On Wed, May 27, 2020 at 7:37 PM Matthew Knepley <knep...@gmail.com> wrote:
> On Wed, May 27, 2020 at 7:34 PM Jed Brown <j...@jedbrown.org> wrote:
>
>> Mark Adams <mfad...@lbl.gov> writes:
>>
>> > Nvidias's NSight with 2D Q3 and bs=10. (attached).
>>
>> Thanks; this is basically the same as a CPU -- the cost is searching the
>> sorted rows for the next entry. I've long thought we should optimize
>> the implementations to fast-path when the next column index in the
>> sparse matrix equals the next index in the provided block. It'd just
>> take a good CPU test to demonstrate that payoff.
>>
>
> So you first check whether the next index is the one in the set passed in,
> and otherwise
> fall back on the search? Good idea.
>
The existing code seems to have *a line (low=i+1) *that seems to be
trying to exploit consecutive indices but it is not quite right, I don't
think,
This is the existing code fragment (this code has been cloned many times
and there are several instances of this kernel).
I've *added code* that I think might make this do the right thing.
if (col <= lastcol) low = 0;
else high = nrow;
lastcol = col;
while (high-low > 5) {
t = (low+high)/2;
*if (rp[low] == col) high = low+1;else * if (rp[t] > col) high = t;
else low = t;
}
for (i=low; i<high; i++) {
if (rp[i] > col) break; *// delete this check if you don't add new
columns*
if (rp[i] == col) {
ap[i] += value;
* low = i + 1;* goto noinsert;
}
}
I'll experiment with this.
Thanks,
Mark
>
> Matt
>
> --
> What most experimenters take for granted before they begin their
> experiments is infinitely more interesting than any results to which their
> experiments lead.
> -- Norbert Wiener
>
> https://www.cse.buffalo.edu/~knepley/
> <http://www.cse.buffalo.edu/~knepley/>
>
