTUAL = 48
C
C     DOT: Both vectors in cache
C
      OPS = 2*DBLE(IMAX1)
      IREP = MAX(1,NINT(AVTIME/(OPS/AVMFLP/1.0D6)))
      S0 = SECOND()
      S1 = SECOND()
      TEMP = 0.0D0
      DO 60 KOUNT = 1, IREP
         DO 40 I = 1, IMAX1
            TEMP = TEMP + X(I)*Y(I)
   40    CONTINUE
   60 CONTINUE
      S2 = SECOND()
      TIME = (S2-S1) - (S1-S0)
      WRITE (6,FMT=99999) OPS*IREP/TIME/1.0D6,
     +  'DOT:   Both vectors in cache'
C
C     AXPY: Both vectors in cache
C
      OPS = 2*DBLE(IMAX1)
      IREP = MAX(1,NINT(AVTIME/(OPS/AVMFLP/1.0D6)))
      S0 = SECOND()
      S1 = SECOND()
      TEMP = 0.0D0
      DO 100 KOUNT = 1, IREP
         DO 80 I = 1, IMAX1
            Y(I) = Y(I) + TEMP*X(I)
   80    CONTINUE
  100 CONTINUE
      S2 = SECOND()
      TIME = (S2-S1) - (S1-S0)
      WRITE (6,FMT=99999) OPS*IREP/TIME/1.0D6,
     +  'AXPY:  Both vectors in cache'
C
C     DOT: both from memory with stride 1
C
      OPS = 2*DBLE(IMAX2)
      IREP = MAX(1,NINT(AVTIME/(OPS/AVMFLP/1.0D6)))
      S0 = SECOND()
      S1 = SECOND()
      DO 140 KOUNT = 1, IREP
         TEMP = 0.0D0
         DO 120 I = 1, IMAX2
            TEMP = TEMP + X(I)*Y(I)
  120    CONTINUE
  140 CONTINUE
      S2 = SECOND()
      TIME = (S2-S1) - (S1-S0)
      WRITE (6,FMT=99999) OPS*IREP/TIME/1.0D6,
     +  'DOT:   Both from memory with stride 1'
C
C     DOT: both from memory, x with stride 1, y with stride 16
C
      OPS = 2*DBLE(IMAX2)
      IREP = MAX(1,NINT(AVTIME/(OPS/AVMFLP/1.0D6)))
      S0 = SECOND()
      S1 = SECOND()
      DO 180 KOUNT = 1, IREP
         TEMP = 0.0D0
         DO 160 I = 1, IMAX2
            TEMP = TEMP + X(I)*B1(1,I)
  160    CONTINUE
  180 CONTINUE
      S2 = SECOND()
      TIME = (S2-S1) - (S1-S0)
      WRITE (6,FMT=99999) OPS*IREP/TIME/1.0D6,
     +  'DOT:   Both from memory, x with stride 1, y with stride 16'
C
C     DOT: both from memory, both with stride 16
C
      OPS = 2*DBLE(IMAX2)
      IREP = MAX(1,NINT(AVTIME/(OPS/AVMFLP/1.0D6)))
      S0 = SECOND()
      S1 = SECOND()
      DO 220 KOUNT = 1, IREP
         TEMP = 0.0D0
         DO 200 I = 1, IMAX2
            TEMP = TEMP + B1(1,I)*B2(1,I)
  200    CONTINUE
  220 CONTINUE
      S2 = SECOND()
      TIME = (S2-S1) - (S1-S0)
      WRITE (6,FMT=99999) OPS*IREP/TIME/1.0D6,
     +  'DOT:   Both from memory, both with stride 16'
C
C     AXPY: Both from memory with stride 1
C
      OPS = 2*DBLE(IMAX2)
      IREP = MAX(1,NINT(AVTIME/(OPS/AVMFLP/1.0D6)))
      S0 = SECOND()
      S1 = SECOND()
      TEMP = 0.0D0
      DO 260 KOUNT = 1, IREP
         DO 240 I = 1, IMAX2
            Y(I) = Y(I) + TEMP*X(I)
  240    CONTINUE
  260 CONTINUE
      S2 = SECOND()
      TIME = (S2-S1) - (S1-S0)
      WRITE (6,FMT=99999) OPS*IREP/TIME/1.0D6,
     +  'AXPY:  Both from memory with stride 1'
C
C     AXPY: Both from memory, y with stride 1, x with stride 16
C
      OPS = 2*DBLE(IMAX2)
      IREP = MAX(1,NINT(AVTIME/(OPS/AVMFLP/1.0D6)))
      S0 = SECOND()
      S1 = SECOND()
      DO 300 KOUNT = 1, IREP
         DO 280 I = 1, IMAX2
            Y(I) = Y(I) + TEMP*B1(1,I)
  280    CONTINUE
  300 CONTINUE
      S2 = SECOND()
      TIME = (S2-S1) - (S1-S0)
      WRITE (6,FMT=99999) OPS*IREP/TIME/1.0D6,
     +  'AXPY:  Both from memory, y with stride 1, x with stride 16'
C
C     AXPY: Both from memory, y with stride 16, x with stride 1
C
      OPS = 2*DBLE(IMAX2)
      IREP = MAX(1,NINT(AVTIME/(OPS/AVMFLP/1.0D6)))
      S0 = SECOND()
      S1 = SECOND()
      DO 340 KOUNT = 1, IREP
         DO 320 I = 1, IMAX2
            B1(1,I) = B1(1,I) + TEMP*X(I)
  320    CONTINUE
  340 CONTINUE
      S2 = SECOND()
      TIME = (S2-S1) - (S1-S0)
      WRITE (6,FMT=99999) OPS*IREP/TIME/1.0D6,
     +  'AXPY:  Both from memory, y with stride 16, x with stride 1'
