Ogg Vorbis optimized for speed

Topic: Ogg Vorbis optimized for speed (Read 257149 times) previous topic - next topic

0 Members and 1 Guest are viewing this topic.

Ogg Vorbis optimized for speed

Reply #225 – 2006-05-30 06:40:07

Is it a chance for any version in near future to work correctly after compilation by GCC (preferably by 4.x branches)?
Latest version which correctly works after compile by gcc 3.3.6 is 20051121 (tested on Athlon XP 2200+ with SSE only support in Ubuntu Linux). All versions after this one give differrent bitrate in generated .ogg (compared to 'standard' - aoTuV b4.51).
Ogg Vorbis is standard lossy codec in Linux world and SSE(2,3) optimized version for Linux is a good support for the community.

Ogg Vorbis optimized for speed

Reply #226 – 2006-05-30 09:28:24

If the bitrate difference is only slight, this is probably normal. IIRC, even the P3-Optimised version of the original AoTuV-encoder gives slightly different results than the generic build.

Ogg Vorbis optimized for speed

Reply #227 – 2006-05-30 12:35:03

The difference is significant, about 2-3 kbps for -q 3 (on my test.wav it gives 112.3 kbps instead of 115.3 kbps in aoTuV 4.51).
I've tried to compile aoTuV by gcc 3.3/3.4/4.0 with or without compiler SSE optimization (-march=athlon-xp -mfpmath=sse) and bitrate was different only in hundredths. So this is definitely a bug.

Ogg Vorbis optimized for speed

Reply #228 – 2006-05-30 14:02:01

Remember that standard aoTuV uses the FPU and Lancer uses SSE. They have different bit-length to represent real nums, and may thus cause different compression.

If in doubt, ABX.

Ogg Vorbis optimized for speed

Reply #229 – 2006-05-30 15:22:09

Already compiled oggenc2/oggDropXPd for Windows give the same bitrate as unoptimized aoTuV, but its were compiled MSVC or similar, not by GCC. I think GCC just untested in new versions of Lancer.

Ogg Vorbis optimized for speed

Reply #230 – 2006-06-15 16:47:00

Lancer 20060616 released.

Edit: Fixed bug in decoding(SSE2)

Ogg Vorbis optimized for speed

Reply #231 – 2006-06-17 09:53:56

Translated by google:
2006/06/16 Lancer 20060616
In one for AMD CPU replacing the CPU distinction processing of the DLL file
Optimizing vorbis_oggpack_look with the inline assembler
Adding SSE3 optimization processing to _mm_add_horz*
SSE optimizing oggdec, it adds
Correcting the trouble of the SSE2 optimization of ov_read_float2pcm
The decoding section of oggdropXPd SSE optimization
Optimization profile for multithread operation for single thread and joint ownership conversion

Ogg Vorbis optimized for speed

Reply #232 – 2006-07-14 11:20:25

Didn't use Vorbis for some time but now I needed an encoder for some previews and tried the Lancer (2006 06 16th) one.

The speed is just sick, thanks to all involved in that
Works great on my Athlon XP, normal SSE version.

Ogg Vorbis optimized for speed

Reply #233 – 2006-07-14 19:14:45

LOL yeah I still got the warm-fuzzy-feeling everytime I encode using Lancer

Ogg Vorbis optimized for speed

Reply #234 – 2006-07-22 19:24:55

New release out today.

Code: [Select]

Changes:
* inline assembly replaces as much as possible to intrinsic
* abolish original memory transfer code in block.c
* bitreverse use looking up table
* fix speed down vorbis_book_decodevv_add's regression in lancer
  20060529
* remove optimization prevention code in vorbis_book_decodevv_add
* pre-calculate tables for triggers in mdct
* simplifying a code in which high frequency removed by mdct_backward
* add decode only funcs: mdct_butterflies_backward,
  dct_butterfly_first_backward
* improve SSE optimization: bark_noise_hybridmp
* add SSE optimization: render_line, vorbis_noise_normalize,
  _vp_noise_normalize
* add SSE3 optimization: mdct_bitreverse
* add pre-calculation code: seed_loop, max_seeds
* optimize: seed_chase
* add SORT16 to psy.c
* auto loop unrolling: SORT8, SORT32 in psy.c
* use lddqu in non SSE environment for unaligned memory load
* improve loop condiution code in inline assembly code
* add t option for oggdec benchmarks (without outputting file)

(courtesy of pub at cyanet.jp)

Good to see the asm being replaced with intrinsics.

