Wanted: Intel AVX support

[Roman Marynchak]

http://software.intel.com/en-us/avx/

Already present in Sandy Bridge.

AMD promises to support the same in Bulldozer.

I know that there is cl-simd, but I plan to investigate this from scratch.

Ideally, SBCL should figure out whether CPU has AVX or not, and turn the corresponding support in the compiler on/off.

[Stas Boukarev (stassats)]

This shouldn't be limited to AVX. SBCL doesn't use even more common extensions, like SSE2 on x86 (although it does for x86_64). So what needs to be done is some general way to specialize architecture more finely, like -march for GCC

[Christophe Rhodes (csr21-cantab)]

Stas Boukarev <email address hidden> writes:

> This shouldn't be limited to AVX. SBCL doesn't use even more common
> extensions, like SSE2 on x86 (although it does for x86_64). So what
> needs to be done is some general way to specialize architecture more
> finely, like -march for GCC

There exists such a general way: sb-c::*backend-subfeatures*.

Christophe

[Nikodemus Siivola (nikodemus)]

And customize-backend-subfeatures.lisp-expr. They should be documented somewhat better, to say the least, though.

Re. x86/SSE2. I with Paul about this at some point, and IIRC we concluded that we should take a page from linux and have src/compiler/x86-common/ to avoid duplicating things.

[Stas Boukarev (stassats)]

backend-subfeatures is too low level, I meant something like specifying "core 2 duo" and it automatically selects the right subfeatures for that architecture.

[Nikodemus Siivola (nikodemus)]

I think it would be reasonable to have (sb-ext:detect-backend-subfeatures) that uses CPUid to decide.

I can't see us maintaining a processor name -> available subfeatures mapping, though.

[Nikodemus Siivola (nikodemus)]

Similarly, it would be nice to be able to have two easy build options:

* Local build: check processor and detect available subfeatures.

* Compatible build: use the lowest common denominator -- ie. no optional subfeatures.

[Paul Khuong (pvk)]

Looking at the plans in <http://software.intel.com/en-us/forums/showpost.php?p=152451>, AVX2 looks like a full-blown fresh port, with support for three-GPR instructions as well.

[Roman Marynchak (roman-marynchak)]

> Re. x86/SSE2. I with Paul about this at some point, and IIRC we concluded that we should take a page from linux and have
> src/compiler/x86-common/ to avoid duplicating things.

I agree. Will look into this in the scope of this project.

> I think it would be reasonable to have (sb-ext:detect-backend-subfeatures) that uses CPUid to decide.

Right, manual architecture tweaks are not required for SBCL.

> AVX2 looks like a full-blown fresh port

My current goal is to support AVX in Win7 x64 build (but I think that it should work on other platforms too). In case I have something to show (just started this evening, so there is nothing to show), I will mail sbcl-devel or update this page. The plan is to put 8-12 hrs/week into this, so the pace will be a bit slow :)

[Douglas Katzman (dougk)]

removing "in progress" state for lack of attention in 9 years