C
C     AXPY: Both from memory, both with stride 16
C
      OPS = 2*DBLE(IMAX2)
      IREP = MAX(1,NINT(AVTIME/(OPS/AVMFLP/1.0D6)))
      S0 = SECOND()
      S1 = SECOND()
      DO 380 KOUNT = 1, IREP
         DO 360 I = 1, IMAX2
            B1(1,I) = B1(1,I) + TEMP*B2(1,I)
  360    CONTINUE
  380 CONTINUE
      S2 = SECOND()
      TIME = (S2-S1) - (S1-S0)
      WRITE (6,FMT=99999) OPS*IREP/TIME/1.0D6,
     +  'AXPY:  Both from memory, both with stride 16'
C
C     DGEMV: In cache, A not transposed, DOT
C
      OPS = 2*DBLE(NACTUAL)**2
      IREP = MAX(1,NINT(AVTIME/(OPS/AVMFLP/1.0D6)))
      S0 = SECOND()
      S1 = SECOND()
      DO 440 KOUNT = 1, IREP
         DO 420 I = 1, NACTUAL
            TEMP = Y(I)
            DO 400 J = 1, NACTUAL
               TEMP = TEMP + A(I,J)*X(J)
  400       CONTINUE
            Y(I) = TEMP
  420    CONTINUE
  440 CONTINUE
      S2 = SECOND()
      TIME = (S2-S1) - (S1-S0)
      WRITE (6,FMT=99999) OPS*IREP/TIME/1.0D6,
     +  'DGEMV: In cache, A not transposed, DOT'
C
C     DGEMV: In cache, A not transposed, DOT unrolled to depth 2
C
      OPS = 2*DBLE(NACTUAL)**2
      IREP = MAX(1,NINT(AVTIME/(OPS/AVMFLP/1.0D6)))
      S0 = SECOND()
      S1 = SECOND()
      DO 500 KOUNT = 1, IREP
         DO 480 I = 1, NACTUAL, 2
            T1 = Y(I)
            T2 = Y(I+1)
            DO 460 J = 1, NACTUAL
               T1 = T1 + A(I,J)*X(J)
               T2 = T2 + A(I+1,J)*X(J)
  460       CONTINUE
            Y(I) = T1
            Y(I+1) = T2
  480    CONTINUE
  500 CONTINUE
      S2 = SECOND()
      TIME = (S2-S1) - (S1-S0)
      WRITE (6,FMT=99999) OPS*IREP/TIME/1.0D6,
     +  'DGEMV: In cache, A not transposed, DOT unrolled to depth 2'
C
C     DGEMV: In cache, A not transposed, DOT unrolled to depth 3
C
      OPS = 2*DBLE(NACTUAL)**2
      IREP = MAX(1,NINT(AVTIME/(OPS/AVMFLP/1.0D6)))
      S0 = SECOND()
      S1 = SECOND()
      DO 560 KOUNT = 1, IREP
         DO 540 I = 1, NACTUAL, 3
            T1 = Y(I)
            T2 = Y(I+1)
            T3 = Y(I+2)
            DO 520 J = 1, NACTUAL
               T1 = T1 + A(I,J)*X(J)
               T2 = T2 + A(I+1,J)*X(J)
               T3 = T3 + A(I+2,J)*X(J)
  520       CONTINUE
            Y(I) = T1
            Y(I+1) = T2
            Y(I+2) = T3
  540    CONTINUE
  560 CONTINUE
      S2 = SECOND()
      TIME = (S2-S1) - (S1-S0)
      WRITE (6,FMT=99999) OPS*IREP/TIME/1.0D6,
     +  'DGEMV: In cache, A not transposed, DOT unrolled to depth 3'
C
C     DGEMV: In cache, A not transposed, DOT unrolled to depth 4
C
      OPS = 2*DBLE(NACTUAL)**2
      IREP = MAX(1,NINT(AVTIME/(OPS/AVMFLP/1.0D6)))
      S0 = SECOND()
      S1 = SECOND()
      DO 620 KOUNT = 1, IREP
         DO 600 I = 1, NACTUAL, 4
            T1 = Y(I)
            T2 = Y(I+1)
            T3 = Y(I+2)
            T4 = Y(I+3)
            DO 580 J = 1, NACTUAL
               T1 = T1 + A(I,J)*X(J)
               T2 = T2 + A(I+1,J)*X(J)
               T3 = T3 + A(I+2,J)*X(J)
               T4 = T4 + A(I+3,J)*X(J)
  580       CONTINUE
            Y(I) = T1
            Y(I+1) = T2
            Y(I+2) = T3
            Y(I+3) = T4
  600    CONTINUE
  620 CONTINUE
      S2 = SECOND()
      TIME = (S2-S1) - (S1-S0)
      WRITE (6,FMT=99999) OPS*IREP/TIME/1.0D6,
     +  'DGEMV: In cache, A not transposed, DOT unrolled to depth 4'
C
C     DGEMV: In cache, A not transposed, AXPY
C
      OPS = 2*DBLE(NACTUAL)**2
      IREP = MAX(1,NINT(AVTIME/(OPS/AVMFLP/1.0D6)))
      S0 = SECOND()
      S1 = SECOND()
      DO 680 KOUNT = 1, IREP
         DO 660 J = 1, NACTUAL
            DO 640 I = 1, NACTUAL
               Y(I) = Y(I) + A(I,J)*X(J)
  640       CONTINUE
  660    CONTINUE
  680 CONTINUE
      S2 = SECOND()
