_mm_set1_ps 和 _mm_set_ps1 有什么区别？

Question

这些功能有什么区别吗？如果不是，为什么？

__m128 __mm_set1_ps(float a)
__m128 __mm_set_ps1(float a)

Intel Intrinsics Guide网站上的两种描述是相同的。谢谢你。

Answer 1

有零差异。 _mm_set1_ps是惯用的，使用它。

例如，clang 的 xmmintrin.h 定义set_ps1为set1_ps：

static __inline__ __m128 __DEFAULT_FN_ATTRS
_mm_set_ps1(float __w)
{
    return _mm_set1_ps(__w);
}

历史猜测

My guess is that Intel just hadn't settled on a naming scheme back in the early days of SSE1, and switched to _mm_set1_type going forward. But if they'd already documented _mm_set_ps1, they couldn't take it back.

Note that there is no _mm_set_epi321 or _mm_set_ep81 (fortunately)! Thus _mm_set1_ps is idiomatic and follows the same pattern as the other broadcast intrinsics, while _mm_set_ps1 is unusual and will surprise human readers. There is _mm_set_pd1 and _mm_load_pd1, though, and presumably they were introduced at the same time (with SSE2).

I only know about it because I stumbled over it the other day while looking for an intrinsic that would do a strict-aliasing-safe broadcast load, like you could with vpbroadcastd in asm. (There isn't a portable one that compiles efficiently everywhere; compiler support for intrinsics is a mixed bag of braindead pessimizations and missing intrinsics when you try to do anything complicated. Maybe in a few more years _mm_loadu_si32(void*) to zero extend will at least be widely supported..) /end off topic rant.