Ogg Vorbis optimized for speed

Reply #235 – 2006-07-22 19:57:08

Quote

* add SSE optimization: render_line, vorbis_noise_normalize,
_vp_noise_normalize

SSE optimizations to the noise normalization code ey? that's interesting. Must be very fast

Ogg Vorbis optimized for speed

Reply #236 – 2006-07-24 21:44:53

Hope this guy will work on theora or dirac in the future !
I also hope to see SSE/SSE2/SS3 builds merged together and autoselects the optimizations on fly (like FLAC...)

Ogg Vorbis optimized for speed

Reply #237 – 2006-08-01 22:09:27

Lancer 20060722 is temporarily unavailable due to memory issue(unconfirmed).

Ogg Vorbis optimized for speed

Reply #238 – 2006-08-02 08:53:20

What kind of memory issue? I've been using 20060722 with no problems.

Ogg Vorbis optimized for speed

Reply #239 – 2006-08-02 11:59:54

Quote from: rudefyet on 2006-08-02 08:53:20

What kind of memory issue? I've been using 20060722 with no problems.

It's impossible to download from the page.
If someone could post Lancer 20060722 release (at least "oggenc2.83"), that would be great. Or just a link for dl, of course.

Ogg Vorbis optimized for speed

Reply #240 – 2006-08-02 12:09:14

Lancer 20060802 released.

This is bug fixed release.
20060722 had a memory leak according to author.

Ogg Vorbis optimized for speed

Reply #241 – 2006-08-03 17:36:29

And the crash issue of Lancer DLL is going to be fixed.

Ogg Vorbis optimized for speed

Reply #242 – 2006-08-05 18:23:34

Experimental Lancer 20060806 is up.

Altavista says:
"Being heap memory access error occurs with vorbis_oggpack_write it abolishes, the optimization module in oggpack_write movement
oggpack_look SSE optimization of optimization
_ve_amp cash control processing of modification
accumulate_fit being imperfect with correction
MDCT-RELATED cash control rearranging unnecessary zero data exception processing the SSE optimization description section of deletion _encodepart with correction inspect_error"

Err.. right.

Ogg Vorbis optimized for speed

Reply #243 – 2006-08-05 19:09:10

Wow! Optimization of optimization, that's gotta be fast!

Ogg Vorbis optimized for speed

Reply #244 – 2006-08-05 19:27:09

Quote from: skelly831 on 2006-08-05 19:09:10

Wow! Optimization of optimization, that's gotta be fast!

None of the many "recent" optimizations provided a major speedup for me. Maybe it's 20x encoding with an old lancer and 22x with the sse2-optimized version. I have a Celeron M 1400Mhz, so the MT speedups don't help here. Still, the speed is great. What are your experiences/speeds/setups? Just curious...

Ogg Vorbis optimized for speed

Reply #245 – 2006-08-05 21:31:34

Quote from: MedO on 2006-08-05 19:27:09

What are your experiences/speeds/setups? Just curious...

I'm experiencing a speedup of about ~1.7x with Lancer 20060806 compared to the latest OggEnc from rarewares (average speed is something around 29x vs 17x, significantly depending on the sound material).

EDIT: My system is WinXP SP2 on an Athlon 64 3400+ (Venice).

Ogg Vorbis optimized for speed

Reply #246 – 2006-08-05 23:05:54

If anybody has a Core 2 Duo, I would be interested how fast the SSE2 version is, as this new CPU doesn't break up SSE2 instructions intto 2 parts and thus can compute them directly.

Ogg Vorbis optimized for speed

Reply #247 – 2006-08-05 23:44:34

Athlon XP 3200+ (2.2GHz), SSE:

Lancer 20060616 - 37.85x
Lancer 20060805 - 38.56x

So it's about 1.9% faster, nice but hardly significant.

I also recall that a previous version of lancer was about .5x faster than 06/16. Like linux kernels, not every next version is faster on every system.

The SSE3 multithreading version should fly on Core2. I wouldn't be surprised if it encodes over 100x.

Ogg Vorbis optimized for speed

Reply #248 – 2006-08-06 15:09:32

Quote from: MedO on 2006-08-05 19:27:09

What are your experiences/speeds/setups? Just curious...

Machine: AMD Athlon64 X2 3800+ (2GHz) @ 2.4GHz
Track: Machinae Supremacy - Elite.wav (4m 24.0s)
Options: -q 5

OggEnc (vorbis-tools Rev.10381): 13.9284x (19s)
OggEnc v2.83 (Lancer [20060805](SSE3MT) based on aoTuV b4b): 57.605572x (4.594s)

Ogg Vorbis optimized for speed

Reply #249 – 2006-08-06 18:29:37

It seems that there was something wrong with the latest Lancer-version (Lancer 20060802(Based on aotuv-b4.51_20051117)) for vorbis, because all files were taken (scratched) from the download-server. See actually his page.

Notice