      TIME = (S2-S1) - (S1-S0)
      WRITE (6,FMT=99999) OPS*IREP/TIME/1.0D6,
     +  'DGEMV: In cache, A not transposed, AXPY'
C
C     DGEMV: In cache, A not transposed, AXPY unrolled to depth 2
C
      OPS = 2*DBLE(NACTUAL)**2
      IREP = MAX(1,NINT(AVTIME/(OPS/AVMFLP/1.0D6)))
      S0 = SECOND()
      S1 = SECOND()
      DO 740 KOUNT = 1, IREP
         DO 720 J = 1, NACTUAL, 2
            DO 700 I = 1, NACTUAL
               Y(I) = Y(I) + A(I,J)*X(J) + A(I,J+1)*X(J+1)
  700       CONTINUE
  720    CONTINUE
  740 CONTINUE
      S2 = SECOND()
      TIME = (S2-S1) - (S1-S0)
      WRITE (6,FMT=99999) OPS*IREP/TIME/1.0D6,
     +  'DGEMV: In cache, A not transposed, AXPY unrolled to depth 2'
C
C     DGEMV: In cache, A not transposed, AXPY unrolled to depth 3
C
      OPS = 2*DBLE(NACTUAL)**2
      IREP = MAX(1,NINT(AVTIME/(OPS/AVMFLP/1.0D6)))
      S0 = SECOND()
      S1 = SECOND()
      DO 800 KOUNT = 1, IREP
         DO 780 J = 1, NACTUAL, 3
            DO 760 I = 1, NACTUAL
               Y(I) = Y(I) + A(I,J)*X(J) + A(I,J+1)*X(J+1) + A(I,J+2)
     +                *X(J+2)
  760       CONTINUE
  780    CONTINUE
  800 CONTINUE
      S2 = SECOND()
      TIME = (S2-S1) - (S1-S0)
      WRITE (6,FMT=99999) OPS*IREP/TIME/1.0D6,
     +  'DGEMV: In cache, A not transposed, AXPY unrolled to depth 3'
C
C     DGEMV: In cache, A not transposed, AXPY unrolled to depth 4
C
      OPS = 2*DBLE(NACTUAL)**2
      IREP = MAX(1,NINT(AVTIME/(OPS/AVMFLP/1.0D6)))
      S0 = SECOND()
      S1 = SECOND()
      DO 860 KOUNT = 1, IREP
         DO 840 J = 1, NACTUAL, 4
            DO 820 I = 1, NACTUAL
               Y(I) = Y(I) + A(I,J)*X(J) + A(I,J+1)*X(J+1) + A(I,J+2)
     +                *X(J+2) + A(I,J+3)*X(J+3)
  820       CONTINUE
  840    CONTINUE
  860 CONTINUE
      S2 = SECOND()
      TIME = (S2-S1) - (S1-S0)
      WRITE (6,FMT=99999) OPS*IREP/TIME/1.0D6,
     +  'DGEMV: In cache, A not transposed, AXPY unrolled to depth 4'
C
C     DGEMV: In cache, A**T, DOT
C
      OPS = 2*DBLE(NACTUAL)**2
      IREP = MAX(1,NINT(AVTIME/(OPS/AVMFLP/1.0D6)))
      S0 = SECOND()
      S1 = SECOND()
      DO 920 KOUNT = 1, IREP
         DO 900 I = 1, NACTUAL
            TEMP = Y(I)
            DO 880 J = 1, NACTUAL
               TEMP = TEMP + A(J,I)*X(J)
  880       CONTINUE
            Y(I) = TEMP
  900    CONTINUE
  920 CONTINUE
      S2 = SECOND()
      TIME = (S2-S1) - (S1-S0)
      WRITE (6,FMT=99999) OPS*IREP/TIME/1.0D6,
     +  'DGEMV: In cache, A**T, DOT'
C
C
C     DGEMV: In cache, A**T, DOT unrolled to depth 2
C
      OPS = 2*DBLE(NACTUAL)**2
      IREP = MAX(1,NINT(AVTIME/(OPS/AVMFLP/1.0D6)))
      S0 = SECOND()
      S1 = SECOND()
      DO 980 KOUNT = 1, IREP
         DO 960 I = 1, NACTUAL, 2
            T1 = Y(I)
            T2 = Y(I+1)
            DO 940 J = 1, NACTUAL
               T1 = T1 + A(J,I)*X(J)
               T2 = T2 + A(J,I+1)*X(J)
  940       CONTINUE
            Y(I) = T1
            Y(I+1) = T2
  960    CONTINUE
  980 CONTINUE
      S2 = SECOND()
      TIME = (S2-S1) - (S1-S0)
      WRITE (6,FMT=99999) OPS*IREP/TIME/1.0D6,
     +  'DGEMV: In cache, A**T, DOT unrolled to depth 2'
C
C     DGEMV: In cache, A**T, DOT unrolled to depth 3
C
      OPS = 2*DBLE(NACTUAL)**2
      IREP = MAX(1,NINT(AVTIME/(OPS/AVMFLP/1.0D6)))
      S0 = SECOND()
      S1 = SECOND()
      DO 1040 KOUNT = 1, IREP
         DO 1020 I = 1, NACTUAL, 3
            T1 = Y(I)
            T2 = Y(I+1)
            T3 = Y(I+2)
            DO 1000 J = 1, NACTUAL
               T1 = T1 + A(J,I)*X(J)
               T2 = T2 + A(J,I+1)*X(J)
               T3 = T3 + A(J,I+2)*X(J)
 1000       CONTINUE
            Y(I) = T1
            Y(I+1) = T2
            Y(I+2) = T3
 1020    CONTINUE
 1040 CONTINUE
      S2 = SECOND()
