Hi,

[email protected] wrote:
>Funny enough patched MacLucasFFTW works slower on QS20 (than PS3) and
>during the execution it uses only 8 SPUs form 16 available ! ?
>I tried using function fftw_cell_set_nspe(16) but still fftw runs on
>maximum 8 spes.

http://www.fftw.org/cell/cellblade/
>IBM QS20 Cell Blade
>IBM Cell Blade: 2 Cell Processors, 3.2GHz Cell Broadband Engine, IBM Cell SDK 
>2.0, Linux 2.6.18, 8 SPEs available.
>The benchmark uses one processor only. 

fftw 8 SPUs max.

I guss fftw_cell_set_nspe(6) or fftw_cell_set_nspe(7) is best on (now) your 
system.

              w = ALLOC_DOUBLES(n/2);
              fftw_cell_set_nspe(4);
              forw =  fftw_plan_dft_1d(n/2,x,x,FFTW_BACKWARD,FFTW_ESTIMATE);

call fftw_cell_set_nspe() before first call fftw_plan_...().



http://netnews.gotdns.org/WallStreet/6351/gfn/lucas.ps3/MacLucasFFTW.ps3.6.patch
ps3$ time ./MacLucasFFTW 216091
M( 216091 )P, n = 16384, MacLucasFFTW v8.1  Ballester

real    4m17.501s
user    4m13.994s
sys     0m3.453s

ps3$ time ./MacLucasFFTW 859433
M( 859433 )P, n = 65536, MacLucasFFTW v8.1  Ballester

real    65m59.341s
user    64m26.565s
sys     1m25.307s
Each sample counts as 0.01 seconds.
  %   cumulative   self              self     total           
 time   seconds   seconds    calls  ms/call  ms/call  name    
 38.87   1442.65  1442.65   859430     1.68     1.68  normalize
 32.90   2663.58  1220.93   859431     1.42     1.42  rftfsub
 28.03   3703.96  1040.38                             fftw_cell_spe_wait_all

ps3$ time ./MacLucasFFTW 32582657
 1 2097152 
 10001 2097152 
^C
real    29m15.234s
user    28m38.115s
sys     0m35.813s
Each sample counts as 0.01 seconds.
  %   cumulative   self              self     total           
 time   seconds   seconds    calls  ms/call  ms/call  name    
 43.78    730.24   730.24    10010    72.95    72.95  normalize
 36.60   1340.67   610.43                             fftw_cell_spe_wait_all  
<- fftw cost
 19.57   1667.10   326.43    10010    32.61    32.61  rftfsub

 .. .  .   .    .     .       .         .          .           .
 Shoichiro Yamada [email protected]
_______________________________________________
Prime mailing list
[email protected]
http://hogranch.com/mailman/listinfo/prime

Reply via email to