[petsc-users] questions about vectorization

[Jed Brown]

Mark Adams <mfadams at lbl.gov> writes:

> On Sun, Nov 12, 2017 at 11:35 PM, Xiangdong <epscodes at gmail.com> wrote:
>
>> Hello everyone,
>>
>> Can someone comment on the vectorization of PETSc? For example, for the
>> MatMult function, will it perform better or run faster if it is compiled
>> with avx2 or avx512?
>>
>
> There are no AVX instructions that I know of in PETSCc, but there are
> kernels from 3rd parties that are: 1) the 'hypre' AMG solver and 2) the new
> MKL sparse matrix class (wraps MKL, which is probably vectorized).

There are explicit AVX/AVX512F instructions for the SELL matrix format.
Many other operations will be hit-or-miss whether the compiler actually
produces vectorized code or whether there would be any benefit in doing
so (most operations in PETSc are limited by memory bandwidth rather than
vectorization; that includes matrix-vector products on most current
architectures except KNL with MCDRAM, where the SELL format is
beneficial).