      TIME = (S2-S1) - (S1-S0)
      WRITE (6,FMT=99999) OPS*IREP/TIME/1.0D6,
     +  'DGEMV: In cache, A**T, DOT unrolled to depth 3'
C
C     DGEMV: In cache, A**T, DOT unrolled to depth 4
C
      OPS = 2*DBLE(NACTUAL)**2
      IREP = MAX(1,NINT(AVTIME/(OPS/AVMFLP/1.0D6)))
      S0 = SECOND()
      S1 = SECOND()
      DO 1100 KOUNT = 1, IREP
         DO 1080 I = 1, NACTUAL, 4
            T1 = Y(I)
            T2 = Y(I+1)
            T3 = Y(I+2)
            T4 = Y(I+3)
            DO 1060 J = 1, NACTUAL
               T1 = T1 + A(J,I)*X(J)
               T2 = T2 + A(J,I+1)*X(J)
               T3 = T3 + A(J,I+2)*X(J)
               T4 = T4 + A(J,I+3)*X(J)
 1060       CONTINUE
            Y(I) = T1
            Y(I+1) = T2
            Y(I+2) = T3
            Y(I+3) = T4
 1080    CONTINUE
 1100 CONTINUE
      S2 = SECOND()
      TIME = (S2-S1) - (S1-S0)
      WRITE (6,FMT=99999) OPS*IREP/TIME/1.0D6,
     +  'DGEMV: In cache, A**T, DOT unrolled to depth 4'
C
C     DGEMV: In cache, A**T, DOT unrolled to depth 4
C
      OPS = 2*DBLE(NACTUAL)**2
      IREP = MAX(1,NINT(AVTIME/(OPS/AVMFLP/1.0D6)))
      S0 = SECOND()
      S1 = SECOND()
      DO 1160 KOUNT = 1, IREP
         DO 1140 I = 1, NACTUAL, 8
            T1 = Y(I)
            T2 = Y(I+1)
            T3 = Y(I+2)
            T4 = Y(I+3)
            T5 = Y(I+4)
            T6 = Y(I+5)
            T7 = Y(I+6)
            T8 = Y(I+7)
            DO 1120 J = 1, NACTUAL
               T1 = T1 + A(J,I)*X(J)
               T2 = T2 + A(J,I+1)*X(J)
               T3 = T3 + A(J,I+2)*X(J)
               T4 = T4 + A(J,I+3)*X(J)
               T5 = T5 + A(J,I+4)*X(J)
               T6 = T6 + A(J,I+5)*X(J)
               T7 = T7 + A(J,I+6)*X(J)
               T8 = T8 + A(J,I+7)*X(J)
 1120       CONTINUE
            Y(I) = T1
            Y(I+1) = T2
            Y(I+2) = T3
            Y(I+3) = T4
            Y(I+4) = T5
            Y(I+5) = T6
            Y(I+6) = T7
            Y(I+7) = T8
 1140    CONTINUE
 1160 CONTINUE
      S2 = SECOND()
      TIME = (S2-S1) - (S1-S0)
      WRITE (6,FMT=99999) OPS*IREP/TIME/1.0D6,
     +  'DGEMV: In cache, A**T, DOT unrolled to depth 8'
C
C     DGEMV: In cache, A**T, AXPY
C
      OPS = 2*DBLE(NACTUAL)**2
      IREP = MAX(1,NINT(AVTIME/(OPS/AVMFLP/1.0D6)))
      S0 = SECOND()
      S1 = SECOND()
      DO 1220 KOUNT = 1, IREP
         DO 1200 J = 1, NACTUAL
            DO 1180 I = 1, NACTUAL
               Y(I) = Y(I) + A(J,I)*X(J)
 1180       CONTINUE
 1200    CONTINUE
 1220 CONTINUE
      S2 = SECOND()
      TIME = (S2-S1) - (S1-S0)
      WRITE (6,FMT=99999) OPS*IREP/TIME/1.0D6,
     +  'DGEMV: In cache, A**T, AXPY'
C
C     DGEMV: In cache, A**T, AXPY unrolled to depth 2
C
      OPS = 2*DBLE(NACTUAL)**2
      IREP = MAX(1,NINT(AVTIME/(OPS/AVMFLP/1.0D6)))
      S0 = SECOND()
      S1 = SECOND()
      DO 1280 KOUNT = 1, IREP
         DO 1260 J = 1, NACTUAL, 2
            DO 1240 I = 1, NACTUAL
               Y(I) = Y(I) + A(J,I)*X(J) + A(J+1,I)*X(J+1)
 1240       CONTINUE
 1260    CONTINUE
 1280 CONTINUE
      S2 = SECOND()
      TIME = (S2-S1) - (S1-S0)
      WRITE (6,FMT=99999) OPS*IREP/TIME/1.0D6,
     +  'DGEMV: In cache, A**T, AXPY unrolled to depth 2'
C
C     DGEMV: In cache, A**T, AXPY unrolled to depth 3
C
      OPS = 2*DBLE(NACTUAL)**2
      IREP = MAX(1,NINT(AVTIME/(OPS/AVMFLP/1.0D6)))
      S0 = SECOND()
      S1 = SECOND()
      DO 1340 KOUNT = 1, IREP
         DO 1320 J = 1, NACTUAL, 3
            DO 1300 I = 1, NACTUAL
               Y(I) = Y(I) + A(J,I)*X(J) + A(J+1,I)*X(J+1) + A(J+2,I)
     +                *X(J+2)
 1300       CONTINUE
 1320    CONTINUE
 1340 CONTINUE
      S2 = SECOND()
      TIME = (S2-S1) - (S1-S0)
      WRITE (6,FMT=99999) OPS*IREP/TIME/1.0D6,
