Re: Allocation sinking in git HEAD
- From: Mike Pall <mike-1207@xxxxxxxxxx>
- To: luajit@xxxxxxxxxxxxx
- Date: Thu, 5 Jul 2012 01:07:42 +0200
Adam Strzelecki wrote:
> > The logical first step is to add SIMD builtins and hand-vectorize
> > the code. The SIMD builtins etc. are one of the future extensions
> > mentioned in the roadmap.
>
> Cool, thanks for the info. C++ autovec is for me portable and
> forward compatible alternative to non-portable intrinsics or
> vector classes lacking of decent standard. So I believe SIMD
> built-ins for Lua will be even better and more flexible than
> autovectorization.
The plan is to add a higher-level "ffi.vec" module, so you don't
need to use CPU-specific SIMD-builtins in your code.
E.g. if you request a vector type consisting of four doubles (v4d),
you'll get the 256 bit AVX ops. That is, if your CPU has AVX.
Otherwise everything is split into two 128 bit SSE2 ops on two
doubles each (2*v2d). OTOH if you were to try this on ARM, you'd
get four scalar ops on plain doubles (ARM only has SIMD ops for
floats/ints).
So 'v3 = v2 + v1' may be turned into very different code, depending
on the capabilities of the CPU you're running this on. All of that
should happen transparently -- except for performance differences,
of course.
Ok, so you still have to choose a certain SIMD type for your code.
But (hopefully) you won't have to deal with AVX vs. SSE vs. VFP
vs. VMX etc.
--Mike
Other related posts:
