Ogg Vorbis acceleration project, Is it dead? Options

[john33]

Here is an x64 compile of oggenc2.85 with FLAC support. Currently this is based on the standard 1.2.3 libs, but if some kind people would like to test this and confirm whether it's OK, or not, I'll go ahead and build an aoTuV version.

http://www.rarewares.org/files/ogg/oggenc2.85-1.2.3-x64.zip

Any chance to have an x64 compile of this Lancer mod too ?

I'm afraid not as it would require heavy modification of the Lancer code which is something I have neither the time, nor the skills to do!

[lvqcl]

I don't have x64 OS yet... so I cannot even test my x64 compile, but here it is (MSVS9 compile, SSE3): [obsolete; removed]

Maybe it even works. It lacks FLAC and SRC support, though.


This post has been edited by lvqcl: Feb 20 2012, 17:20

[forart.eu]

I don't have x64 OS yet... so I cannot even test my x64 compile, but here it is (MSVS9 compile, SSE3): [attachment=5536:oggenc2_x64_test.7z]

Maybe it even works. It lacks FLAC and SRC support, though.

Cool, it encodes correctly (on WinXP64 SP2 - E2200 @ 2,35 GHz - 1 Gb).

Now testing speeds...

[forart.eu]

Very fast graphic reply of tests (i'll explain better tomorrow):



This post has been edited by forart.eu: Dec 12 2009, 13:29

[forart.eu]

OK, here's the encoders explaination:

• All vencs are - of course - the latest build from aoTuV test page by Aoyumi;
• All oggenc2SSE* are from the top of this 3ad by lvqcl;
• oggenc2LM is the patched john33's encoder in this 3ad too;
• oggenc64 is the cited john33's unofficial build;
• oggenc264 is the last build posted here by lvqcl;

Hope that inspires...

[lvqcl]

john33, I tested your Lancer compile with the following input:
8...48kHz/16 bit/mono or stereo, -q2, -q1, ... -q10;
44.1kHz/16bit/stereo, -q2, -q1.9, -q1.8, ...-q9.9, -q10.

Everything is fine but its ENCODE_VENDOR_STRING ends with "\n" again

[john33]

john33, I tested your Lancer compile with the following input:
8...48kHz/16 bit/mono or stereo, -q2, -q1, ... -q10;
44.1kHz/16bit/stereo, -q2, -q1.9, -q1.8, ...-q9.9, -q10.

Everything is fine but its ENCODE_VENDOR_STRING ends with "\n" again

Amended, recompiled and uploaded on same link. Hopefully all is now well!!

[forart.eu]

More details about my above encoding tests:

• Original file: CDImage.wav, 44100 Hz/16 bit, 606MB (636 223 100 bytes) - 1:00:06.707 (159055764 samples), 1411 kbps
• Encoding options: -q0 (~64Kbps)
• Coded files:
  28,3 MB (29.710.652 byte) - oggenc264
  28,3 MB (29.710.723 byte) - oggenc2LM, oggenc264, oggenc2SSE*
  29,1 MB (30.571.832 byte) - all vencs

Ask me for any other doubt/question.

This post has been edited by forart.eu: Dec 15 2009, 09:00

[Compact Dick]

I'll leave you to do the maths, but the speed gain is somewhat alarming!!

Alarming in the most pleasing manner possible

If this appears stable, and I'd like feedback regarding this, then I'll make it generally available on Rarewares.

I've been using your builds for a few weeks now, everything's performing as expected. Thanks, john33 [and lvqcl] for doing the hard work. I'm loving it

[AshenTech]

cant wait to see somebody do an assimbly enhanced version for x64 cpu's, my athlon II 435@phenom II b35 3.5gz loves 64bit

[Dukers]

I completely forgot about this topic....

Fantastic job lvqcl and john33!

[AshenTech]

lvqcl if you still need an x64 os get ahold of me, I may be able to help you out there, I would really like to see a nice well tested/tweaked x64 version of vorbis(not that i dont like the x64 one that we already have, but the more the better)

[lvqcl]

As you can see from the picture in post #29, Lancer doesn't benefit from 64-bitness.

[forart.eu]

No Lancer x64 in the above tests !!!

BTW, according to my tests, x64 build can achieve a similar performances compared to Lancer.
So it would be really interesting to encode with an x64 + Lancer optimizations Vorbis build...

It would also be very interesting to compare compilers' performances, IMHO.

This post has been edited by forart.eu: Jun 15 2010, 09:06

[ilikedirtthe2nd]

nice improvements, but i had a slightly hard time figuring this thead out... could you put the stuff that works up on rarewares.org? maybe with a litte explanation of the different versions.

thanks

[lvqcl]

No Lancer x64 in the above tests !!!

Buuuuut... you said

oggenc264 is the last build posted here by lvqcl;

I think you mean the attachment in the post #27. And this is -- Lancer x64 build.

[forart.eu]

I think you mean the attachment in the post #27. And this is -- Lancer x64 build.

I haven't noticed that it was Lancer build !!!
BTW, if so, it's not a good build 'cause it MUST achieve at least the same performances of x86 build (Lancer has ASM optimizations...).

A question: it's builded from the same oggenc2LM patched sources ?

This post has been edited by forart.eu: Jun 16 2010, 11:22

[lvqcl]

john33's oggenc2LM was compiled with Intel (optimizing) compiler, and mine - with MSVS compiler. That's the only difference.

Also, x64 MSVS compile (aka oggenc264) is slightly faster than x86 MSVS compile (oggenc2SSE2/oggenc2SSE3).

[forart.eu]

@john33: Lame x64 officially available @ RareWares, what about Vorbis ???

[john33]

@john33: Lame x64 officially available @ RareWares, what about Vorbis ???

Essentially because the compiles that I've produced work fine on XP x64, but instantly crash on Windows 7 x64 and I haven't yet managed to figure out why!! If anyone has any ideas, please fire away.

[lvqcl]

Once I encountered a problem (with debug compile) with the following code in encode.c:

CODE

        /* Next 3 lines added to add padding bytes into comment header for tagging space.
         */
        header_comments.packet = realloc(header_comments.packet, header_comments.bytes + opt->padding);
        memset(header_comments.packet + header_comments.bytes, 0, opt->padding);
        header_comments.bytes += opt->padding;

And oggenc2 stopped crashing in debug mode after removing this code.

[john33]

Once I encountered a problem (with debug compile) with the following code in encode.c:

CODE

        /* Next 3 lines added to add padding bytes into comment header for tagging space.
         */
        header_comments.packet = realloc(header_comments.packet, header_comments.bytes + opt->padding);
        memset(header_comments.packet + header_comments.bytes, 0, opt->padding);
        header_comments.bytes += opt->padding;

And oggenc2 stopped crashing in debug mode after removing this code.

You're absolutely right! Although, why this should be OK in XP x64 and not Windows 7 x64 beats me.

Anyway, the following download is a standard 1.3.1 lib compile for x64 and does work on XP x64 and Win 7 x64:

http://www.rarewares.org/files/oggenc2.87-1.3.1-x64.zip

Regarding the above, I have obviously removed the 'Comment Padding' option. Unless there are any tagging programs out there that will add/amend vorbis comments via update-in-place, and I don't believe there are, then this was always of little value anyway.

Edit: If there are no reports of any issues with this, then I'll add it to the normal Rarewares ogg page.

This post has been edited by john33: Jul 4 2010, 15:03

[lvqcl]

IIRC... the address that stored in header_comments.packet is also stored in some other variable. So if realloc doesn't change this address then oggenc works but if it returns different address -- oggenc crashes.

This post has been edited by lvqcl: Jul 4 2010, 16:13

[john33]

IIRC... the address that stored in header_comments.packet is also stored in some other variable. So if realloc doesn't change the address oggenc works but if it returns different address -- oggenc crashes.

Makes sense. Thanks very much for the assist.

[forart.eu]

If there are no reports of any issues with this, then I'll add it to the normal Rarewares ogg page.

Just tested under both x64 Windows (7 and XP), it works.

Another question: what about merging lancer optimizations ?