     +  'DGEMV: In cache, A**T, AXPY unrolled to depth 3'
C
C
C     DGEMV: In cache, A**T, AXPY unrolled to depth 4
C
      OPS = 2*DBLE(NACTUAL)**2
      IREP = MAX(1,NINT(AVTIME/(OPS/AVMFLP/1.0D6)))
      S0 = SECOND()
      S1 = SECOND()
      DO 1400 KOUNT = 1, IREP
         DO 1380 J = 1, NACTUAL, 4
            DO 1360 I = 1, NACTUAL
               Y(I) = Y(I) + A(J,I)*X(J) + A(J+1,I)*X(J+1) + A(J+2,I)
     +                *X(J+2) + A(J+3,I)*X(J+3)
 1360       CONTINUE
 1380    CONTINUE
 1400 CONTINUE
      S2 = SECOND()
      TIME = (S2-S1) - (S1-S0)
      WRITE (6,FMT=99999) OPS*IREP/TIME/1.0D6,
     +  'DGEMV: In cache, A**T, AXPY unrolled to depth 4'
C
C     DGEMV: From memory, A not transposed, DOT
C
      OPS = 2*DBLE(NACTUAL)**2
      IREP = MAX(1,NINT(AVTIME/(OPS/AVMFLP/1.0D6+TFLUSH)))
      S0 = SECOND()
      S1 = SECOND()
      DO 1460 KOUNT = 1, IREP
         CALL FLUSH(IWORDS)
         DO 1440 I = 1, NACTUAL
            TEMP = Y(I)
            DO 1420 J = 1, NACTUAL
               TEMP = TEMP + A(I,J)*X(J)
 1420       CONTINUE
            Y(I) = TEMP
 1440    CONTINUE
 1460 CONTINUE
      S2 = SECOND()
      TIME = (S2-S1) - (S1-S0) - TFLUSH*IREP
      WRITE (6,FMT=99999) OPS*IREP/TIME/1.0D6,
     +  'DGEMV: From memory, A not transposed, DOT'
C
C     DGEMV: From memory, A not transposed, DOT unrolled to depth 2
C
      OPS = 2*DBLE(NACTUAL)**2
      IREP = MAX(1,NINT(AVTIME/(OPS/AVMFLP/1.0D6+TFLUSH)))
      S0 = SECOND()
      S1 = SECOND()
      DO 1520 KOUNT = 1, IREP
         CALL FLUSH(IWORDS)
         DO 1500 I = 1, NACTUAL, 2
            T1 = Y(I)
            T2 = Y(I+1)
            DO 1480 J = 1, NACTUAL
               T1 = T1 + A(I,J)*X(J)
               T2 = T2 + A(I+1,J)*X(J)
 1480       CONTINUE
            Y(I) = T1
            Y(I+1) = T2
 1500    CONTINUE
 1520 CONTINUE
      S2 = SECOND()
      TIME = (S2-S1) - (S1-S0) - TFLUSH*IREP
      WRITE (6,FMT=99999) OPS*IREP/TIME/1.0D6,
     +  'DGEMV: From memory, A not transposed, DOT unrolled to depth 2'
C
C     DGEMV: From memory, A not transposed, DOT unrolled to depth 3
C
      OPS = 2*DBLE(NACTUAL)**2
      IREP = MAX(1,NINT(AVTIME/(OPS/AVMFLP/1.0D6+TFLUSH)))
      S0 = SECOND()
      S1 = SECOND()
      DO 1580 KOUNT = 1, IREP
         CALL FLUSH(IWORDS)
         DO 1560 I = 1, NACTUAL, 3
            T1 = Y(I)
            T2 = Y(I+1)
            T3 = Y(I+2)
            DO 1540 J = 1, NACTUAL
               T1 = T1 + A(I,J)*X(J)
               T2 = T2 + A(I+1,J)*X(J)
               T3 = T3 + A(I+2,J)*X(J)
 1540       CONTINUE
            Y(I) = T1
            Y(I+1) = T2
            Y(I+2) = T3
 1560    CONTINUE
 1580 CONTINUE
      S2 = SECOND()
      TIME = (S2-S1) - (S1-S0) - TFLUSH*IREP
      WRITE (6,FMT=99999) OPS*IREP/TIME/1.0D6,
     +  'DGEMV: From memory, A not transposed, DOT unrolled to depth 3'
C
C     DGEMV: From memory, A not transposed, DOT unrolled to depth 4
C
      OPS = 2*DBLE(NACTUAL)**2
      IREP = MAX(1,NINT(AVTIME/(OPS/AVMFLP/1.0D6+TFLUSH)))
      S0 = SECOND()
      S1 = SECOND()
      DO 1640 KOUNT = 1, IREP
         CALL FLUSH(IWORDS)
         DO 1620 I = 1, NACTUAL, 4
            T1 = Y(I)
            T2 = Y(I+1)
            T3 = Y(I+2)
            T4 = Y(I+3)
            DO 1600 J = 1, NACTUAL
               T1 = T1 + A(I,J)*X(J)
               T2 = T2 + A(I+1,J)*X(J)
               T3 = T3 + A(I+2,J)*X(J)
               T4 = T4 + A(I+3,J)*X(J)
 1600       CONTINUE
            Y(I) = T1
            Y(I+1) = T2
            Y(I+2) = T3
            Y(I+3) = T4
 1620    CONTINUE
 1640 CONTINUE
      S2 = SECOND()
      TIME = (S2-S1) - (S1-S0) - TFLUSH*IREP
      WRITE (6,FMT=99999) OPS*IREP/TIME/1.0D6,
     +  'DGEMV: From memory, A not transposed, DOT unrolled to depth 4'
