[FFmpeg-devel] pre discussion around Blackfin dct_quantize_bfin routine

[Guillaume Poirier]

Hi,
Reimar Doeffinger wrote:
> Hello,
> On Mon, Jun 11, 2007 at 10:58:00PM -0400, Marc Hoffman wrote:
>> This is my rough draft it works well with the small precision errors  
>> due to 16bit arithmetic, actually its the same (as far as I can tell)  
>> as the MMX quantizer. I ran my reference fixpoint against the mmx and  
>> it seems to work on the cases I tried, its 1 bit off which seems  
>> about right considering the truncation of the coeffs from 22 to  
>> 16bits.  Can someone look over my shoulder and give a peak. I know I  
>> have to clean up the codes a bit but I wanted to get someone else to  
>> review my work up until now.  I have used inline asm, Michael this  
>> should be pretty straight forward for you to get your head around.   
>> This doubles the performance roughly.
> 
> Hmm. Ok, most is cleanup suggestions so I might be telling you things
> you already know, still...
> 
>> #define clock()      ({ int _t; asm volatile ("%0=cycles;" : "=d" (_t)); _t; })
> 
> If you leave that in the final version you should probably use a less
> collision-freindly name like bfin_clock or so.
> Also that ({ }) construction is a nasty gcc-ism, you could avoid it by
> making it a static always_inline fuction, though that increases the risk
> of the compiler messing up...
> And actually, maybe you could just implement START_TIMER and STOP_TIMER
> in libavutil (or does that do something else)?
> If not, maybe naming it START_PROF and END_PROF is at least more
> consistent, and least the EPROF name is not quite obvious in its meaning


In case you didn't know Marc (I'm sure you do though), you only need
to implement BlackFin version of read_time(void) as found in
libavutil/common.h to get a working support of START_TIMER and
STOP_TIMER on you plateform.

Guillaume