_mm256_setr_epi32()的延迟和吞吐量

时间:2018-07-05 14:17:53

标签: c++ intel latency avx throughput

intel intrinsics guide上没有有关_mm256_setr_epi32()的延迟和吞吐量的信息。
有谁知道它或知道一种计算方法?

非常感谢!

1 个答案:

答案 0 :(得分:4)

它不是由Intel本身指定的,这里的https://www.intel.com/content/dam/www/public/us/en/documents/manuals/64-ia-32-architectures-software-developer-instruction-set-reference-manual-325383.pdf表C.2 COMPOSITE INTRINSIC部分暗示将有一组指令,具体取决于编译器的输入。

在以下情况下,当编译器无法预测输入时,它会根据以下说明进行优化编译:

    volatile int a = 1, b = 2, c = 3, d = 4, e = 5, f = 6, g = 7, h = 8;
    00007FF71DE91013  mov         dword ptr [rbp+1Ch],1  
    00007FF71DE9101A  mov         dword ptr [rbp+18h],2  
    00007FF71DE91021  mov         dword ptr [rbp+14h],3  
    00007FF71DE91028  mov         dword ptr [rbp+10h],4  
    00007FF71DE9102F  mov         dword ptr [rbp+0Ch],5  
    00007FF71DE91036  mov         dword ptr [rbp+8],6  
    00007FF71DE9103D  mov         dword ptr [rbp+4],7  
    00007FF71DE91044  mov         dword ptr [rbp],8  

    volatile __m256i reg = _mm256_setr_epi32(a,b,c,d,e,f,g,h);
    00007FF71DE9104B  mov         ebx,dword ptr [rbp]  
    00007FF71DE9104E  mov         r11d,dword ptr [g]  
    00007FF71DE91052  mov         r10d,dword ptr [f]  
    00007FF71DE91056  mov         r9d,dword ptr [e]  
    00007FF71DE9105A  mov         r8d,dword ptr [d]  
    00007FF71DE9105E  mov         edx,dword ptr [c]  
    00007FF71DE91061  mov         ecx,dword ptr [b]  
    00007FF71DE91064  mov         eax,dword ptr [a]  
    00007FF71DE91067  vmovd       xmm1,eax  
    00007FF71DE9106B  vpinsrd     xmm1,xmm1,ecx,1  
    00007FF71DE91071  vpinsrd     xmm1,xmm1,edx,2  
    00007FF71DE91077  vmovd       xmm0,r9d  
    00007FF71DE9107C  vpinsrd     xmm0,xmm0,r10d,1  
    00007FF71DE91082  vpinsrd     xmm0,xmm0,r11d,2  
    00007FF71DE91088  vpinsrd     xmm1,xmm1,r8d,3  
    00007FF71DE9108E  vpinsrd     xmm0,xmm0,ebx,3  
    00007FF71DE91094  vinsertf128 ymm0,ymm1,xmm0,1  
    00007FF71DE9109A  vmovdqu     ymmword ptr [rbp+20h],ymm0 

但是如果编译器知道输入,那么它看起来会更短...

    volatile __m256i reg = _mm256_setr_epi32(1,2,3,4,5,6,7,8);
    00007FF7789C100F  vmovdqu     ymm0,ymmword ptr [__ymm@0000000800000007000000060000000500000004000000030000000200000001 (07FF7789C2200h)]  
    00007FF7789C1017  vmovdqu     ymmword ptr [rbp],ymm0

因此,延迟和周期甚至无法大致得知。我认为在任何情况下,汇编器参考中的右侧部分都是VINSERTI128描述(如果按照上面的链接,当前为1670页)