Floating point performance

Floating point performance

[Matej Kupljen]

Hi

I've built soft float toolchain (with crosstool) and then build
MPlayer with it. The performance is very low. I cannot even
play the mp3 file with MPlayer on DBAU1200 with 400MHz CPU!

I used
GCC: 3.3.5
GLIBC: 2.3.5
BINUTILS: 2.16.1

I did some profiling and this is what I get:
-----------------------------------------------------------------------------------
Flat profile:

Each sample counts as 0.01 seconds.
  %   cumulative   self              self     total           
 time   seconds   seconds    calls   s/call   s/call  name    
 18.09     44.70    44.70                             __pack_f
 17.30     87.44    42.74                             _fpadd_parts
 15.82    126.54    39.10                             __mulsf3
 15.28    164.31    37.77                             __unpack_f
  6.49    180.36    16.05                             _fpadd_parts
  5.30    193.45    13.09    17424     0.00     0.00  dct64_1
  4.75    205.18    11.73                             __subsf3
  4.04    215.16     9.98                             __addsf3
  3.68    224.26     9.10    17424     0.00     0.00  synth_1to1
  1.90    228.96     4.70                             __pack_d
  1.20    231.92     2.96                             __unpack_d
  1.17    234.80     2.88                             __muldf3
-----------------------------------------------------------------------------------

As it seems, all the time is spent in the FP routines. 
I decided to use SF toolchain, because they should be faster
then FPU emulator (at least it is on ARM), but did no test
this with emulator.

Does anybody use MPlayer on MIPS? 
What performance do you get?
Any other suggestions?

BR,
Matej

Re: Floating point performance

[Jerry]

>[In reply to "Floating point performance" from Matej Kupljen <matej.kupljen@ultra.si> to linux-mips@linux-mips.org <linux-mips@linux-mips.org>  27.09.2005 11:17]

> Hi

> I've built soft float toolchain (with crosstool) and then build
> MPlayer with it. The performance is very low. I cannot even
> As it seems, all the time is spent in the FP routines.
> I decided to use SF toolchain, because they should be faster
> then FPU emulator (at least it is on ARM), but did no test
> this with emulator.

Decoding of mp3 (and/or other media) with FP routines is extremely
slow on systems w/o fp processor. No matter is it "soft-float" or
"kernel-trap emulator". Both cases it is the emulator.(but SF is a
litthe faster because it has no "trap" overhead).

> Does anybody use MPlayer on MIPS? 
> What performance do you get?

I used mplayed with libmad (libmad uses only integer ops to decode
mp3). It consumes about 10-15% of cpu time (au1200) to play mp3 file.



   ()_()
 -( ^,^ )- -[21398845]- -<The Bat! 3.51.10>- -<27.09.2005 11:33>-
  (") (")

RE: Floating point performance

[Ulrich Eckhardt]

(Sorry if this mail is garbled, I'm forced to use a sub-par client)

Matej Kupljen wrote:
> I've built soft float toolchain (with crosstool) and then build
> MPlayer with it. The performance is very low. I cannot even play the
> mp3 file with MPlayer on DBAU1200 with 400MHz CPU!
[...]
> Any other suggestions?

I'm not sure what you are doing, but if you only want to play music, I'd use Ogg Vorbis instead, which has a decoder that only uses integer arithmetic for exactly the case of FPU-less machines and the Au1200. I could also imagine an MP3 decoder written for integer only being written somewhere, but I don't know anything about it.

Uli

RE: Floating point performance

[Ulrich Eckhardt]

(Sorry if this mail is garbled, I'm forced to use a sub-par client)

Matej Kupljen wrote:
> I've built soft float toolchain (with crosstool) and then build
> MPlayer with it. The performance is very low. I cannot even play the
> mp3 file with MPlayer on DBAU1200 with 400MHz CPU!
[...]
> Any other suggestions?

I'm not sure what you are doing, but if you only want to play music, I'd use Ogg Vorbis instead, which has a decoder that only uses integer arithmetic for exactly the case of FPU-less machines and the Au1200. I could also imagine an MP3 decoder written for integer only being written somewhere, but I don't know anything about it.

