RFR: 8265783: Create a separate library for x86 Intel SVML assembly intrinsics

[Paul Sandoz]

Hi Andrew,

I’ll let Sandhya talk more about the provenance and numerical accuracy. I think we can add more comments/details in that respect.

IMO this is a reasonable compromise, at least for incubation with follow on investigation to determine if we can leverage possible enhancements to Panama FFM (see JEP 414 section on SVML). We would like encourage experimentation of numerical data-parallel algorithms. The performance gains using SVML are compelling in that regard.

Note that the code is only accessed by the Vector API and is not present in a JDK image if the incubating vector module is not present.

Paul. 


> On May 15, 2021, at 3:29 AM, Andrew Haley <aph at redhat.com> wrote:
> 
> On 4/22/21 11:27 PM, Sandhya Viswanathan wrote:
>> Intel Short Vector Math Library (SVML) based intrinsics in native x86 assembly provide optimized implementation for Vector API transcendental and trigonometric methods.
>> These methods are built into a separate library instead of being part of libjvm.so or jvm.dll.
> 
> Is this really acceptable code quality for OpenJDK? No comments, no
> explanation of the derivation of algorithms, no explanation or proofs
> of accuracy. There doesn't even seem to be any source code, just compiler
> output.
> 
> -- 
> Andrew Haley  (he/him)
> Java Platform Lead Engineer
> Red Hat UK Ltd. <https://www.redhat.com>
> https://keybase.io/andrewhaley
> EAC8 43EB D3EF DB98 CC77 2FAD A5CD 6035 332F A671
>