W dniu 2010-09-20 16:03, Mattias Gärtner pisze:
Zitat von Dariusz Mazur <[email protected]>:
[...]
Most of programs are huge count of simple computing. There are many
places, that things may run parallel without pain. But very often
this places are computing short, thus very often overhead of starting
thread is bigger than using second core. But CILK potentially (I can
prof that, only imagine) reduce this overhead. Second : its very
readability solution, and compiler potentially can warning about some
danger . If we try to concurrency smaller things, without loss
efficiency, danger will be minimize
Start a wiki page with a proposal for the syntax.
Create a unit like mtprocs that demonstrates CILK and how the syntax
should be converted by the compiler.
Find someone to implement and, more important, maintain this.
Then there is a good chance to get this into the fpc main branch.
I know, every one say this. But without help (or even permission) of
core developer its none chance to achieve this. I can prepare some of
units, but because i;m rather poor with documentation its little chance
to achieve this.
[...]
Even the mtprocs does that.
http://wiki.lazarus.freepascal.org/Parallel_procedures#Features
realy?
cite:" a procedure or method is executed with an Index running from
an arbitrary StartIndex to an arbitrary EndIndex"
assigning iteration to threads are done on start loop, eq. we must
know count of iteration.
what when iteration 1 is ten times longer than rest?
Then the other threads will sleep OR if this is a nested sub procedure
the other threads will continue on the other work.
For example:
for i:=1 to 10 paralleldo
for j:=1 to 10 paralleldo begin
if j=1 then DoAlotOfWork else DoLittleWork;
end;
With mtprocs all threads will work until the end when the everything
was done except the last DoAlotOfWork, which is then executed by only
one thread.
mtprocs is just a unit, it can not auto parallelize DoAlotOfWork as a
compiler could.
No, its just by architecture. Iterations are divide between threads.
Because thread are limited and costly we must limit granularity.
CILK may work the same, also with loops, but main difference is under.
[...]
I thinking about Fifo queue, I use one Lock free,
its need 16 byte per spawn (func addres, param and result)
time of swich less than 1us
Are you sure about the 1us switch? The algorithm looks more dramatic.
Which algorithm? I say about pop and push form lock free queue. I use
this heavy with my webserver to distributing messages.
--
Darek
--
_______________________________________________
Lazarus mailing list
[email protected]
http://lists.lazarus.freepascal.org/mailman/listinfo/lazarus
