theora dev - Aug 2008 - GSoC - Theora multithread decoder

Hi,

This email is to inform what I have been doing since the mid-term.

After the mid-term I worked on a pipeline implementation with OpenMP.

As I said before I did a pipelined implementation of these functions:
(c_dec_dc_unpredict_mcu_plane + oc_dec_frags_recon_mcu_plane) and
(oc_state_loop_filter_frag_rows + oc_state_borders_fill_rows) as
explained in my previous email.
But the results were not good. They were equal the implementation
without pipeline.

http://lampiao.lsc.ic.unicamp.br/~piga/gsoc_2008/comparison.png
http://lampiao.lsc.ic.unicamp.br/~piga/gsoc_2008/speedup.png
http://lampiao.lsc.ic.unicamp.br/~piga/gsoc_2008/systime.png


Next I tried to do some improvements, but without success. I think
that the implementation as it is, could not be improved more
with parallelism.

Another approach could be tried to decode the next frame, as the
current frame is decoding, but this is very chanlenging, because most
of the frames has data dependency of the frame before it.

The gain adding parallelism is 13% with OpenMP. Not so bad.

My project is with few code, I tried to do my best. When I proposed
this project I knew that it would be very difficult. It was
challenging and I have learned a lot, specially  about the OpenMP
library.

Also, I will understand if you don't approve me. The project was very
difficult, more than I expected. I wanted to obtain better results but
I couldn't. I have tried a lot.

-- 
Leonardo de Paula Rosa Piga
http://lampiao.lsc.ic.unicamp.br/~piga

On 14-Aug-08, at 5:49 PM, Leonardo de Paula Rosa Piga wrote:
> This email is to inform what I have been doing since the mid-term.
Thanks, it's great to hear a report. A few comments/questions below.
> As I said before I did a pipelined implementation of these functions:
> (c_dec_dc_unpredict_mcu_plane + oc_dec_frags_recon_mcu_plane) and
> (oc_state_loop_filter_frag_rows + oc_state_borders_fill_rows) as
> explained in my previous email.
> But the results were not good. They were equal the implementation  
> without pipeline.
That's too bad. Do you have any hope that explicit threading would do  
better?
> http://lampiao.lsc.ic.unicamp.br/~piga/gsoc_2008/comparison.png
Why are the error bars so large on three points on this plot? There  
data doesn't distinguish anything in that mcu range.
> My project is with few code, I tried to do my best. When I proposed  
> this project I knew that it would be very difficult. It was  
> challenging and I have learned a lot, specially  about the OpenMP  
> library.
You shouldn't be so discouraged. We were all aware this was a  
difficult proposal, but actually trying things is still useful work.

You should also commit all the code for these tests. That way others  
can reproduce and build on your work. One advantage of open source  
code repositories is that you can publish negative results. Google  
also wants all your code at the end of the program, and it's easier  
if they can pull it from our repository.

  -r