------- Comment #1 from jv244 at cam dot ac dot uk 2007-03-08 11:11 -------
The following is (for me) an even more interesting example, as it times only
the loop that thus the actual multiply / add but also tricks my version of
ifort into generating the expected asm. Ifort is about twice as fast as
gfortran on it.
SUBROUTINE collocate_core_2_2_0_0(jg,cmax)
IMPLICIT NONE
integer, INTENT(IN) :: jg,cmax
INTEGER, PARAMETER :: wp = SELECTED_REAL_KIND ( 14, 200 )
INTEGER, PARAMETER :: N=10,Nit=100000000
TYPE vec
real(wp) :: a(2)
END TYPE vec
TYPE(vec) :: dpy(1000)
TYPE(vec) :: pxy(1000)
TYPE(vec) :: s(02)
integer :: i,j
DO i=1,N
pxy(i)%a=0.0_wp
ENDDO
DO i=1,N
dpy(i)%a=0.0_wp
ENDDO
s(01)%a(1)=0.0_wp
s(01)%a(2)=0.0_wp
s(02)%a(1)=0.0_wp
s(02)%a(2)=0.0_wp
CALL USE(dpy,pxy,s)
DO j=1,Nit
DO i=1,N
s(01)%a(:)=s(01)%a(:)+pxy(i)%a(:)*dpy(i)%a(1)
s(02)%a(:)=s(02)%a(:)+pxy(i)%a(:)*dpy(i)%a(2)
ENDDO
ENDDO
CALL USE(dpy,pxy,s)
END SUBROUTINE
[EMAIL PROTECTED]:/data/vondele/extracted_collocate/test> gfortran -O2
-march=native -ftree-vectorize -ffast-math test.f90
[EMAIL PROTECTED]:/data/vondele/extracted_collocate/test> ./a.out
4.288268
[EMAIL PROTECTED]:/data/vondele/extracted_collocate/test> ifort -O2 -xT test.f90
test.f90(16) : (col. 8) remark: LOOP WAS VECTORIZED.
test.f90(19) : (col. 8) remark: LOOP WAS VECTORIZED.
test.f90(31) : (col. 6) remark: LOOP WAS VECTORIZED.
test.f90(31) : (col. 6) remark: LOOP WAS VECTORIZED.
test.f90(32) : (col. 6) remark: LOOP WAS VECTORIZED.
test.f90(32) : (col. 6) remark: LOOP WAS VECTORIZED.
[EMAIL PROTECTED]:/data/vondele/extracted_collocate/test> ./a.out
1.944121
The inner loop asm looks, with ifort, also the way I was hoping it to look
like:
.B2.7: # Preds ..B2.7 ..B2.6
movddup -16+collocate_core_2_2_0_0_$DPY.0.0(%rcx), %xmm2 #31.41
movddup -8+collocate_core_2_2_0_0_$DPY.0.0(%rcx), %xmm3 #32.41
addq $16, %rdx #33.4
movapd collocate_core_2_2_0_0_$PXY.0.0(%rdx), %xmm4 #31.6
mulpd %xmm4, %xmm2 #31.39
mulpd %xmm3, %xmm4 #32.39
addpd %xmm2, %xmm1 #31.7
addpd %xmm4, %xmm0 #32.7
addq $16, %rcx #33.5
cmpq $160, %rcx #33.4
jle ..B2.7 # Prob 90% #33.4
# LOE rdx rcx rbx rbp r12 r13 r14 r15 eax xmm0
xmm1
--
http://gcc.gnu.org/bugzilla/show_bug.cgi?id=31079
