[vector] Unary/Ternary ops

[Al K]

Sounds good, I'd figured escape analysis would reduce the varargs down.
I'll keep that in mind when I start working on masked variants later.
Reduction op currently use Function, but I don't think a masked variant is
required, so could just replace that reference with ToLongFunction.

On Wed, Feb 14, 2018 at 1:03 PM, Paul Sandoz <paul.sandoz at oracle.com> wrote:

> I suggest we create explicit functions, kind of like what you have done,
> but without varargs, be explicit in the arity (that’s what i did for the
> Java implementation for the scalar functions).
>
> VectorUnaryOp
> VectorUnaryMaskOp
> VectorBinaryOp
> VectorBinaryMaskOp
> VectorTernaryOp
> VectorTernaryMaskOp
> VectorBinaryTest // returns boolean
>
> (Being a little verbose in the names does not matter so much as they will
> only appear in the interface declarations and VectorIntrinsic method
> signatures).
>
> I would like to ensure we avoid boxing and packing into arrays to be more
> efficient for the cases where an intrinsic is not supported (what we are
> referring to as graceful degradation).
>
> We could do all that as a follow on patch.
>
> Paul.
>
> On Feb 14, 2018, at 10:47 AM, Al K <someusername3 at gmail.com> wrote:
>
> Any idea what the interfaces for masked variants would look like? I
> imagine we'd have to create additional functional interfaces for each
> op to accept mask argument as well (f.e. unop would be a binop with
> 1 vector 1 mask, binop would be ternop with 2 vectors 1 mask, etc.).
> Would it be better to just have 2 functional interfaces, f.e.,
>
> interface VecOperator<V> {
>     V apply(V... xn);
> }
>
> interface VecMaskOperator<V, M> {
>     V apply(M mask, V... xn);
> }
>
> Something like that work?
>
> Cheers,
> Al
>
> On Wed, Feb 14, 2018 at 12:26 PM, Paul Sandoz <paul.sandoz at oracle.com>
> wrote:
>
>>
>>
>> > On Feb 14, 2018, at 10:12 AM, Al K <someusername3 at gmail.com> wrote:
>> >
>> > Oops, wrong patch link, see
>> > http://cr.openjdk.java.net/~vlivanov/panama/al_miftah/vector
>> _ops/webrev.00/index.html
>> >
>>
>> One comment:
>>
>> - For fma I recommend creating a package private TriFunction rather than
>> packing into an array, then i think the additional casts will be redundant.
>>
>> Thanks,
>> Paul.
>>
>> > On Wed, Feb 14, 2018 at 11:54 AM, Al K <someusername3 at gmail.com> wrote:
>> >
>> >> http://cr.openjdk.java.net/~vlivanov/panama/vector.generaliz
>> ed_intrinsics/webrev.06/
>> >>
>> >>
>> >> This extends the generalized intrinsics changeset with the following:
>> >>
>> >> 1. General ternary ops - currently only fma
>> >>
>> >> 2. General unary ops - abs/neg/sqrt
>> >>
>> >>
>> >> I've replaced the previous inline_vector_binary_operation() in
>> opto/library_call
>> >>
>> >> with a generic inline_vector_nary_operation() to avoid writing
>> specialized impls
>> >>
>> >> for the various n-ary ops.
>> >>
>> >>
>> >> There's also a change in os/linux/os_linux to fix the pointer
>> comparison so clang
>> >>
>> >> would stop complaining.
>> >>
>> >>
>> >> I've signed the OCA, awaiting for it to be reflected on the site.
>> >>
>> >>
>> >> Cheers,
>> >>
>> >> Al
>> >>
>> >>
>>
>>
>
>