I would love a more "standard" approach for this.
I would aslo love to see the list of other headers that you have
found I only saw 1 one of them. (not counting macstl c++ template
approach)
• I'm not an expert on the matter and maybe I only once a year I
write this (or any assembler) kind of code, and it always takes me
while to get used to the SIMD coding style and to remember the
consequence of 15-bit multiplication with a 16bit multiplication
etc.. (the vec_mradds discussion..) etc..
• Also some of these kind of headers try to generalize, but it would
good to mention different sets that work maybe better on SSE and
other better altivec.
For example today I looked at:
vSInt16 = vec_sl(vSInt16,vSInt16) translating to SSE
and I notice that it would be better to use
vSInt16 = _mm_slli_epi16(vSInt16, int)
So no vector at the end but an integer that shifts all the 8 SInt16
numbers with the same value.
of course static inline functions would be better for type checking
but currently doing the quick and dirty way...
So it would be helpful to have some kind of wiki Doc kind of website
that everybody could add some notes because the documentation that I
found online (microsoft and intel) is very rudimentary and if you
don't do these kind of things every day it takes you about 3 hours
for one line of code. So in my case I would have added some extra
notes to the wiki doc that in some case it is also good to consider
_mm_slli_epi16 in certain cases)
Because I believe more clearer documentation could make it easier for
people that aren't used to this, to consider using it.
so a s
On 23 Nov 2006, at 03:16, Ian Ollmann wrote:
It is probably worth having some sort of discussion on putting
together a collaborative group of some kind on this problem. I've
seen a couple of these headers now from multiple sources. They each
seem to excel in some areas, but none seems to have arrived at the
optimum translation for every instruction, so while one clearly
this problem is one that needs a single solution, there doesn't
seem to be one publicly available yet that does everything well.
Ultimately, the best place for something like this is probably gcc,
provided they are interested, but some other distribution mechanism
might work fine too.
_______________________________________________
Do not post admin requests to the list. They will be ignored.
PerfOptimization-dev mailing list (email@hidden)
Help/Unsubscribe/Update your Subscription:
This email sent to email@hidden
