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
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
|
On olympus.extreme.indiana.edu (sparc-sun-solaris2.6):
egcs 1.1b
Initial version, with -O2 -ftemplate-depth-30 -O2 -funroll-loops
-fstrict-aliasing
ctime1 17.7 0.9
ctime2 25.7 1.2
ctime3 52.0 2.1
ctime4 sleep
With -fno-gcse:
ctime1 17.3 1.0
ctime2 26.3 1.3
ctime3 1:02.0 2.1
ctime4 sleep
With -O:
ctime1 17.3 0.8
ctime2 24.4 1.2
ctime3 51.5 2.1
ctime4 sleep
With -O -fno-inline:
ctime1 16.9 0.8
ctime2 20.0 1.0
ctime3 24.7 1.2
ctime4 31.2 1.6
Woohoo. Okay, obviously inlining is the key.
Now try new expression templates:
With -O -funroll-loops -DBZ_NEW_EXPRESSION_TEMPLATES
ctime1 14.1 0.9
ctime2 22.3 1.2
ctime3 58.8 2.2
With -O -funroll-loops -DBZ_NEW_EXPRESSION_TEMPLATES -DBZ_NO_INLINE_ET
ctime1 14.1 0.9
ctime2 21.1 1.0
ctime3 45.4 1.9
With -O -funroll-loops -DBZ_NEW_EXPRESSION_TEMPLATES -DBZ_NO_INLINE_ET -DBZ_ETPARMS_CONSTREF
ctime1 14.6 0.8
ctime2 20.7 1.1
ctime3 41.6 2.1
ctime4 1:27.7 3.0
Things to try:
-fno-inline
Just -O (this will turn off -funroll-all-loops)
-fno-expensive-optimizations
-fno-unroll-all-loops
-fno-strength-reduce
-fno-rerun-cse-after-loop
On hgar1.cwru.edu (alpha), with KCC:
With +K3 -O3 -DBZ_NEW_EXPRESSION_TEMPLATES -DBZ_NO_INLINE_ET -DBZ_ETPARMS_CONSTREF:
ctime1 13.1 0.8
ctime2 20.9 1.0
ctime3 27.3 1.0
ctime4 36.2 1.1
ctime5 48.7 1.2
With just +K3 -O3:
ctime1 15.8 0.9
ctime2 25.3 1.0
ctime3 46.2 1.2
ctime4 79.9 1.5
So a speed up of about X 2 with KCC, not counting the overhead.
Here are the results for <valarray>:
ctime1 0.9 0.2
ctime2 2.1 0.2
ctime3 9.4 0.3
ctime4 33.2 0.4
ctime5 1:13 0.6
For C code:
ctime5 0.35 0.08
Pretty terrible.
|
