[Ffmpeg-devel] [PATCH] SSE counterpart of ff_imdct_calc_3dn2

Dave Dodge dododge
Thu Aug 24 19:10:18 CEST 2006


On Thu, Aug 24, 2006 at 12:12:22PM -0400, Rich Felker wrote:
> On Thu, Aug 24, 2006 at 08:50:18AM -0400, Dave Dodge wrote:
> > As an aside: unless they've changed it recently, on IA64 icc provides
> > only intrinsics and doesn't allow inline assembly at all.
> 
> This is idiotic.

Well, it was certainly surprising.  However if you know anything about
the nightmarish world of IA64 assembly it becomes at least a little
bit understandable.

> Inline asm is not just for vector ops. Often you want it for
> 'normal' opcodes too.

Intel does provide a ton of non-vector intrinsics: I/O, atomic
operations, basic loads and stores, etc. are all available.  The
problem is that IA64 has a bunch of fancy instruction hints and some
of those (or at least some combinations of those) don't seem to have
full coverage.  For example there's an intrinsic that will produce the
"st8.rel" instruction, but none that can produce "st8.rel.nta" to
suggest alternate cache behavior.

> Of course icc sucks so it's no big deal.

IA64 compilation is an ongoing research project.  I get about the same
performance from gcc and icc but I don't do much floating point, which
I think is icc's strength.

                                                  -Dave Dodge




More information about the ffmpeg-devel mailing list