[②ARM Neon]: Neon指令归纳

前言
博主在[①ARM Neon]: Neon Intrinsic简介一文中介绍了ARM Neon指令的基本知识，在本篇博客中就归纳总结一些常用的，或者一些特殊好用的Neon指令，持续更新。
Neon指令的查询可以参考使用下面网址：
https://developer.arm.com/architectures/instruction-sets/intrinsics/#f:@navigationhierarchiessimdisa=[Neon]
https://gcc.gnu.org/onlinedocs/gcc-4.6.4/gcc/ARM-NEON-Intrinsics.html

加载存储操作

• vld1/vst1 variants，顺序加载存储，例如vld1q_f32，vst1q_f32:

/* Element/structure load 
   ini{f0, f1, f2, f3} ->  in.val[0] = {f0, f1}
                           in.val[1] = {f2, f3}
*/
  float32x2x2_t in = vld1q_f32((const float*)ini);
  vst1q_f32((float*)ini, in);

• vld2/vst2 variants，交错加载存储，例如vld2q_f32，vst2q_f32:

/* Element/structure load 
   ini{f0, f1, f2, f3} ->  in.val[0] = {f0, f2}
                           in.val[1] = {f1, f3}
*/
  float32x2x2_t in = vld2q_f32((const float*)ini);
  vst2q_f32((float*)ini, in);

• vld3/vst3 variants，交错加载存储，例如vld3q_f32，vst3q_f32:

/* Element/structure load 
   ini{f0, f1, f2, f3,     -> in.val[0] = {f0, f3, f6, f9}
       f4, f5, f6, f7,        in.val[1] = {f1, f4, f7, f10}
       f8, f9, f10, f11}      in.val[2] = {f2, f5, f8, f11}
                          
*/
  float32x4x3_t in = vld2q_f32((const float*)ini);
  vst2q_f32((float*)ini, in);

• 同理也有vld4/vst4 variants，适用于例如 float32x4x4_t 返回数据类型，这里就不作更多复述了。

特殊操作

• vdupq_n，set all lanes to the same value，例如vdupq_n_f32，将一个 float32x4x2_t 向量数据初始化为0：

float32x4x2_t a_val;
/* a_val -> {{0.f, 0.f, 0.f, 0.f},
             {0.f, 0.f, 0.f, 0.f}}
             
*/

a_val.val[0] = vdupq_n_f32(0.f);
a_val.val[1] = vdupq_n_f32(0.f);

基本运算

• vnegq，negate，例如vnegq_f32，将一个 float32x4x2_t 向量数据中的一组数取反：

float32x4x2_t a_val;
/* a_val{{f0, f1, f2, f3},  -> {{ f0,  f1,  f2,  f3},
         {f4, f5, f6, f7}}      {-f4, -f5, -f6, -f7}}
             
*/

a_val.val[1] = vnegq_f32(a_val.val[1]);

• vmlaq/vmlsq，multiply-accumulate/multiply-subtract，例如vmlaq_f32/vmlsq_f32，做一组数的相乘后相加或相减:

float32x4x2_t a_val, b_val, c_val;
/* a_val{{a0, a1, a2, a3},      b_val{{b0, b1, b2, b3},     c_val{{c0, c1, c2, c3},
         {a4, a5, a6, a7}}            {b4, b5, b6, b7}}           {c4, c5, c6, c7}} 
         
   RESULT[i] = a[i] + (b[i] * c[i]), for i= 0 to 3
   
   a_val -> {a0 + b0 * c0, a1 + b1 * c1, a2 + b2 * c2, a2 + b2 * c2},
                       a4,           a5,           a6,           a7}}
   
*/
a_val.val[0] = vmlaq_f32(a_val.val[0], b_val.val[0], c_val.val[0]);

/*
   RESULT[i] = a[i] - (b[i] * c[i]), for i= 0 to 3
    
   a_val -> {a0 - b0 * c0, a1 - b1 * c1, a2 - b2 * c2, a2 - b2 * c2},
                       a4,           a5,           a6,           a7}}
*/
a_val.val[0] = vmlsq_f32(a_val.val[0], b_val.val[0], c_val.val[0]);

• vfmaq_laneq_f32，floating-point fused multiply-add to accumulator (vector)。将某个float32x4_t与另外一个float32x4_t中的单个数据进行broadcast乘法，将结果跟第三个float32x4_t相加。

float32x4_t acc_val, a_val, b_val;
/* acc_val{c0, c1, c2, c3},      a_val{a0, a1, a2, a3},     b_val{b0, b1, b2, b3},

// arguments: (float32x4_t a, float32x4_t b, float32x4_t v, const int lane)
acc_val = vfmaq_laneq_f32(acc_val, a_val, b_val, 1); // lane = 1, 取 b_val 中的 b1

acc_val -> {c0 + a0 * b1, c1 + a1 * b1, c2 + a2 * b1, c3 + a3 * b1}