[Tremor] LOW_ACCURACY mode and MDCT improvements

[Daniel Kahn Gillmor]

On 03/23/2010 11:53 PM, Mike Giacomelli wrote:
>> Configure is dumb, second guess it :)

i "second-guessed" it by pre-defining CFLAGS="-D_ARM_ASSEM_" for the moment.

here's with _ARM_ASSEM_ :

65690+1 records in
65690+1 records out
33633600 bytes (34 MB) copied, 27.6992 s, 1.2 MB/s

real	0m27.768s
user	0m23.420s
sys	0m4.320s


and here's without:

65690+1 records in
65690+1 records out
33633600 bytes (34 MB) copied, 39.0579 s, 861 kB/s

real	0m39.157s
user	0m35.150s
sys	0m3.570s


So that's certainly a speedup, about 25%.  not bad.

But!  i then tried to install the ARM_ASSEM-built .so at /usr/lib
(rather than invoking it with LD_PRELOAD or LD_LIBRARY_PATH) and the
speed of the result ends up back where we started (~39s), even with a
binary-identical library in either place.

Any thoughts about why that might be happening?  It leaves me
unconvinced that the speedup i'm seeing has anything to do with the
assembly (maybe it has to do with the linker?)

Sorry this profiling is taking so many twists and turns.

Is there a way i could inspect the built library to tell whether the
assembler routines were used?

> Actually if removing the dash so its "arm*" instead of "arm-*" runs for you then we should probably just submit a patch to fix configure.
> 
> arm-elf-gcc and such are common names for gcc/arm compilers, but not the only ones and I doubt there are non-arm gcc versions out there that begin with "arm" . . .  		 	   		  

Hrm, i don't think i know enough to propose such a patch, or even to
whom i would propose it.

Suggestions welcome,

	--dkg

-------------- next part --------------
A non-text attachment was scrubbed...
Name: signature.asc
Type: application/pgp-signature
Size: 891 bytes
Desc: OpenPGP digital signature
Url : http://lists.xiph.org/pipermail/tremor/attachments/20100325/e97e8ffa/attachment.pgp