SBR600: To Thumb? or Not to Thumb? (v0.1)

** CONTENT OF THIS ENTRY IS INCORRECT **

I have got the root access to a brand new beagleXM machine. This machine is named cdot-guru-4-1 and it's running on fedora-arm armv5tel architecture.

I was benchmarking this machine with and without Thumb instruction set. All the compiling was done with rpmbuild and source rpm files.

For compiling options for rpm, the macros file is located at /usr/lib/rpm/platform/armv5tel-linux/macros. I enabled Thumb by giving -mthumb and -mthumb-interwork switchs to %optflags inside macro file.

NLED

	-mthumb	-mthumb-interwork	none
Compiled rpm size	30788 (31K)	30788 (31K)	30792 (31K)
Compile time	real 0m15.633s user 0m13.920s sys 0m1.250s	real 0m15.671s user 0m13.750s sys 0m1.410s	real 0m15.739s user 0m13.740s sys 0m1.440s

Surprisingly, I didn't see enough difference. I had a bunch of thoughts in my head after testing with NLED. I started believing that I might be doing something incorrectly. I've also thought NLED being tiny in size could be the reason.

gzip

	-mthumb	-mthumb-interwork	none
Compiled rpm size	116548 (114K)	116544 (114K)	116556 (114K)
Compile time	real 1m16.543s user 0m49.060s sys 0m16.190s	real 1m22.706s user 0m49.260s sys 0m15.810s	real 1m21.846s user 0m48.880s sys 0m16.130s

calc

	-mthumb	-mthumb-interwork	none
Compiled rpm size	587920 (575K)	587904 (575K)	587912 (575K)
Compile time	real 5m39.970s user 4m33.500s sys 0m30.300s	real 5m32.401s user 4m33.790s sys 0m30.670s	real 5m33.676s user 4m34.350s sys 0m31.080s

No real difference. I began to notice the performance gap between regular size PC and the beagleXM.

gedit

	-mthumb	-mthumb-interwork	none
Compiled rpm size	3374912 (3.3M)	3374868 (3.3M)	3374880 (3.3M)
Compile time	real 31m25.956s user 28m21.630s sys 1m39.280s	real    31m32.184s user    28m25.480s sys    1m38.960s	real 31m37.174s user 28m11.660s sys 1m38.880s

I tried testing with a bigger source file as well. When I tried compiling GIMP, which was a 16MB file, it took way too long and I could not wait. So I ended up stopping at the 92 minutes mark. Note that in my previous blog post, the same file got compiled in less than 6 minutes on my desktop PC.