_mm256_setr_epi32() 的延迟和吞吐量

Question

intel intrinsics guide.
上没有关于 _mm256_setr_epi32() 的延迟和吞吐量的信息有谁知道或知道计算方法吗？

非常感谢！

Answer 1

intel自己没有说明，这里https://www.intel.com/content/dam/www/public/us/en/documents/manuals/64-ia-32-architectures-software-developer-instruction-set-reference-manual-325383.pdf table C.2 COMPOSITE INTRINSIC section暗示会有一组指令，这取决于编译器的输入。

在下面的情况下，当编译器的输入是不可预测的table时，它会根据以下指令进行优化编译：

    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

但如果编译器知道输入，它看起来会短得多...

    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

所以延迟和周期甚至无法粗略地知道。在任何情况下，查看汇编程序参考的正确部分我认为是 VINSERTI128 描述（如果按照我上面的 link，当前页 1670）

_mm256_setr_epi32() 的延迟和吞吐量

Latency and Throughput of _mm256_setr_epi32()

c++

intel

latency

avx

throughput