ctime-results

来自「A C++ class library for scientific compu」· 代码 · 共 95 行

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.