Uli

RE: Floating point performance

[Matej Kupljen]

Hi

> > I've built soft float toolchain (with crosstool) and then build
> > MPlayer with it. The performance is very low. I cannot even play the
> > mp3 file with MPlayer on DBAU1200 with 400MHz CPU!
> [...]
> > Any other suggestions?
> 
> I'm not sure what you are doing, but if you only want to play music, 
> I'd use Ogg Vorbis instead, which has a decoder that only uses integer 
> arithmetic for exactly the case of FPU-less machines and the Au1200. 
> I could also imagine an MP3 decoder written for integer only being 
> written somewhere, but I don't know anything about it.

Yes, I can use madplay (libmad) for music only, which uses int
arithmetics (also special version for MIPS).

But I also want to play video and currently I am testing this with
MPlayer (maybe I'll add support for MAE, sometime in the future).
Then I found out, that MPlayer can use libmad for MP3 and it
works great know.

Now I'll try to write XV driver for MAE backend so I'll have
HW accelerated Color Space Conversion (form YV12->RGB) and
Scaling. 

I thought that SF *should* be relatively fast, because I have
experience with it on ARM, where Nicolas Pitre wrote amazing 
SF support for the glibc.
How can we speed-up SF on MIPS? 
Does anybody have some suggestions?

BR,
Matej

Re: Floating point performance

[Greg Weeks]

Ulrich Eckhardt wrote:

>(Sorry if this mail is garbled, I'm forced to use a sub-par client)
>
>Matej Kupljen wrote:
>  
>
>>I've built soft float toolchain (with crosstool) and then build
>>MPlayer with it. The performance is very low. I cannot even play the
>>mp3 file with MPlayer on DBAU1200 with 400MHz CPU!
>>    
>>
>[...]
>  
>
>>Any other suggestions?
>>    
>>
>
>I'm not sure what you are doing, but if you only want to play music, I'd use Ogg Vorbis instead, which has a decoder that only uses integer arithmetic for exactly the case of FPU-less machines and the Au1200. I could also imagine an MP3 decoder written for integer only being written somewhere, but I don't know anything about it.
>  
>
mpg123 had a version of the libs for OS-9 that used integer ops only. It 
was in contributions and not the mainline and I don't know how difficult 
it would be to get it running on Linux. It's a dead project now anyway.

Greg Weeks

Re: Floating point performance

[Nigel Stephens]

Matej Kupljen wrote:

>I thought that SF *should* be relatively fast, because I have
>experience with it on ARM, where Nicolas Pitre wrote amazing 
>SF support for the glibc.
>How can we speed-up SF on MIPS? 
>Does anybody have some suggestions?
>  
>

Maybe someone should volunteer to port Nicolas's "amazing SF support" 
from ARM to MIPS. Hint hint.

Nigel

Re: Floating point performance

[Matej Kupljen]

Hi

> Maybe someone should volunteer to port Nicolas's "amazing SF support" 
> from ARM to MIPS. Hint hint.

It is very ARM specific, so I don't know if it would be possible.
I'll look into it, when I find the time.

BR,
Matej

Re: Floating point performance

[Matej Kupljen]

Hi

> Maybe someone should volunteer to port Nicolas's "amazing SF support" 
> from ARM to MIPS. Hint hint.

Take a look at those benchmarks:
http://lists.arm.linux.org.uk/pipermail/linux-arm/2004-March/007214.html
http://lists.arm.linux.org.uk/pipermail/linux-arm/2004-March/007223.html

BR,
Matej

Re: Floating point performance

[Daniel Jacobowitz]

On Thu, Sep 29, 2005 at 12:36:10PM +0100, Nigel Stephens wrote:
> 
> 
> Matej Kupljen wrote:
> 
> >I thought that SF *should* be relatively fast, because I have
> >experience with it on ARM, where Nicolas Pitre wrote amazing 
> >SF support for the glibc.
> >How can we speed-up SF on MIPS? 
> >Does anybody have some suggestions?
> > 
> >
> 
> Maybe someone should volunteer to port Nicolas's "amazing SF support" 
> from ARM to MIPS. Hint hint.

