是否有一个内在函数可以将 __m128i 向量的最后 n 个字节归零? [英] Is there an intrinsic function to zero out the last n bytes of a __m128i vector?
问题描述
给定 n
,我想将 __m128i
向量的最后 n
个字节归零.
Given n
, I want to zero out the last n
bytes of a __m128i
vector.
例如考虑以下 __m128i
向量:
For instance consider the following __m128i
vector:
<代码> 11111111 11111111 11111111 11111111 11111111 11111111 11111111 11111111 11111111 11111111 11111111 11111111 11111111 11111111 11111111 11111111 代码>
将最后的 n = 4
个字节归零后,向量应如下所示:
After zeroing out the last n = 4
bytes, the vector should look like:
<代码> 11111111 11111111 11111111 11111111 11111111 11111111 11111111 11111111 11111111 11111111 11111111 11111111 00000000 00000000 00000000 00000000 代码>
是否有执行此操作的 SSE 内在函数(通过接受 __128i
向量和 n
作为参数)?
Is there a SSE intrinsic function that does this (by accepting a __128i
vector and n
as parameters)?
推荐答案
有多种不依赖于 AVX512 的选项.例如:
There are various options that don't rely on AVX512. For example:
char mask[32] = { 0, 0, 0, 0, 0, 0, 0, 0,
0, 0, 0, 0, 0, 0, 0, 0,
-1, -1, -1, -1, -1, -1, -1, -1,
-1, -1, -1, -1, -1, -1, -1, -1};
__m128i zeroLowestNBytes(__m128i x, uint32_t n)
{
__m128i m = _mm_loadu_si128((__m128i*)&mask[16 - n]);
return _mm_and_si128(x, m);
}
使用 AVX,负载可以成为 vpand
的内存操作数.如果没有 AVX,它仍然可以,使用 movdqu
和 pand
.
With AVX, the load can become a memory operand of the vpand
. Without AVX it's still fine, with movdqu
and pand
.
未对齐的负载通常不是问题,除非它跨越 4K 边界.如果你可以让 mask
32 对齐,那么这个问题就会消失.负载仍会未对齐,但不会达到特定的边缘情况.
The load being unaligned isn't normally a problem, unless it crosses a 4K boundary. If you can get mask
32-aligned then that problem would go away. The load would still be unaligned, but wouldn't hit that particular edge case.
n
是一个 uint32_t
以避免符号扩展.
n
is an uint32_t
to avoid sign-extension.
__m128i zeroLowestNBytes(__m128i x, int n)
{
__m128i threshold = _mm_set1_epi8(n);
__m128i index = _mm_set_epi8(15, 14, 13, 12, 11, 10, 9, 8, 7, 6, 5, 4, 3, 2, 1, 0);
return _mm_andnot_si128(_mm_cmpgt_epi8(threshold, index), x);
}
这避免了未对齐的负载,但这并不重要.更重要的是,它避免了依赖于输入的负载":在具有未对齐负载的版本中,负载取决于 n
.在这个版本中,负载独立于n
.例如,如果此函数是内联的,则允许编译器将其提升出循环.它还允许乱序执行更自由地尽早开始加载,也许在计算 n
之前.
This avoids the unaligned load, but that shouldn't really matter. More importantly, it avoids the "input-dependent load": in the version with the unaligned load, the load depends on n
. In this version, the load is independent of n
. For example, that allows a compiler to hoist it out of a loop, if this function is inlined. It also allows out-of-order execution more freedom to start the load early, perhaps before n
has been computed.
另一方面,它基本上需要 AVX2 或 SSSE3 才能很好地实现 _mm_set1_epi8(n)
.此外,这通常会花费更多指令,这可能会降低吞吐量.延迟应该更好,因为主链"中没有负载.(有一个负载,但它不在一边,它不会将其延迟添加到计算的延迟中).
The flipside is, it basically requires AVX2 or SSSE3 for a decent realization of _mm_set1_epi8(n)
. Also, this normally costs more instructions, which may be worse for throughput. The latency should be better, since there is no load in the "main chain" (there is a load, but it's off to the side, it doesn't add its latency to the latency of the computation).
这篇关于是否有一个内在函数可以将 __m128i 向量的最后 n 个字节归零?的文章就介绍到这了,希望我们推荐的答案对大家有所帮助,也希望大家多多支持IT屋!