AVX - cast from YMM to XMM without overhead?

[Marc Glisse]

On Sun, 24 Jun 2012, Dag Lem wrote:

> Marc Glisse <marc.glisse@inria.fr> writes:
>
>> On Sat, 23 Jun 2012, Dag Lem wrote:
>
> [...]
>
>>> However, gcc does not seem to allow casting between vectors of
>>> different lengths (why?!).
>>
>> Ugly.
>
> So in your opinion,
>
>  v4sf c = (v4sf)a + b;
>  v4sf d = __builtin_ia32_haddps(c, c);
>  *z = (v2sf)d;
>
> is uglier than
>
>  v4sf c = *((v4sf*)&a) + b;
>  v4sf d = __builtin_ia32_haddps(c, c);
>  __builtin_ia32_storelps(z, d);
>
> ???
>
> I suspect that you must be referring to something else than apparent
> aesthetics; can you please tell me where the ugliness is hidden in the
> first example?

Casting a vector to a vector with a different number of elements: I am not 
sure what that is supposed to mean. Casting a pointer: I know this means 
reinterpreting what is in memory there. In C++ you could also consider 
casting to a reference to v4sf (equivalent to the pointer cast) but I am 
not sure that it is supported yet (related to bug 53121).

>>> The problem with this is that in the calculation of "c = a + b", gcc
>>> generates an intermediate instruction (vmovdqa) to store the contents
>>> of the YMM register holding "a" to memory, instead of directly
>>> accessing the low 128 bits via the corresponding XMM register.
>>
>> http://gcc.gnu.org/bugzilla/show_bug.cgi?id=53101
>
> OK, so *((v4sf*)&a) is not supposed to be a no-op, but presently
> isn't?

Strike the "not" in your sentence, and yes. It manages it for *(float*)&a 
(taking the first element) but not for subvectors (yet).

>>> I have also looked into using inline assembly to avoid the overhead,
>>> however as far as I can tell it is not possible to use YMM registers
>>> as parameters.
>>
>> Yes it is, you may want to re-read the doc remembering that the YMM
>> registers are the XMM registers (you just use them more fully).
>
> Hmm, I have both read and re-read the documentation, and I am still at
> a loss on how to do "a + b" in a sane way when "a" is a YMM register
> and "b" is an XMM register, and I want the result in an XMM
> register. I am able to emit something like "vaddps %ymm0, %xmm1,
> %xmm1", but this is not valid assembly.

Ah... I see. Can't think of a nice solution right now, sorry.

>>> Is there any way to cast from a YMM to an XMM register without
>>> incurring any performance penalty?
>>
>> ISTR that _mm256_extractf128_ps(*,0) is properly optimized to nothing
>> when possible (I may misremember).
>
> I tried this now, and gcc simply replaces vextractf128 with an equally
> superfluous vmovaps.

:-(
Similar builtins sometimes work. You could try to file a bug about it if 
there isn't one already.

A single optimistic note: the extra moves often have negligible cost, even 
in a tight loop of a dozen instructions.

-- 
Marc Glisse