Vc-XeonPhi

[Matthias Kretz]

On Monday 26 August 2013 13:59:59 Kulakov, Igor wrote:
> 1. I get 13.6% without VC_IMLP definition at all. But in this case
> float_v::Size == 4.

Oh, you're saying Vc with ICC defaults to SSE even if AVX is supported. I'll 
have to make a test for this.

> 2. Same with VC_IMPL=SSE.
> 3. VC_IMPL=AVX gives same result. I check - float_v::Size == 8.

OK, at least Vc with ICC is consistent there.

> 4. on MIC I have seg fault with Release. and Assertion `hitIndex < NHits' 
> in Debug.

That assertion is a good hint why it segfaults in Release mode. You'd need to 
find out whether the assertion is hit because of an incorrect assumption about 
the vector size or because of a bug in Vc.

> 5. on MIC with VC_IMPL=Scalar it works nice.

:)

> 6. We have only gcc 4.4.7 on the node. Vc-0.99 requires at least GCC 4.6.

You can take any other node as well to test Vc-0.99 with GCC. But the GCC 
results are really important. If GCC and ICC agree then there's a high chance 
for a bug in Vc.

> 
> > Maybe I can see a pattern if you give me more results. Also, if you can
> > identify small sections of code where it goes wrong we could try to
> > enhance
> > the Vc unit tests.
> 
> I will look into this later.
> 
> Regards,
> Igor
> 
> On Aug 26, 2013, at 9:28 AM, Matthias Kretz wrote:
> > I'm currently looking into the MIC failures. The tests actually pass, but
> > libstdc++ segfaults after main returned...
> > 
> > Let me make sure I understand your remaining issue:
> > You get 13.6% clones with Vc-XeonPhi using ICC and VC_IMPL=SSE, right?
> > What you you get if you use GCC 4.7/4.8 instead?
> > Any difference for AVX?
> > And how does it look on MIC?
> > 
> > Maybe I can see a pattern if you give me more results. Also, if you can
> > identify small sections of code where it goes wrong we could try to
> > enhance
> > the Vc unit tests.
> > 
> > -Matthias
> > 
> > On Monday 26 August 2013 13:20:45 Kulakov, Igor wrote:
> >> Matthias, thank you.
> >> 
> >> It has improved the situation.
> >> 
> >> 1. I have only 4 test fails with Vc (not 18 as before), all of them are
> >> mic
> >> tests, none are sse. 2. The tracker runs now in Release (do not crashes)
> >> with sse. But with same 13.6% clones.
-- 
Dipl.-Phys. Matthias Kretz

Web:   http://compeng.uni-frankfurt.de/?mkretz

SIMD easy and portable: http://compeng.uni-frankfurt.de/?vc