Mudanças entre as edições de "Timing"

Edição das 02h10min de 20 de junho de 2012

Testando vetorização:

! http://goparallel.sourceforge.net/optimizing-loops-vectorization/
program Vectorization
use portlib
real(4),dimension(:),allocatable  ::  x,y,z
integer                           ::  len=150*1024*1024 ! 154 MiB=150MB
real(4)                           ::  timing

allocate( x(len) ,stat=ierr)
allocate( y(len) ,stat=ierr)
allocate( z(len) ,stat=ierr)

do j=1,10
timing = secnds(0.0)
do i=1,len
  z(i)=sqrt(x(i))+sqrt(y(i))
end do
timing = secnds(timing)*1000
print *,' Timing =',timing,'/1000 s'
end do
end program

• Memory: using performance monitor from windows

 Maior problema alocável: 150MiB * 3*4 = 1.75GiB = 1.88GB

 real(4)                               = 1.85 GB 
 real(8)                               = 3.69 GB
 real(16)                              = 7.39 GB

• Results real(4):

Debug   (x32)        2.13  s
Debug   (x64)        2.00  s
Release (x32) /O2    0.143 s
Release (x64) /O2    0.140 s   <=========

Release(x64)
  Threshold for vectorization   0  0.140 s
  Threshold for parallelization 0  0.140 s
  /Qvec-                           0.909 s
  /Qvec- /Qparallelization         0.171 s   usa 8 processors
  Inline directive                 0.145 s
  /Ob1                                  use 4 processors
  /Qvec- /Qparallelization         0.232 s

• Resultados:

Release (x64)                      0.461 s  (default optimization)
  /Qvec-                           0.909 s
  /Qvec- /Qparallelization         0.356 s  (auto-paralelizacao) <==========

• real(16): As coisas ficam lll...e...n.t...as.. . .. . . (vectorization doesn't help)

Release (x64)                      7.00  s 
  /Qvec-   (sem vetorizacao)       7.02  s
  /Qvec- /Qparallelization         1.75  s
  /Qparallelization                1.75  s