[petsc-dev] error with flags PETSc uses for determining AVX

[Jed Brown]

Jacob Faibussowitsch <jacob.fai at gmail.com> writes:

>> working out dispatch in MatCreate_XXX() instead of for each function.
>
> Or use compiler extensions for multiversioned functions (I recall GCC has something similar):
> https://clang.llvm.org/docs/AttributeReference.html#target <https://clang.llvm.org/docs/AttributeReference.html#target>

Yes, if you want multiple kernel variants to live in the same source file.

https://gcc.gnu.org/onlinedocs/gcc/Common-Function-Attributes.html#index-target-function-attribute

I believe you can do something like the following to specialize "portable" code.

static inline void f_portable(size_t n, double *x) {
  for (size_t i=0; i<n; i++) x[i] += 1.;
}

void f_haswell(size_t n, double *x) __attribute__((__target__("arch=haswell"))) {
  f_portable(n, x);
}

void f_skylakeavx512(size_t n, double *x) __attribute__((__target__("arch=skylake-avx512"))) {
  f_portable(n, x);
}