Re: arch-specific template code

[Marc Glisse <marc dot glisse at inria dot fr>]

On Tue, 28 Aug 2012, Ulrich Drepper wrote:

> On Tue, Aug 28, 2012 at 8:33 AM, Marc Glisse <marc.glisse@inria.fr> wrote:
> > Actually, it looks to me like most of it can be rewritten using gcc's vector
> > extensions. _mm_mul_pd(a,b) is just a*b, a[0] gives the first element,
>
> It does?  How?
>
> If I use
>
> typedef double v2df __attribute__ ((vector_size (16)));
> v2df a;
>
> then I cannot use a[0].

Then you are probably not using a current mainline (it works in C since 
4.7 I think, but the C++ support was only added recently (still missing 
constexpr support)).

> gcc has an arch-specific builtin __builtin_ia32_vec_ext_v2df().  There 
> is nothing else I have found which can take its place.

pointer casts work too (and that's what [] is expanded to).

> The horizontal add you already noticed.
> What is also missing is control over memory accesses.  I.e., loads and
> stores from scalar pointers that are possibly unaligned.

Hmm, *mem=vec probably assumes alignment, and memcpy possibly loses too 
much information (although in __builtin_memcpy the compiler has the 
information, so it could optimize). Unions are always horrible. Using 
mem[0]=vec[0]; mem[1]=vec[1]; is the least bad I could find, and indeed it 
isn't using unaligned instructions. Good point.

> There are big gaps in the vector extensions.  In other functions sqrt
> is needed, masking is required, etc.  None of this works (at least
> with the same efficiency) as using the Intel intrinsics.  Waiting for
> the compiler to catch up isn't good IMO.

I can understand that, but it is nice to know exactly what is missing so 
we have a chance to get there eventually. __builtin_sqrt is a good example 
indeed. I thought masking was supposed to work (depending on what that 
means), although possibly only in C.

> As you have seen in the code, it should be easy enough for the SSE code 
> to be converted to use the gcc vector extensions if and when they catch 
> up.

Indeed. I just believe vector extensions should be a goal to keep in mind 
(for later. I hope it is clear all of my questions are for my education 
and are not meant to block your improvements in any way (besides, I am 
not a maintainer)).

Thanks again for the time you took answering,

--
Marc Glisse