[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

Re: FORTH plusminus, was Re: [f-cpu] Winograd DCT on my seul.org account

hi !

cedric wrote:
> > What i need is a DataFlow Graph analysis tool that allows me to
> > handle large chuncks of code with all the automated tedious tasks
> > that i usually do by hand (register allocation, reallocation,
> > interleaving...)
> What did you mean by reallocation is i memory operation or register
> reallocation ? If it's the second, the problem is the same as register
> allocation, and it exist some good algorithm for risc machine (with a lot of
> registers), that can do that better than a lot of human ;-)

do you speak about register coloring ? the complexity grows at least O(n^2)
with the number of registers. the "manual" algo i use (and which can be automated)
is O(n), or linear complexity. give it more registers and it will be even happier
(contrarily to register coloring).

>         For interleaving, it's, if I correctly understand what you mean scheduling,
instruction scheduling, to be exact (look at the end of the DCT paper)

> the problem is not easy,
what i do is straight-forward : it's like taking a string aaaaaaaaaaa and bbbbbbbbbbb
and interleaving it : abababababababababababababababab...

> and need a lot of time to have the right answer.
at least the time to type it ;-)

i know that i am probably the only one who understands myself, and it's
just of problem of communication from my part. However i wish that
my knowledge becomes useful to others.

> But what we can do is detecting where we lost cycle.
> (It's what I plan to do in my asm).

it's like calling the doctor who says : "you're sick" and goes away.

writing asm is ok for small code (like the DCT) but not for large code,
you need to not have to worry about the representation, only about the
function, and asm programming does not make that possible.
peephole optimisation might be easier with your tool, but that's all.
it doesn't solve the "big problems" which often come from superficial

btw, do you come to the Libre Software Meeting in Bordeaux ?
will you present some of our work ?

> A+
>   Cedric
To unsubscribe, send an e-mail to majordomo@seul.org with
unsubscribe f-cpu       in the body. http://f-cpu.seul.org/