1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
|
C Initial Fortran 77 version of the Lattice QCD benchmark.
subroutine qcdf(M, res, src, V, iters)
integer V, iters, i, site, spin, col
complex*16 M(3,3,V), res(3,2,V), src(3,2,V)
DO i=1,iters
DO site=1,V
DO spin=1,2
DO col=1,3
res(col,spin,site) = M(col,1,site) * src(1,spin,site)
. + M(col,2,site) * src(2,spin,site)
. + M(col,3,site) * src(3,spin,site)
ENDDO
ENDDO
ENDDO
ENDDO
return
end
C Hand-tuned version
C Changes:
C o Ordering of array altered to improve layout of data in memory
C o col and spin loops unwound; it was found that unwinding the
C col loop inside the spin loop was marginally faster (by 1.1%)
C o Unwinding both loops was faster than unwinding just one.
subroutine qcdf2(M, res, src, V, iters)
integer V, iters, i, site
complex*16 M(3,3,V), res(3,2,V), src(3,2,V)
DO i=1,iters
DO site=1,V
C col=1, spin=1
res(1,1,site) = M(1,1,site) * src(1,1,site)
. + M(1,2,site) * src(2,1,site)
. + M(1,3,site) * src(3,1,site)
C col=2, spin=1
res(2,1,site) = M(2,1,site) * src(1,1,site)
. + M(2,2,site) * src(2,1,site)
. + M(2,3,site) * src(3,1,site)
C col=3, spin=1
res(3,1,site) = M(3,1,site) * src(1,1,site)
. + M(3,2,site) * src(2,1,site)
. + M(3,3,site) * src(3,1,site)
C col=1, spin=2
res(1,2,site) = M(1,1,site) * src(1,2,site)
. + M(1,2,site) * src(2,2,site)
. + M(1,3,site) * src(3,2,site)
C col=2, spin=2
res(2,2,site) = M(2,1,site) * src(1,2,site)
. + M(2,2,site) * src(2,2,site)
. + M(2,3,site) * src(3,2,site)
C col=3, spin=2
res(3,2,site) = M(3,1,site) * src(1,2,site)
. + M(3,2,site) * src(2,2,site)
. + M(3,3,site) * src(3,2,site)
ENDDO
ENDDO
return
end
|
