我有一个带有双精度数据的向量存储在AVX寄存器中{a0,a1,a2,a3}.现在,我想制作三个新的向量(也应该存储为__m256d变量),比如b,c和d,包含以下元素:
b = {a0,a1,a0,a0}
c = {a1,a2,a2,a1}
d = {a2,a3,a3,a2}
Run Code Online (Sandbox Code Playgroud)
我怎么能用AVX做到这一点?
使用gcc向量扩展来查看编译器如何执行此操作-march=core-avx2
#include <stdint.h>
typedef double v4df __attribute__((vector_size(32)));
typedef uint64_t v4u64 __attribute__((vector_size(32)));
void foo(v4df a, v4df *b, v4df *c, v4df *d)
{
*b = __builtin_shuffle(a, (v4u64){0, 1, 0, 0});
*c = __builtin_shuffle(a, (v4u64){1, 2, 2, 1});
*d = __builtin_shuffle(a, (v4u64){2, 3, 3, 2});
}
Run Code Online (Sandbox Code Playgroud)
gcc-4.8.4 -march = core-avx2 -O3产生:
vpermpd $4, %ymm0, %ymm1
vmovapd %ymm1, (%rdi)
vpermpd $105, %ymm0, %ymm1
vpermpd $190, %ymm0, %ymm0
vmovapd %ymm1, (%rsi)
vmovapd %ymm0, (%rdx)
vzeroupper
ret
Run Code Online (Sandbox Code Playgroud)
gcc-4.8.4 -march = corei7-avx -O3产生:
vinsertf128 $1, %xmm0, %ymm0, %ymm1
vpermilpd $2, %ymm1, %ymm1
vmovapd %ymm1, (%rdi)
vperm2f128 $33, %ymm0, %ymm0, %ymm1
vshufpd $9, %ymm1, %ymm0, %ymm1
vmovapd %ymm1, (%rsi)
vperm2f128 $49, %ymm0, %ymm0, %ymm0
vpermilpd $6, %ymm0, %ymm0
vmovapd %ymm0, (%rdx)
vzeroupper
ret
Run Code Online (Sandbox Code Playgroud)