Hi @seiko2plus and @Qiyu8.

I am posting this from NumPy slack:

1. How and when to set a macro at dispatch based on intrinsic, ex: libdivide(Note, that macro is not only for including header, functions are defined based on that below in file):

   numpy/numpy/core/include/numpy/libdivide/libdivide.h

   Lines 29 to 35 in e538e11

   	#if defined(LIBDIVIDE_AVX512)
   	#include <immintrin.h>
   	#elif defined(LIBDIVIDE_AVX2)
   	#include <immintrin.h>
   	#elif defined(LIBDIVIDE_SSE2)
   	#include <emmintrin.h>
   	#endif

   , now if I set it say in arithmetic.h in sse, its a bit late and is not picked up.
   Just to bypass this, I am doing it in loops.c.src, doesn't look right though:

   23 /* Use Libdivide for faster division */
   24 #ifdef NPY_HAVE_SSE2_INTRINSICS
   25     #define LIBDIVIDE_SSE2 1
   26 #endif
   27 #include "numpy/libdivide/libdivide.h"

2. What is the best way to do the following with NumPy intrinsics:
   Say I have an array npy_long* of 11 elements. Each element is 8 bytes.
   One __m128i can hold 2 npy_longs. I can use npyv_load_s64 to load into a __m128i(i.e.npyv_s32).
   q2.1: How to best handle the last case with one extra element? Do we use npyv_load_till_s64 ?
   q2.2: How to load result(npyv_s32) into a npy_long* pointer back? npyv_store_s32 npyv_store_s64[EDIT1] is giving me some junk values on write. It's probably just with the way I am sliding but wanted to know what is the right way to do it. [EDIT2] I forgot to reduce the loop length :).
   Example that gave junk:

  891             #ifdef NPY_HAVE_SSE2_INTRINSICS // Just some POC code to try stuff
  892                  is1*=2; // Slide by 2 elements cause of above point
  893                  os1*=2;
  xxx                   n/=2; // [EDIT2], kinda works after this is added.
  894                  BINARY_LOOP_SLIDING {
  895                      npyv_s32 num_as_simdvector = npyv_load_s64((npy_long*)ip1);
  896                      npyv_s32 volatile res = npyv_div_s32(num_as_simdvector, &fast_d1); // Till here works
  897                      npyv_store_s64((npy_long*)op1, res); // [EDIT1] from 32->64
  898                  }
  899             #elif

3. General question: When is npyv_storen_s32 used with strides?

cc: @seberg @mattip