Unless you've got a spare ASE lying around with conditional execution
and a barrel shifter, I don't think this is in the cards.

Which isn't to say that someone couldn't write a good MIPS-specific
implementation, if they were a sufficiently good FP guru.  But
my feeling is that ARM is more prone than MIPS to clever tricks that
are hard for a compiler to generate.

-- 
Daniel Jacobowitz
CodeSourcery, LLC

Re: Floating point performance

[David Daney]

Matej Kupljen wrote:
> Hi
> 
> 
>>>I've built soft float toolchain (with crosstool) and then build
>>>MPlayer with it. The performance is very low. I cannot even play the
>>>mp3 file with MPlayer on DBAU1200 with 400MHz CPU!
>>
>>[...]
>>
>>>Any other suggestions?
>>
>>I'm not sure what you are doing, but if you only want to play music, 
>>I'd use Ogg Vorbis instead, which has a decoder that only uses integer 
>>arithmetic for exactly the case of FPU-less machines and the Au1200. 
>>I could also imagine an MP3 decoder written for integer only being 
>>written somewhere, but I don't know anything about it.
> 
> 
> Yes, I can use madplay (libmad) for music only, which uses int
> arithmetics (also special version for MIPS).
> 
> But I also want to play video and currently I am testing this with
> MPlayer (maybe I'll add support for MAE, sometime in the future).
> Then I found out, that MPlayer can use libmad for MP3 and it
> works great know.
> 

We are using libmad and can play 128kbps MP3s on a mips 300MHz 4kc code 
with about 10-15% CPU utilization.

However I think anything but very low resolution low frame rate video is 
out of the question.  You need several orders of magnitude more 
processing power to make it work.  All of the MIPS based video systems 
that I am aware of (TiVo, Sony and Samsung HDTVs etc.) have dedicated 
hardware video decoders.  There are people talking about doing the 
decoding purely in software, but they would be using something with the 
processing power of a 2GHz Pentium4 with special DSP extensions.  As far 
as I know these are just paper designs and are not yet in production.

I guess what I am trying to say is that even if you could make the soft 
float library four times faster, you would still be no where close to 
being able to decode video.

I declare it impossible, but encourage you to prove me wrong.

David Daney.

Re: Floating point performance

[Dominik 'Rathann' Mierzejewski]

On Thu, Sep 29, 2005 at 01:16:40PM +0200, Matej Kupljen wrote:
> Hi
> 
> > > I've built soft float toolchain (with crosstool) and then build
> > > MPlayer with it. The performance is very low. I cannot even play the
> > > mp3 file with MPlayer on DBAU1200 with 400MHz CPU!
> > [...]
> > > Any other suggestions?
> > 
> > I'm not sure what you are doing, but if you only want to play music, 
> > I'd use Ogg Vorbis instead, which has a decoder that only uses integer 
> > arithmetic for exactly the case of FPU-less machines and the Au1200. 
> > I could also imagine an MP3 decoder written for integer only being 
> > written somewhere, but I don't know anything about it.
> 
> Yes, I can use madplay (libmad) for music only, which uses int
> arithmetics (also special version for MIPS).
> 
> But I also want to play video and currently I am testing this with
> MPlayer (maybe I'll add support for MAE, sometime in the future).
> Then I found out, that MPlayer can use libmad for MP3 and it
> works great know.
> 
> Now I'll try to write XV driver for MAE backend so I'll have
> HW accelerated Color Space Conversion (form YV12->RGB) and
> Scaling. 

If you're interested in video playback using MPlayer, you may
want to port the assembly parts of its code to MIPS. We (MPlayer
developers) would be grateful if you could contribute your code
back to MPlayer. I'm sure that you can get assistance from our
coders if you have any problems and questions about the code.

Regards,
R.

-- 
Dominik 'Rathann' Mierzejewski <rathann*at*icm.edu.pl>
Interdisciplinary Centre for Mathematical and Computational Modelling
Warsaw University  |  http://www.icm.edu.pl  |  tel. +48 (22) 5540810