C
C     DGEMV: From memory, A not transposed, AXPY
C
      OPS = 2*DBLE(NACTUAL)**2
      IREP = MAX(1,NINT(AVTIME/(OPS/AVMFLP/1.0D6+TFLUSH)))
      S0 = SECOND()
      S1 = SECOND()
      DO 1700 KOUNT = 1, IREP
         CALL FLUSH(IWORDS)
         DO 1680 J = 1, NACTUAL
            DO 1660 I = 1, NACTUAL
               Y(I) = Y(I) + A(I,J)*X(J)
 1660       CONTINUE
 1680    CONTINUE
 1700 CONTINUE
      S2 = SECOND()
      TIME = (S2-S1) - (S1-S0) - TFLUSH*IREP
      WRITE (6,FMT=99999) OPS*IREP/TIME/1.0D6,
     +  'DGEMV: From memory, A not transposed, AXPY'
C
C     DGEMV: From memory, A not transposed, AXPY unrolled to depth 2
C
      OPS = 2*DBLE(NACTUAL)**2
      IREP = MAX(1,NINT(AVTIME/(OPS/AVMFLP/1.0D6+TFLUSH)))
      S0 = SECOND()
      S1 = SECOND()
      DO 1760 KOUNT = 1, IREP
         CALL FLUSH(IWORDS)
         DO 1740 J = 1, NACTUAL, 2
            DO 1720 I = 1, NACTUAL
               Y(I) = Y(I) + A(I,J)*X(J) + A(I,J+1)*X(J+1)
 1720       CONTINUE
 1740    CONTINUE
 1760 CONTINUE
      S2 = SECOND()
      TIME = (S2-S1) - (S1-S0) - TFLUSH*IREP
      WRITE (6,FMT=99999) OPS*IREP/TIME/1.0D6,
     +  'DGEMV: From memory, A not transposed, AXPY unrolled to depth 2'
C
C     DGEMV: From memory, A not transposed, AXPY unrolled to depth 3
C
      OPS = 2*DBLE(NACTUAL)**2
      IREP = MAX(1,NINT(AVTIME/(OPS/AVMFLP/1.0D6+TFLUSH)))
      S0 = SECOND()
      S1 = SECOND()
      DO 1820 KOUNT = 1, IREP
         CALL FLUSH(IWORDS)
         DO 1800 J = 1, NACTUAL, 3
            DO 1780 I = 1, NACTUAL
               Y(I) = Y(I) + A(I,J)*X(J) + A(I,J+1)*X(J+1) + A(I,J+1)
     +                *X(J+2)
 1780       CONTINUE
 1800    CONTINUE
 1820 CONTINUE
      S2 = SECOND()
      TIME = (S2-S1) - (S1-S0) - TFLUSH*IREP
      WRITE (6,FMT=99999) OPS*IREP/TIME/1.0D6,
     +  'DGEMV: From memory, A not transposed, AXPY unrolled to depth 3'
C
C     DGEMV: From memory, A not transposed, AXPY unrolled to depth 4
C
      OPS = 2*DBLE(NACTUAL)**2
      IREP = MAX(1,NINT(AVTIME/(OPS/AVMFLP/1.0D6+TFLUSH)))
      S0 = SECOND()
      S1 = SECOND()
      DO 1880 KOUNT = 1, IREP
         CALL FLUSH(IWORDS)
         DO 1860 J = 1, NACTUAL, 4
            DO 1840 I = 1, NACTUAL
               Y(I) = Y(I) + A(I,J)*X(J) + A(I,J+1)*X(J+1) + A(I,J+2)
     +                *X(J+2) + A(I,J+3)*X(J+3)
 1840       CONTINUE
 1860    CONTINUE
 1880 CONTINUE
      S2 = SECOND()
      TIME = (S2-S1) - (S1-S0) - TFLUSH*IREP
      WRITE (6,FMT=99999) OPS*IREP/TIME/1.0D6,
     +  'DGEMV: From memory, A not transposed, AXPY unrolled to depth 4'
C
C    DGEMV: From memory, A**T, DOT
C
      OPS = 2*DBLE(NACTUAL)**2
      IREP = MAX(1,NINT(AVTIME/(OPS/AVMFLP/1.0D6+TFLUSH)))
      S0 = SECOND()
      S1 = SECOND()
      DO 1940 KOUNT = 1, IREP
         CALL FLUSH(IWORDS)
         DO 1920 I = 1, NACTUAL
            TEMP = Y(I)
            DO 1900 J = 1, NACTUAL
               TEMP = TEMP + A(J,I)*X(J)
 1900       CONTINUE
            Y(I) = TEMP
 1920    CONTINUE
 1940 CONTINUE
      S2 = SECOND()
      TIME = (S2-S1) - (S1-S0) - TFLUSH*IREP
      WRITE (6,FMT=99999) OPS*IREP/TIME/1.0D6,
     +  'DGEMV: From memory, A**T, DOT'
C
C     DGEMV: From memory, A**T, DOT unrolled to depth 2
C
      OPS = 2*DBLE(NACTUAL)**2
      IREP = MAX(1,NINT(AVTIME/(OPS/AVMFLP/1.0D6+TFLUSH)))
      S0 = SECOND()
      S1 = SECOND()
      DO 2000 KOUNT = 1, IREP
         CALL FLUSH(IWORDS)
         DO 1980 I = 1, NACTUAL, 2
            T1 = Y(I)
            T2 = Y(I+1)
            DO 1960 J = 1, NACTUAL
               T1 = T1 + A(J,I)*X(J)
               T2 = T2 + A(J,I+1)*X(J)
 1960       CONTINUE
            Y(I) = T1
            Y(I+1) = T2
 1980    CONTINUE
 2000 CONTINUE
      S2 = SECOND()
