Re: Calc research & funcalc ...

Only in very recent times (the last generation) has typical GPU
hardware become capable of running multiple kernels simultaneously
and/or pre-empting running kernels. This leads to amusing situations -
whereby moving the mouse while a long running sheet calculates would
simply not be able to render - until a Windows / TDR was triggered. We
had to come up with heuristics to break down the CL workload into
bite-sized chunks to avoid this. More modern hardware doesn't have this
issue though.

Interesting stuff :)

And yes, we use CL when we think it makes sense - based on weights and
complexity of the relevant formulae. Otherwise we use the old
interpreter (or now its threaded variant - again depending on complexity).

I like the practical approach using weights and formula complexity. We currently use a full

big-step cost semantics for the formula language to estimate cost in the static partitioning

algorithm. Unfortunately, assigning costs also results in quite a slowdown in the partitioning algorithm.

As one of my colleagues said though, you can think of the process as compiling the spread-

sheet into a partition that can scheduled to run on a set of multicore processors.

The threaded variant you refer to is the one from the slides right? As I understand, the interpreter

is not fully parallel but runs cell arrays/formula groups in parallel.

Mvh/Best regards,

Alexander Asp Bock, PhD student

Computer Science Department

IT University of Copenhagen

On 18 Sep 2018, at 11.55, Michael Meeks <michael.meeks@xxxxxxxxxxxxx> wrote:

Hi Alexander,

On 18/09/18 10:26, Alexander Bock wrote:

I would be delighted to join one of the hackfests if time allows. Is
there a schedule available somewhere?

We typically have one in Hamburg at some stage in the year - which
would be near you; the ESC minutes have details on all of those as they
come up (posted to this list weekly). We also have a larger hackfest in
Brussels before or after FOSDEM - which is an excellent conference to
attend anyway =)

I know of EUSPRIG as well and their horror stories
<http://eusprig.org/horror-stories.htm>

Some good stories there =) Thanks for the list of conferences.

Do you run any of the generated OpenCL kernels in parallel or do you run
a normal sequential recalculation and call the kernel code as necessary?
I would suspect the latter given the information you have provided so far :)

Only in very recent times (the last generation) has typical GPU
hardware become capable of running multiple kernels simultaneously
and/or pre-empting running kernels. This leads to amusing situations -
whereby moving the mouse while a long running sheet calculates would
simply not be able to render - until a Windows / TDR was triggered. We
had to come up with heuristics to break down the CL workload into
bite-sized chunks to avoid this. More modern hardware doesn't have this
issue though.

And yes, we use CL when we think it makes sense - based on weights and
complexity of the relevant formulae. Otherwise we use the old
interpreter (or now its threaded variant - again depending on complexity).

HTH,

Michael.

--
michael.meeks@xxxxxxxxxxxxx <><, GM Collabora Productivity
Hangout: mejmeeks@xxxxxxxxx, Skype: mmeeks
(M) +44 7795 666 147 - timezone usually UK / Europe