Re: [AD] _aligned_malloc()

[AJ <aj@xxxxxxxxxx>]

Elias Pschernig wrote:
> On Sat, 2005-07-23 at 16:06 +1000, AJ wrote:
>
> > A quick search thru the allegro src code, i find lots of malloc() calls.
> >
> > I have mentioned this before, i'd like to replace them with 
> > _aligned_malloc() as this yields greater performance on the newer 
> > (heavily alignment dependant) CPUs.
> >
> > As i dont think anyones going to agree to a simple find-n-replace; how 
> > about replacing them with a macro or a function call, that could have 
> > build specific code to allow me to use _aligned_malloc() and default 
> > back to normal malloc() for non-specific builds.
>
>
> Yes, since quite probably, it will need to be different on different
> platforms. I don't think my libc here has an _aligned_malloc function.
> And from what I remember, there's already a proposal among the A5
> documents which has such an malloc layer, so definitely a good idea.
>
> Do you have some benchmarks which shows the improvement?


i have no evidence of it yet; but i have read plenty of documents that 
suggest it.  Most of the SSE(2) stuff requires aligned data, else a 
performance hit should be expected.
I dont think i'd have much trouble in finding that text in the AMD/Intel 
CPU software design guides published by AMD/Intel, which should be 
enough to convince you of its nessesity.



here is a snippet from  http://bmagic.sourceforge.net/bmsse2opt.html
"SIMD enabled compilers usually provide functions for aligned memory 
allocation. For Intel C++ use the _mm_malloc/_mm_free intrinsics to 
allocate and free aligned blocks of memory. In MSVC v.7 
_aligned_malloc/_aligned_free serves the same purpose. Don't make a 
mistake by calling general purpose free to deallocate pointer created by 
_mm_malloc. It will cause unpredictable behavior."


 Maybe could do
> something like for the asm code on allegro.cc, where it got clear that
> the C version is faster on newer CPUs. If we get something similiar with
> malloc, then we should make the change.