Slow MATMUL

[Tobias Burnus]

Hello,

I am not sure what to do but using the NAG or Intel compiler, a MATMUL 
can be 4 times faster than the gfortran version. Using a the MATMUL from 
Intel's MKL, the performance is regained.

Thus, a possible solution would be to see whether one can (by 
algorithmic changes in libgfortran or by middle end changes?) make the 
built-in mathmul faster.

If one tries the NETLIB version directly, the performance is improved:

gfortran -O3 -march=native -fexternal-blas test.f90 dgemm.f lsame.f xerbla.f
  Time, MATMUL:    1.0120630
  dgemm:    1.0160630

However, as soon as I enable LTO (-flto -fwhole-program), even with 
--param hot-bb-frequency-fraction=2000, I am back to the slow 
performance of:
  Time, MATMUL:    1.4760920
  dgemm:    1.4760920


Based on 
http://groups.google.com/group/comp.lang.fortran/browse_thread/thread/1cba8e6ce5080197

The following program needs with GCC 2.5 (Athlon64) to 4.1 (Core2) times 
as long as with the Intel compiler (all: x86-64-linux).

On an Intel Core2 Duo E8400 (intel = ifort 11.1; gfortran = GCC 4.6.0):
(number = time in seconds)
   0.4579300   ifort -O2
   0.1929710   ifort -xHost -O3
  [0.1509770   when using dgemm (MKL) manually]

   0.76688403   gfortran -O2
   0.79987901   gfortran -march=native -Ofast -funroll-loops
   0.80287802   gfortran -funroll-loops -floop-strip-mine -floop-block \
                -floop-interchange -march=native -Ofast -funroll-loops
  [0.17097399   when using dgemm (MKL) via -fexternal-blas]
  [0.15197700   when using dgemm (MKL) manually]

On a Athlon 64 X2 Dual Core Processor 4800+ (2.4 GHz):
   0.9760600    ifort -O2
   0.6840420    ifort -O3 -xHost
  [0.4360271    using dgemm (MKL) manually]
  [1.584099     using dgemm (BLAS, gfortran -O2 compiled)]

   1.6761041    gfortran -O2
   1.6721040    gfortran -O3 -march=native -funroll-loops
   1.6801040    gfortran -flto -fwhole-program -floop-strip-mine \
                -floop-block -floop-interchange -ffast-math \
                -O3 -march=native -funroll-loops
   MKL BLAS (call via -O2
   1.5800979    gfortran -march=native -fexternal-blas -O3 -lblas

   gfortran -O2 compiled BLAS:
   0.43602699   gfortran -march=native -fexternal-blas -O3 -lblas
                using dgemm (BLAS, gfortran -O2 compiled)]

   0.4200000    nagf95 -O3 -kind=byte


Note: gfortran calls _gfortran_matmul_r8, NAG calls __NAGf90_dgemm. The 
version which manually calls dgemm (cf. c.l.f) is with a normal 
"gfortran -O2"-compiled LAPACK also not faster.


! "mult_test.f90"
program t2
implicit none

REAL time_begin, time_end
integer, parameter :: n=200;
REAL(8) :: a(n,n), b(n,n), c(n,n)
integer, parameter :: m = 100
integer :: i

call RANDOM_NUMBER(a)
call RANDOM_NUMBER(b)

call cpu_time(time_begin)

do i = 1,m
     a(1,1) = a(1,1) + 0.1
     c = MATMUL(a,b)
enddo

call cpu_time(time_end)

PRINT *, 'Time, MATMUL: ',time_end-time_begin

! call cpu_time(time_begin)
! do i = 1,m
!    a(1,1) = a(1,1) + 0.1
!    call dgemm('n','n',n, n, n, dble(1.0), a, n, b, n, dble(0.0), c, n)
! enddo
!
! call cpu_time(time_end)
! PRINT *, 'dgemm: ',time_end-time_begin

end program



Tobias