Welcome to the MacNN Forums.

oeyvind · Jan 21, 2002, 09:44 AM

Anyone know if the default configure and compiling of LAME from source do make use of the G4's AltiVec unit? If not, how do I patch the source so that it's AltiVec aware?

Thanks in advance

NoAuthoritaw · Jan 21, 2002, 08:20 PM

Originally posted by oeyvind:
Anyone know if the default configure and compiling of LAME from source do make use of the G4's AltiVec unit? If not, how do I patch the source so that it's AltiVec aware?

Thanks in advance

I don't know how much benefit will be seen simply by enabling the compiler's default AltiVec optimization, but the compiler directive that turns it on is:

-faltivec

To see any serious benefit I'd imagine you'd have to go in and create your own process branch in LAME that has AltiVec specific function calls. But hey tell me how it goes for you just flipping the toggle switch.

[ 01-21-2002: Message edited by: NoAuthoritaw ]

oeyvind · Jan 21, 2002, 08:53 PM

% CFLAGS=-faltivec; export CFLAGS
% CPPFLAGS=-faltivec; export CPPFLAGS
% ./configure
% make

hmmm, seem a bit faster, I didn't do much benchmark, but seem to be a bit faster (or is it all in my imagination)?

LAME version 3.92 Alpha 1 (20020121)

[ 01-31-2002: Message edited by: oeyvind ]

NoAuthoritaw · Jan 21, 2002, 08:56 PM

Originally posted by oeyvind:
% CFLAGS=-altivec; export CFLAGS
% CPPFLAGS=-altivec; export CPPFLAGS
% ./configure
% make

hmmm, seem a bit faster, I didn't do much benchmark, but seem to be a bit faster (or is it all in my imagination)?

LAME version 3.92 Alpha 1 (20020121)

[ 01-21-2002: Message edited by: oeyvind ]

Only way to tell is to benchmark it

oeyvind · Jan 21, 2002, 10:02 PM

LAME version 3.92 alpha 1 (released 20020122)

Configure option used:
./configure --disable-nls --enable-mp3x --enable-brhist

Track used: Track 1 of Buena Vista Social Club: Chan Chan

A (without CFLAGS & CPPFLAG with the extra -faltivec)

lame option: --r3mix -b 112 (histogram removed)
<BLOCKQUOTE>code:<HR><pre>
Encoding as 44.1 kHz VBR(q=1) j-stereo MPEG-1 Layer III (ca. 6.5x) qval=2
Frame | CPU time/estim | REAL time/estim | play/CPU | ETA
9873/9876 (100%)| 2:18/ 2:18| 2:29/ 2:29| 1.8596x| 0:00
average: 173.3 kbps LR: 2160 (21.87%) MS: 7716 (78.13%)
[/code]

lame option: -h -b 160
<BLOCKQUOTE>code:<HR><pre>
Encoding as 44.1 kHz 160 kbps j-stereo MPEG-1 Layer III (8.8x) qval=2
Frame | CPU time/estim | REAL time/estim | play/CPU | ETA
9873/9876 (100%)| 1:27/ 1:27| 1:32/ 1:32| 2.9593x| 0:00
average: 160.0 kbps LR: 1195 (12.10%) MS: 8681 (87.90%)
[/code]

B (WITH CFLAGS & CPPFLAG with the extra -faltivec)

lame option: --r3mix -b 112 (histogram removed)
<BLOCKQUOTE>code:<HR><pre>
Encoding as 44.1 kHz VBR(q=1) j-stereo MPEG-1 Layer III (ca. 6.5x) qval=2
Frame | CPU time/estim | REAL time/estim | play/CPU | ETA
9873/9876 (100%)| 2:16/ 2:16| 2:28/ 2:28| 1.8847x| 0:00
average: 173.3 kbps LR: 2160 (21.87%) MS: 7716 (78.13%)
[/code]

lame option: -h -b 160
<BLOCKQUOTE>code:<HR><pre>
Encoding as 44.1 kHz 160 kbps j-stereo MPEG-1 Layer III (8.8x) qval=2
Frame | CPU time/estim | REAL time/estim | play/CPU | ETA
9873/9876 (100%)| 1:26/ 1:26| 1:30/ 1:30| 2.9795x| 0:00 average: 160.0 kbps LR: 1195 (12.10%) MS: 8681 (87.90%)
[/code]

Apocalypse · Jan 22, 2002, 10:28 PM

from what I can tell, it doesn't look much better. This most likely means that the compiler doesn't actually do anything on its own except enable the altivec code in your source to be compiled (basically by defining the compiler directive you should be checking for). I was pretty sure that you actually had to write the altivec VecLib calls yourself so you might want to try replacing a few of the computational lines with actual vector code and see what happens.

I don't actually know, though. Ask someone on Apple's dev list to find out what they actually have the compiler doing with that option.

Post back if you pursue this,
Jeff.

NoAuthoritaw · Jan 23, 2002, 01:46 AM

Originally posted by Apocalypse:
from what I can tell, it doesn't look much better. This most likely means that the compiler doesn't actually do anything on its own except enable the altivec code in your source to be compiled (basically by defining the compiler directive you should be checking for). I was pretty sure that you actually had to write the altivec VecLib calls yourself so you might want to try replacing a few of the computational lines with actual vector code and see what happens.

I don't actually know, though. Ask someone on Apple's dev list to find out what they actually have the compiler doing with that option.

Post back if you pursue this,
Jeff.

I've not looked at the LAME code so I can't say how difficult it would be to vectorize it with calls to the AltiVec libraries. However if someone does try to simply go in and substitute standard arithmetic operations with vector operations bear in mind that AltiVec is heavily dependent on proper data alignment. This might mean that some amount of work may need to be done to make sure that data gets loaded into the registers properly.

-Nathan

oeyvind · Jan 23, 2002, 04:44 AM

Originally posted by NoAuthoritaw:


I've not looked at the LAME code so I can't say how difficult it would be to vectorize it with calls to the AltiVec libraries. However if someone does try to simply go in and substitute standard arithmetic operations with vector operations bear in mind that AltiVec is heavily dependent on proper data alignment. This might mean that some amount of work may need to be done to make sure that data gets loaded into the registers properly.

-Nathan

Latest lame src is here, really hope someone is actually working to get LAME to use Altivec units in G4...^^

Apocalypse · Jan 23, 2002, 11:37 AM

We need to find someone with VastC/AltiVec. Apparently it will convert all c language arithmetic to AltiVec library calls as well as perform proper parallelization to ensure that none of the potential has been wasted.

If I had a G4 I would look into it. I am very interested in AltiVec and I think that, due to its advantages, Mac OS XI will probably be G4 or higher using all AltiVec enabled OS code. That would be damn sweet for development knowing that even the API was fully optimized.

Hopefully I will get some money and then pick up an AltiVec system soon. Of course, soon will be when we are referring to G5's or G6's since I am only a student.

Hope someone with the means is interested in our little experiment,
Jeff.

oeyvind · Jan 23, 2002, 08:18 PM

Good news, just heard back from Ivnn Cavero Belaunde, one of the LAME developer that there's "some bulk of the work (LAME shared library, QT export component) hasn't been checked in yet."... cool!

and "Someone in the Architecture group at Apple has volunteered to do some Altivec work on the vectorizable parts of LAME, so I expect we'll see some stuff soon, which will be very sweet. He has access to profiling tools us mere mortals do not."

BTW, if you need G4, I think the SourgeForge compile farm has them

[ 01-23-2002: Message edited by: oeyvind ]