Re: [Discuss-gnuradio] patch to gcell fft

[Charles Swiger]

On Wed, 2008-07-09 at 14:26 -0700, Eric Blossom wrote:
> On Wed, Jul 09, 2008 at 04:38:20PM -0400, Charles Swiger wrote:
> > Here is a patch to add windowing to
> > gcell/src/lib/wrapper/spu/gcs_fft_1d_r2.c , seems quick and snappy:

> No need for all this shuffling...
> 

Well, after running 'spu_timing' analysis on my scheme, there must be a
better way because it shows there are lots of operand stalls in my
'quick and snappy' loop:


                                                               .L5:
0d                                              5678            rotmi   
$32,$7,-31
1d 01                                           -6789           lqx     
$23,$8,$9
0                                                 78            andi    $33,$7,1
0                                                  89           ai      
$6,$6,-1       <- decr loop counter
0  012                                              9           shli    
$30,$33,4
0  01                                                           a       
$31,$32,$7
0   12                                                          ai      $7,$7,1
0    2345                                                       rotmai  
$28,$31,-1
0d    34                                                        ai      
$29,$30,32
1d    --567890                                                  lqx     
$26,$29,$sp
0d       6789                                                   shli    
$27,$28,4
1d       ----012345                                             lqx     
$25,$27,$10
1             -----6789                                         shufb   
$24,$25,$11,$26  <- shuffle bytes
0                   ---012345                                   fm      
$22,$23,$24
0d                      1                                       nop     127
1d                      -----678901                             stqx    
$22,$8,$9
0D                            78                                ai      $8,$8,16
                                                               .L13:
1D                            7890                              brnz    $6,.L5


where '-' are stalls and 'd' are where dual-issue is possible, but would
not occur due to operand dependencies.

--Chuck