Hi Ruoyao,
Okay, that makes sense. That's one method I was considering but wasn't
entirely happy with as for testing I use #defines to control which versions
I build and using different compiles would mean a makefile change.
I'm thinking now that I could do my own #define just before each target
code. I still can't use target_clones and have to triplicate the code but
the code would be identical except with a few conditional bits.
Another method I've considered trying was whether I could have target
specific inlined functions that get inlined in the target_clones making
each different.
I'll keep plugging away at it.
Thanks, Colin
On Thu, Feb 14, 2019 at 11:43 PM Xi Ruoyao <xry111@xxxxxxxxxxxxxxxx> wrote:
> On 2019-02-14 22:54 +0800, Colin Hercus wrote:
> > Hi,
> >
> > I have several functions using SIMD intrinsics with target specific
> > versions for SSE2, SSSE3, SSE4.1, AVX. They all use packed 8bit ints and
> > only 128bit data.
> >
> > As I'm using intrinsics, each version target is slightly different, SSSE3
> > has _mm_shuffle_epi8, SSE4.1 has _mm_blendv_epi8. Because I need to use
> > only a few different intrinsics for each target the code is 95%
> > triplicated. Luckily the AVX code can be a target_clone of the SSE4.1
> > version but the others are slightly different.
> >
> > Is it possible to have code conditional on the target like #ifdef
> > __TARGET_AVX2_
> > that might let me have one source code with a few small conditional
> > segments?
>
> No. Because the preprocessor doesn't know what is "target attribute".
>
> > I tried using #ifdef __AVX__ which works if I select an AVX target on the
> > GCC command but doesn't work in a function with __attribute__ ((target
> > ("avx")))
>
> I suggest to compile one source file with -msse3, -mavx, -mavx2, etc. to
> get
> multiple object files.
> --
> Xi Ruoyao <xry111@xxxxxxxxxxxxxxxx>
> School of Aerospace Science and Technology, Xidian University
>
>
