Hi Alessandro and all,
Compiler optimizations can be helpful when tuning code for good
performance, but I am surprised that you see anything like a 2x
improvement in decoding speed.
The following table shows the results of a series of tests I made today
for the decoder (the executable program jt9) running on files like the
ones you created (01.wav, 02.wav, ... 10.wav, all copies of the example
file 130610_2343.wav). The first column lists the Fortran compiler
flags used; the numerical column gives the total execution time (wall
clock) for processing the ten files.
FFLAGS Time
------------------------------------------------
-O0 -fbounds-check 42.4
-O1 -fbounds-check 22.8
-Os -fbounds-check 22.8
-O2 -fbounds-check -funroll-all-loops 20.4 *
-O2 -fbounds-check 20.2
-O3 -fbounds-check 19.8
-Ofast -fbounds-check 18.9
-O2 18.4
-O2 -mtune=native 18.4
-O2 -funroll-all-loops 18.2
-O3 18.0
-Ofast 17.8
------------------------------------------------
* Used in the release builds of WSJT-X
As you can see, "-mtune=native" made essentially no difference. The
biggest improvement in execution performance (over the default Release
build) is gained by turning off bounds-checking. A slight additional
improvement is obtained by using -O3 or -Ofast rather than -O2.
However, the total available improvement is less than 15%.
Obviously, such tests will give different results on different machines.
Those described above were done on a machine with a Core2 Duo E6750
CPU, 2.66 Ghz. Here is a similar set of results for a Windows machine
(Core i5-2500, 3.3 GHz):
FFLAGS Time
------------------------------------------------
-O0 -fbounds-check 28.5
-O1 -fbounds-check 18.2
-O2 -fbounds-check -funroll-all-loops 16.6 *
-O2 -fbounds-check 16.2
-O3 -fbounds-check 16.2
-Ofast 15.7
-O3 -m32 -msse -funroll-all-loops 15.4
-O3 -mtune=core2 15.1
-O3 -m32 -msse 15.0
-O3 -mtune=native 15.0
------------------------------------------------
* Used in the release builds of WSJT-X for Windows
The flags we're currently using for Windows Release builds give results
within about 10% of the best one listed.
One way to look at all of this is that the most important optimizations
are those that have already been done, by the programmer. These include
making the best possible choices of data structures, algorithms, loop
ordering, etc., etc.
-- 73, Joe, K1JT
------------------------------------------------------------------------------
Download BIRT iHub F-Type - The Free Enterprise-Grade BIRT Server
from Actuate! Instantly Supercharge Your Business Reports and Dashboards
with Interactivity, Sharing, Native Excel Exports, App Integration & more
Get technology previously reserved for billion-dollar corporations, FREE
http://pubads.g.doubleclick.net/gampad/clk?id=164703151&iu=/4140/ostg.clktrk
_______________________________________________
wsjt-devel mailing list
wsjt-devel@lists.sourceforge.net
https://lists.sourceforge.net/lists/listinfo/wsjt-devel
