问题（zlw）：
一个_mm128型的寄存器包含 (xx3, xx2, xx1, xx0)这样四个float型浮点数.
想做 xx0 + xx1 + xx2 + xx3这样的计算
提问者自己觉得可以用:
xx = _mm_hadd_ps(xx, _mm_set_zero); // to get (0, 0, xx3+ xx2, xx1+xx0)
xx= _mm_add_ss(xx, _mm_shuffle_ps(xx,xx, _MM_SHUFFLE( 0, 0, 0, 1 )) );
_mm_store_ss(&temp,xx);
亦可以用:
xx= _mm_add_ps(xx, _mm_movehl_ps(xx, xx));
xx= _mm_add_ss(xx, _mm_shuffle_ps(xx, xx, 1));
_mm_store_ss( &temp, xx );
问有没有更好的方法？
回答1（jimdempseyatthecove）：
对同一寄存器使用两次hadd（源和目的寄存器为同一个）
// xx = { xx3, xx2, xx1, xx0 }
xx=_mm_hadd_ps(xx,xx);
// xx = { xx3+xx2, xx1+xx0, xx3+xx2, xx1+xx0}
xx=_mm_hadd_ps(xx,xx);
// xx = { xx2+xx3+xx1+xx0, xx3+xx2+xx1+xx0, xx3+xx2+xx1+xx0, xx3+xx2+xx1+xx0}
虽然最终每个32位里都是四个float型数的累加和，但是并不会带来额外开销
需要额外注意的是，如果在比较老的处理器上使用，需要在两次水平累加中间插入一个load或者store或者非sse操作（why？）
回答2（Brandon Hewitt）：
“最好”的方法依赖实现的平台。
（这里用到了 Intel® Cilk™ Plus ）
使用Intel C++ Composer XE Update 2 版本的编译器，产生的默认 Intel® SSE2代码如下 (i.e. icc -S -c test.c), :
movups (%rdi), %xmm0
movaps %xmm0, %xmm1
movhlps %xmm0, %xmm1
addps %xmm1, %xmm0
movaps %xmm0, %xmm2
shufps $245, %xmm0, %xmm2
addss %xmm2, %xmm0
ret
而支持 Intel SSE4.2 (icc -xSSE4.2 -S -c test.c)平台，则会得到
movups (%rdi), %xmm0
haddps %xmm0, %xmm0
haddps %xmm0, %xmm0
ret
这个查看Cilk汇编指令的方法很好。
Intel对自家产品的优化应该尽心尽力，一般手工优化能达到这样的水平应该说还算不错的。
P.S.//20110807
http://software.intel.com/sites/products/documentation/hpc/composerxe/en-us/cpp/mac/optaps/common/optaps_par_cean_prog.htm
C/C++ Extensions for Array Notations Programming Model
Reductions
A reduction combines array section elements to generate a scalar result. Intel® Cilk™ Plus supports reductions on array sections. It defines a generic reduction function that applies a user-defined dyadic function. It also has nine built-in common reduction functions. The built-in functions are polymorphic functions that accept int, float, and other C basic data type arguments. The names and descriptions of reduction functions are summarized in the table below.
| Function Prototypes | Descriptions |
|---|---|
| __sec_reduce(fun, identity, a[:]) | Generic reduction function. Reduces fun across the array a[:] using identity as the initial value. |
| __sec_reduce_add(a[:]) | Built-in reduction function. Adds values passed as arrays |
| __sec_reduce_mul(a[:]) | Built-in reduction function. Multiplies values passed as arrays |
| __sec_reduce_all_zero(a[:]) | Built-in reduction function. Tests that array elements are all zero |
| __sec_reduce_all_nonzero(a[:]) | Built-in reduction function. Tests that array elements are all non-zero |
| __sec_reduce_any_nonzero(a[:]) | Built-in reduction function. Tests for any array element that is non-zero |
| __sec_reduce_min(a[:]) | Built-in reduction function. Determines the minimum value of array elements |
| __sec_reduce_max(a[:]) | Built-in reduction function. Determines the maximum value of array elements |
| __sec_reduce_min_ind(a[:]) | Built-in reduction function. Determines the index of minimum value of array elements |
| __sec_reduce_max_ind(a[:]) | Built-in reduction function. Determines the index of maximum value of array elements |
The reduction operation can reduce on multiple ranks. The number of ranks reduced depends on the execution context. For a given execution context of rank m and a reduction array section argument with rank n, where n>m, the last n-m ranks of the array section argument are reduced.
| Example |
|---|
sum = __sec_reduce_add(a[:][:]); // sum across the whole array a sum_of_column[:] = __sec_reduce_add(a[:][:]); // sum across the column of a |