      TIME = (S2-S1) - (S1-S0) - TFLUSH*IREP
      WRITE (6,FMT=99999) OPS*IREP/TIME/1.0D6,
     +  'DGEMV: From memory, A**T, DOT unrolled to depth 2'
C
C     DGEMV: From memory, A**T, DOT unrolled to depth 3
C
      OPS = 2*DBLE(NACTUAL)**2
      IREP = MAX(1,NINT(AVTIME/(OPS/AVMFLP/1.0D6+TFLUSH)))
      S0 = SECOND()
      S1 = SECOND()
      DO 2060 KOUNT = 1, IREP
         CALL FLUSH(IWORDS)
         DO 2040 I = 1, NACTUAL, 3
            T1 = Y(I)
            T2 = Y(I+1)
            T3 = Y(I+2)
            DO 2020 J = 1, NACTUAL
               T1 = T1 + A(J,I)*X(J)
               T2 = T2 + A(J,I+1)*X(J)
               T3 = T3 + A(J,I+2)*X(J)
 2020       CONTINUE
            Y(I) = T1
            Y(I+1) = T2
            Y(I+2) = T3
 2040    CONTINUE
 2060 CONTINUE
      S2 = SECOND()
      TIME = (S2-S1) - (S1-S0) - TFLUSH*IREP
      WRITE (6,FMT=99999) OPS*IREP/TIME/1.0D6,
     +  'DGEMV: From memory, A**T, DOT unrolled to depth 3'
C
C     DGEMV: From memory, A**T, DOT unrolled to depth 4
C
      OPS = 2*DBLE(NACTUAL)**2
      IREP = MAX(1,NINT(AVTIME/(OPS/AVMFLP/1.0D6+TFLUSH)))
      S0 = SECOND()
      S1 = SECOND()
      DO 2120 KOUNT = 1, IREP
         CALL FLUSH(IWORDS)
         DO 2100 I = 1, NACTUAL, 4
            T1 = Y(I)
            T2 = Y(I+1)
            T3 = Y(I+2)
            T4 = Y(I+3)
            DO 2080 J = 1, NACTUAL
               T1 = T1 + A(J,I)*X(J)
               T2 = T2 + A(J,I+1)*X(J)
               T3 = T3 + A(J,I+2)*X(J)
               T4 = T4 + A(J,I+3)*X(J)
 2080       CONTINUE
            Y(I) = T1
            Y(I+1) = T2
            Y(I+2) = T3
            Y(I+3) = T4
 2100    CONTINUE
 2120 CONTINUE
      S2 = SECOND()
      TIME = (S2-S1) - (S1-S0) - TFLUSH*IREP
      WRITE (6,FMT=99999) OPS*IREP/TIME/1.0D6,
     +  'DGEMV: From memory, A**T, DOT unrolled to depth 4'
C
C     DGEMV: From memory, A**T, AXPY
C
      OPS = 2*DBLE(NACTUAL)**2
      IREP = MAX(1,NINT(AVTIME/(OPS/AVMFLP/1.0D6+TFLUSH)))
      S0 = SECOND()
      S1 = SECOND()
      DO 2180 KOUNT = 1, IREP
         CALL FLUSH(IWORDS)
         DO 2160 J = 1, NACTUAL
            DO 2140 I = 1, NACTUAL
               Y(I) = Y(I) + A(J,I)*X(J)
 2140       CONTINUE
 2160    CONTINUE
 2180 CONTINUE
      S2 = SECOND()
      TIME = (S2-S1) - (S1-S0) - TFLUSH*IREP
      WRITE (6,FMT=99999) OPS*IREP/TIME/1.0D6,
     +  'DGEMV: From memory, A**T, AXPY'
C
C     DGEMV: From memory, A**T, AXPY unrolled to depth 2
C
      OPS = 2*DBLE(NACTUAL)**2
      IREP = MAX(1,NINT(AVTIME/(OPS/AVMFLP/1.0D6+TFLUSH)))
      S0 = SECOND()
      S1 = SECOND()
      DO 2240 KOUNT = 1, IREP
         CALL FLUSH(IWORDS)
         DO 2220 J = 1, NACTUAL, 2
            DO 2200 I = 1, NACTUAL
               Y(I) = Y(I) + A(J,I)*X(J) + A(J+1,I)*X(J+1)
 2200       CONTINUE
 2220    CONTINUE
 2240 CONTINUE
      S2 = SECOND()
      TIME = (S2-S1) - (S1-S0) - TFLUSH*IREP
      WRITE (6,FMT=99999) OPS*IREP/TIME/1.0D6,
     +  'DGEMV: From memory, A**T, AXPY unrolled to depth 2'
C
C     DGEMV: From memory, A**T, AXPY unrolled to depth 3
C
      OPS = 2*DBLE(NACTUAL)**2
      IREP = MAX(1,NINT(AVTIME/(OPS/AVMFLP/1.0D6+TFLUSH)))
      S0 = SECOND()
      S1 = SECOND()
      DO 2300 KOUNT = 1, IREP
         CALL FLUSH(IWORDS)
         DO 2280 J = 1, NACTUAL, 3
            DO 2260 I = 1, NACTUAL
               Y(I) = Y(I) + A(J,I)*X(J) + A(J+1,I)*X(J+1) + A(J+2,I)
     +                *X(J+2)
 2260       CONTINUE
 2280    CONTINUE
 2300 CONTINUE
      S2 = SECOND()
      TIME = (S2-S1) - (S1-S0) - TFLUSH*IREP
      WRITE (6,FMT=99999) OPS*IREP/TIME/1.0D6,
     +  'DGEMV: From memory, A**T, AXPY unrolled to depth 3'
C
C     DGEMV: From memory, A**T, AXPY unrolled to depth 4
C
      OPS = 2*DBLE(NACTUAL)**2
