Re: [AD] malloc() vs align_malloc()

[AJ <aj@xxxxxxxxxx>]

> Note that for bitmaps, you'll want to align the line member. You
> shouldn't care about alignment of the dat member, really.

so if i was to ensure the dat was large enough, then set all the line 
ptrs to aligned addresses, the rest of allegro should work just fine ?
or do some functions use  dat[y*pitch+w]  arithmetic?




> In fact, aligned_malloc isn't all that useful to Allegro.


it will be, it will be.



> > -----Original Message-----
> > From: alleg-developers-admin@xxxxxxxxxx [mailto:alleg-
> > developers-admin@xxxxxxxxxx] On Behalf Of AJ
> > Sent: Saturday, February 19, 2005 11:57 PM
> > To: alleg-developers@xxxxxxxxxx
> > Subject: Re: [AD] malloc() vs align_malloc()
> >
> > further research for this topic:
> >
> > these funcs also exist for msvc7
> >
> > _aligned_free()
> > _aligned_malloc()
> > _aligned_realloc()
> >
> > so i guess we need to match eahc aligned_malloc() with an
>
> aligned_free()
>
> > my mingw (gcc 3.4.2) has this:
> >
> >
> > /* These require msvcr70.dll or higher. */
> > #if __MSVCRT_VERSION__ >= 0x0700
> > _CRTIMP void * __cdecl _aligned_offset_malloc(size_t, size_t, size_t);
> > _CRTIMP void * __cdecl _aligned_offset_realloc(void*, size_t, size_t,
> > size_t);
> >
> > _CRTIMP void * __cdecl _aligned_malloc (size_t, size_t);
> > _CRTIMP void * __cdecl _aligned_realloc (void*, size_t, size_t);
> > _CRTIMP void __cdecl _aligned_free (void*);
> > #endif /* __MSVCRT_VERSION__ >= 0x0700 */
> >
> >
> > which means it can only be defined for systems that are using the
>
> msvc7
>
> > runtime.  not sure if this is a requirement for MSVC6 as the MSDN
> > documentaion claims its for systems back to win98.
> >
> > this makes if a bit more difficult to implement cleanly, however if it
> > potentially can improve things like draw_sprite and masked_blit() by
>
> as
>
> > much as 50% its worth doing.
> >
> >
> >
> >
> >
> > AJ wrote:
> >
> > > before i go to the trouble of patching allegro i would like to hear
>
> of
>
> > > any arguments against using align_malloc() instead of malloc()
> > >
> > > SSE/SSE2 instructions require aligned data for performance reasons.
> > >
> > > i'd like to see at minimum all bitmap->dat*  mallocs be aligned to
>
> 16
>
> > > byte boundaries (SSE requirement).
> > >
> > > on Windows, its called _aligned_malloc
> > > details here:
> > > http://msdn.microsoft.com/library/default.asp?url=/library/en-
> >
> > us/vclib/html/vclrf_aligned_malloc.asp
> >
> > > a quick look through allegro, i found many instances were an aligned
> > > malloc *may* improve performace, many of these would not have any
>
> side
>
> > > effects as they are internally viewable vars, and would not break
>
> user
>
> > > viewable structs etc.
> > >
> > > anyone from linux land able to offer the linux view or this
>
> proposal?
>
> > > aj