NPY_FINLINE npyv_u32 npyv_loadn_u32(const npy_uint32 *ptr, npy_intp stride)
{
    assert(llabs(stride) <= NPY_SIMD_MAXLOAD_STRIDE32);
    const __m256i steps = _mm256_setr_epi32(0, 1, 2, 3, 4, 5, 6, 7);
    const __m256i idx = _mm256_mullo_epi32(_mm256_set1_epi32((int)stride), steps);
    return _mm256_i32gather_epi32((const int*)ptr, idx, 4);
}

NPY_FINLINE npyv_s32 npyv_loadn_s32(const npy_int32 *ptr, npy_intp stride)
{
    __m128i a = _mm_cvtsi32_si128(*ptr);
#ifdef NPY_HAVE_SSE41
    a = _mm_insert_epi32(a, ptr[stride],   1);
    a = _mm_insert_epi32(a, ptr[stride*2], 2);
    a = _mm_insert_epi32(a, ptr[stride*3], 3);
#else
    __m128i a1 = _mm_cvtsi32_si128(ptr[stride]);
    __m128i a2 = _mm_cvtsi32_si128(ptr[stride*2]);
    __m128i a3 = _mm_cvtsi32_si128(ptr[stride*3]);
    a = _mm_unpacklo_epi32(a, a1);
    a = _mm_unpacklo_epi64(a, _mm_unpacklo_epi32(a2, a3));
#endif
    return a;
}