      IREP = MAX(1,NINT(AVTIME/(OPS/AVMFLP/1.0D6+TFLUSH)))
      S0 = SECOND()
      S1 = SECOND()
      DO 2360 KOUNT = 1, IREP
         CALL FLUSH(IWORDS)
         DO 2340 J = 1, NACTUAL, 4
            DO 2320 I = 1, NACTUAL
               Y(I) = Y(I) + A(J,I)*X(J) + A(J+1,I)*X(J+1) + A(J+2,I)
     +                *X(J+2) + A(J+3,I)*X(J+3)
 2320       CONTINUE
 2340    CONTINUE
 2360 CONTINUE
      S2 = SECOND()
      TIME = (S2-S1) - (S1-S0) - TFLUSH*IREP
      WRITE (6,FMT=99999) OPS*IREP/TIME/1.0D6,
     +  'DGEMV: From memory, A**T, AXPY unrolled to depth 4'
C
C    DGEMM: From cache, C <- C + AB, DOT unrolled twice
C
      AVMFLP = 40.0D0
      NACTUAL = 16
      OPS = 2*DBLE(NACTUAL)**2*MAXN
      IREP = MAX(1,NINT(AVTIME/(OPS/AVMFLP/1.0D6)))
      S0 = SECOND()
      S1 = SECOND()
      DO 2440 KOUNT = 1, IREP
         DO 2420 J = 1, NACTUAL, 2
            DO 2400 I = 1, NACTUAL, 2
               T11 = C(I,J)
               T12 = C(I,J+1)
               T21 = C(I+1,J)
               T22 = C(I+1,J+1)
               DO 2380 K = 1, MAXN
                  T11 = T11 + A(I,K)*B(K,J)
                  T12 = T12 + A(I,K)*B(K,J+1)
                  T21 = T21 + A(I+1,K)*B(K,J)
                  T22 = T22 + A(I+1,K)*B(K,J+1)
 2380          CONTINUE
               C(I,J) = T11
               C(I,J+1) = T12
               C(I+1,J) = T21
               C(I+1,J+1) = T22
 2400       CONTINUE
 2420    CONTINUE
 2440 CONTINUE
      S2 = SECOND()
      TIME = (S2-S1) - (S1-S0)
      WRITE (6,FMT=99999) OPS*IREP/TIME/1.0D6,
     +  'DGEMM: From cache, C <- C + AB, DOT unrolled twice'
C
C     DGEMM: From cache, C <- C + A**T*B, DOT unrolled twice
C
      OPS = 2*DBLE(NACTUAL)**2*MAXN
      IREP = MAX(1,NINT(AVTIME/(OPS/AVMFLP/1.0D6)))
      S0 = SECOND()
      S1 = SECOND()
      DO 2520 KOUNT = 1, IREP
         DO 2500 J = 1, NACTUAL, 2
            DO 2480 I = 1, NACTUAL, 2
               T11 = C(I,J)
               T12 = C(I,J+1)
               T21 = C(I+1,J)
               T22 = C(I+1,J+1)
               DO 2460 K = 1, MAXN
                  T11 = T11 + A(K,I)*B(K,J)
                  T12 = T12 + A(K,I)*B(K,J+1)
                  T21 = T21 + A(K,I+1)*B(K,J)
                  T22 = T22 + A(K,I+1)*B(K,J+1)
 2460          CONTINUE
               C(I,J) = T11
               C(I,J+1) = T12
               C(I+1,J) = T21
               C(I+1,J+1) = T22
 2480       CONTINUE
 2500    CONTINUE
 2520 CONTINUE
      S2 = SECOND()
      TIME = (S2-S1) - (S1-S0)
      WRITE (6,FMT=99999) OPS*IREP/TIME/1.0D6,
     +  'DGEMM: From cache, C <- C + A**T*B, DOT unrolled twice'
      STOP
C
99999 FORMAT (1X,F5.2,' mflops -- ',A)
      END
      SUBROUTINE INIT(A,M,N)
C
C     Initializes an array
C
C     .. Scalar Arguments ..
      INTEGER         M, N
C     .. Array Arguments ..
      DOUBLE PRECISION A(M,N)
C     .. Local Scalars ..
      INTEGER         I, J
C     .. Intrinsic Functions ..
      INTRINSIC       DBLE
C     .. Executable Statements ..
      DO 40 J = 1, N
         DO 20 I = 1, M
            A(I,J) = 0.01D0*DBLE(J+I)
   20    CONTINUE
   40 CONTINUE
      RETURN
      END
      SUBROUTINE FLUSH(IWORDS)
C
C     This subroutine causes IWORDS words to be initialized. If it
C     is called before a code fragment, it will make sure that
C     the cache is flushed completely.
C
C     .. Local Scalars ..
      INTEGER          I
C     .. Local Arrays ..
      DOUBLE PRECISION FOO(8192)
C     .. Scalar Arguments ..
      INTEGER          IWORDS
C     .. Executable Statements ..
      IF (IWORDS.GT.8192) THEN
         WRITE (6,FMT=*)
     +     '*** Increase size of array in subroutine FLUSH'
         STOP
      END IF
      DO 20 I = 1, IWORDS
         FOO(I) = 1.1D0
   20 CONTINUE
      RETURN
      END

.