[Libre-soc-bugs] [Bug 676] FORTRAN MAXLOC SVP64 example

[bugzilla-daemon at libre-soc.org]

https://bugs.libre-soc.org/show_bug.cgi?id=676

--- Comment #4 from Luke Kenneth Casson Leighton <lkcl at lkcl.net> ---
(In reply to Luke Kenneth Casson Leighton from comment #0)

this one is remarkably simple: use the characteristics
of mapreduce mode to overwrite nm from a parallel
comparison

> int m2(int * const restrict a, int n) 
> { 
>    int m, nm; 
>    int i; 
> 
>    m = INT_MIN; 
>    nm = -1; 

first prepare offsets

    setvl vl=16
    sv.svstep offs, srcstep

and nm-potential

    li nm, -1
    li nmbase, 0

next use CTR mode

>    for (i=0; i<n; i++) 
>    { 

    setvl r1,0,16,0,1,1 # VL=r1=MIN(MAXVL,CTR)

perform the load of the vector data, elstrided:

     sv.ld/els *a, 8(nmbase)

>        if (a[i] > m) 
>        { 

comparison:

    sv.cmp cr0.gt, *a, m

next, these two both use mapreduce mode with predication

>            m = a[i]; 

    sv.ori/mr/m=gt m,*a,0

>            nm = i;

add-overwriting base with vector-offset into nm

    sv.add/mr/m=gt nm, nmbase, *offs

>        } 

nmbase must be incremented by vl

    add nmbase, nmbase, r1 # r1 contains copy of vl

branch and subtract VL from CTR if not end of loop

    sv.bc/all 16, *0, loop

>     } 
>     return nm; 

    mr r3, nmbase
    blr

the trick here is the multiple overwrites of nm and m, even though
they are scalar, the usual "termination of looping because scalar result"
is switched *off* in mapreduce mode.
thus due to the predication the last successful conditional overwrite
"wins"

the second trick was to use a base nm which increments by VL, from a
vector of sequential constants.

there will be better/other methods, using data-dependent failfirst
will stop at the first compare-fail but will need "continuation"
(2 nested loops) whereas the mapreduce one is really simple but
relies on WaW hazards to be eliminated

-- 
You are receiving this mail because:
You are on the CC list for the bug.