I'm from the fcpu project (f-cpu.org), the goal is to create a "free
cpu".

One of our main problem is the compiler. I have wrote some algorithme in C to improve matrix multiplication (using gcc on a P3). I have used loop unrolling, software pipelining, a have brake a lot of false dependancies, remove calculation from inner loop. I reorder the loop to align data access and do strip mining. Most of this are processor dependant and should be done by the compiler. I have won an average of 4 times compare to usual algorithme, my best gain was made on an 1.2Ghz athlon processors with 512*512 matrix : x25 !

One of the main problem is the alias problem induice by pointer. In my test, i lose 25% of the performance by using pointer instead of array.

So what in D will be done to improve the use of new instruction ? All new cpu use conditional move (avoid to empty the pipeline), and vector instructions (like MMX and SSE). We can also add the fact to compile for 2 processors and more.

What do you think ?

nicO

Since D supports rectangular arrays, much more aggressive array optimization becomes possible.

nicO wrote in message <3B7DFFFC.7FB87E4F@ifrance.com>...
>I'm from the fcpu project (f-cpu.org), the goal is to create a "free
>cpu".
>
>One of our main problem is the compiler. I have wrote some algorithme in C to improve matrix multiplication (using gcc on a P3). I have used loop unrolling, software pipelining, a have brake a lot of false dependancies, remove calculation from inner loop. I reorder the loop to align data access and do strip mining. Most of this are processor dependant and should be done by the compiler. I have won an average of 4 times compare to usual algorithme, my best gain was made on an 1.2Ghz athlon processors with 512*512 matrix : x25 !
>
>One of the main problem is the alias problem induice by pointer. In my test, i lose 25% of the performance by using pointer instead of array.
>
>So what in D will be done to improve the use of new instruction ? All new cpu use conditional move (avoid to empty the pipeline), and vector instructions (like MMX and SSE). We can also add the fact to compile for 2 processors and more.
>
>What do you think ?
>
>nicO

BTW, Microsoft's compiler has a
"/Oa assume no aliasing"
switch.
kind of playing with fire I assume ;)

"Walter" <walter@digitalmars.com> wrote in message news:9lkk0g$2qdq$2@digitaldaemon.com...
> Since D supports rectangular arrays, much more aggressive array
optimization
> becomes possible.
>
> nicO wrote in message <3B7DFFFC.7FB87E4F@ifrance.com>...
> >I'm from the fcpu project (f-cpu.org), the goal is to create a "free
> >cpu".
> >
> >One of our main problem is the compiler. I have wrote some algorithme in C to improve matrix multiplication (using gcc on a P3). I have used loop unrolling, software pipelining, a have brake a lot of false dependancies, remove calculation from inner loop. I reorder the loop to align data access and do strip mining. Most of this are processor dependant and should be done by the compiler. I have won an average of 4 times compare to usual algorithme, my best gain was made on an 1.2Ghz athlon processors with 512*512 matrix : x25 !
> >
> >One of the main problem is the alias problem induice by pointer. In my test, i lose 25% of the performance by using pointer instead of array.
> >
> >So what in D will be done to improve the use of new instruction ? All new cpu use conditional move (avoid to empty the pipeline), and vector instructions (like MMX and SSE). We can also add the fact to compile for 2 processors and more.
> >
> >What do you think ?
> >
> >nicO
>
>

Walter a écrit :
> 
> Since D supports rectangular arrays, much more aggressive array optimization becomes possible.
> 

Is that enough to introduice strip mining ? Or to change the order of the loop ?

> nicO wrote in message <3B7DFFFC.7FB87E4F@ifrance.com>...
> >I'm from the fcpu project (f-cpu.org), the goal is to create a "free
> >cpu".
> >
> >One of our main problem is the compiler. I have wrote some algorithme in C to improve matrix multiplication (using gcc on a P3). I have used loop unrolling, software pipelining, a have brake a lot of false dependancies, remove calculation from inner loop. I reorder the loop to align data access and do strip mining. Most of this are processor dependant and should be done by the compiler. I have won an average of 4 times compare to usual algorithme, my best gain was made on an 1.2Ghz athlon processors with 512*512 matrix : x25 !
> >
> >One of the main problem is the alias problem induice by pointer. In my test, i lose 25% of the performance by using pointer instead of array.
> >
> >So what in D will be done to improve the use of new instruction ? All new cpu use conditional move (avoid to empty the pipeline), and vector instructions (like MMX and SSE). We can also add the fact to compile for 2 processors and more.
> >
> >What do you think ?
> >
> >nicO

LuigiG wrote in message <9ll862$5ch$1@digitaldaemon.com>...
>BTW, Microsoft's compiler has a
>"/Oa assume no aliasing"
>switch.
>kind of playing with fire I assume ;)


I decided long ago not to support such a switch. It would really only be useful to someone who understood exactly how the internal compiler optimizations really worked, which is likely nobody <g>.

nicO wrote in message <3B7EB156.6D6761E6@ifrance.com>...
>Walter a écrit :
>>
>> Since D supports rectangular arrays, much more aggressive array
optimization
>> becomes possible.
>>
>
>Is that enough to introduice strip mining ? Or to change the order of the loop ?


I think so.

"Walter" <walter@digitalmars.com> wrote in message news:9lmlj0$111a$3@digitaldaemon.com...
>
> LuigiG wrote in message <9ll862$5ch$1@digitaldaemon.com>...
> >BTW, Microsoft's compiler has a
> >"/Oa assume no aliasing"
> >switch.
> >kind of playing with fire I assume ;)
>
>
> I decided long ago not to support such a switch. It would really only be useful to someone who understood exactly how the internal compiler optimizations really worked, which is likely nobody <g>.
>


Yep,
basically, if you know so much about the compiler that you can use the
noalias switch; you don't the noalias switch anymore.

Walter a écrit :
> 
> nicO wrote in message <3B7EB156.6D6761E6@ifrance.com>...
> >Walter a écrit :
> >>
> >> Since D supports rectangular arrays, much more aggressive array
> optimization
> >> becomes possible.
> >>
> >
> >Is that enough to introduice strip mining ? Or to change the order of the loop ?
> 
> I think so.

Good news !

Walter a écrit :
> 
> nicO wrote in message <3B7EB156.6D6761E6@ifrance.com>...
> >Walter a écrit :
> >>
> >> Since D supports rectangular arrays, much more aggressive array
> optimization
> >> becomes possible.
> >>
> >
> >Is that enough to introduice strip mining ? Or to change the order of the loop ?
> 
> I think so.

And what about vector computing ? (for the use of MMX and SSE)

nicO wrote in message <3B801890.5A7216C7@ifrance.com>...
>And what about vector computing ? (for the use of MMX and SSE)

I don't know.