Re: [eigen] Accidental push/merge for s390x SIMD :)

[ Thread Index | Date Index | More Archives ]


sorry for late reply,

On Tue, Apr 5, 2016 at 9:44 PM, Konstantinos Margaritis <markos@xxxxxxxxxxx> wrote:
I accidentally pushed my patches to the main tree, while I wanted to just sync my fork with current
tip and then create a pull request. If you think I should revert the commits, I'll try to do that
and create a proper pull request, however the files touched are pretty much s390x specific and apart
from CMake rules and the includes in Eigen/Core nothing else was touched. 

I looked at the list of changes, and no problem on my side. Great job.
In any case, this pretty much completes a first draft of s390x zEC13 SIMD (ZVector) support. Only
int and double traits are vectorized (ZVector does not offer 32-bit vector float support in
hardware). make check gives "99% tests passed, 10 tests failed out of 716" and those 10 don't really
look SIMD specific, but I'll take a closer look later.

What are those 10 tests? Do you consider this support mature enough to be shipped with the 3.3 release, or shall it be delayed for 3.4?
For float, I was
considering adding the traits that are not arithmetic (ie load/store/binary/transpose/etc)
operations and doing the arithmetic (add/sub/mul/div) in scalar, so that perhaps that might give a
small gain over pure scalar. But I need to test and benchmark this first, it's not a priority right

Since ZVector does not address single precision, I guess that users of such systems do not mind about float at all, so I would not waste time for that.
I'm also going to push some VMX/VSX-related fixes in a while. Sadly, I'm out of arm64 hardware for
the time being, so can't really test much.

ok, perhaps some arm64 users could help, anyone?


Mail converted by MHonArc 2.6.19+