emqbit cfft performance test


I finally got my BB-XM working the dsplink.

I tried to run the emqbit cfft tests on both ARM and DSP, and the
performance was worst in the DSP than in the GPP. Does it make sense?

root@beagleboard:~/emqbit# ./cfft_dsp
N=16,nTimes=100: 0.004 s
N=32,nTimes=100: 0.01 s
N=64,nTimes=100: 0.026 s
N=128,nTimes=100: 0.062 s
N=256,nTimes=100: 0.145 s
N=512,nTimes=100: 0.329 s
N=1024,nTimes=100: 0.737 s
N=2048,nTimes=100: 1.633 s
N=4096,nTimes=100: 3.596 s
N=8192,nTimes=100: 7.893 s
N=16384,nTimes=100: 17.238 s

root@beagleboard:~/emqbit# ./cfft_arm
N=16,nTimes=100: 0.000885 s
N=32,nTimes=100: 0.001923 s
N=64,nTimes=100: 0.005127 s
N=128,nTimes=100: 0.011414 s
N=256,nTimes=100: 0.025605 s
N=512,nTimes=100: 0.05661 s
N=1024,nTimes=100: 0.124237 s
N=2048,nTimes=100: 0.278625 s
N=4096,nTimes=100: 0.601013 s
N=8192,nTimes=100: 1.30441 s
N=16384,nTimes=100: 2.8063 s

I guess the dsp test really used the DSP, because in the GPP one, the
CPU usage was about 99% all the time, while during the DSP test, the
CPU usage was of about 1%.

Anyway, after running the tests, i get a lot of messages like this in
the kernel messages:

[ 394.068176] BUG: Your driver calls ioremap() on system memory. This leads
[ 394.068206] to architecturally unpredictable behaviour on ARMv6+,
and ioremap()
[ 394.068206] will fail in the next kernel release. Please fix your driver.
[ 394.093597] ------------[ cut here ]------------
[ 394.100006] WARNING: at arch/arm/mm/ioremap.c:211

Could this error be the performance bottleneck?

And finally, wasn't the DSP test supposed to be faster?