使用 NEON 内在函数加速浮点的高斯模糊

Question

我想在 arm 中加速算法的某些部分（高斯模糊过滤器），我决定将 SIMD 与 NEON intrinsic 一起使用。 但不幸的是，用 NEON 编写的代码比纯 cpp 代码慢得多。 这些是纯 cpp 代码和 NEON 版本的代码。 可以改进吗？

纯cpp：

inline void GaussianBlur_5x5_row(const float __restrict_arr *in, float __restrict_arr *out, const unsigned int cols)
{
    //Left columns
    out[0] = (in[0]+in[2])*0.054488685f + (in[0]+in[1])*0.24420135f + in[0]*0.40261996f;
    out[1] = (in[0]+in[3])*0.054488685f + (in[0]+in[2])*0.24420135f + in[1]*0.40261996f;

    //Middle columns 
    for (unsigned int j=2; j<cols-2; j+=1)
    {
        out[j] = (in[j-2]+in[j+2])*0.054488685f + (in[j-1]+in[j+1])*0.24420135f + in[j]*0.40261996f;
        out[j+1] = (in[j-1]+in[j+3])*0.054488685f + (in[j]+in[j+2])*0.24420135f + in[j+1]*0.40261996f;


    }
    //Right columns
    out[cols-2] = (in[cols-4]+in[cols-1])*0.054488685f + (in[cols-3]+in[cols-1])*0.24420135f + in[cols-2]*0.40261996f;
    out[cols-1] = (in[cols-3]+in[cols-1])*0.054488685f + (in[cols-2]+in[cols-1])*0.24420135f + in[cols-1]*0.40261996f;
}

与 NEON 内在：

#include <arm_neon.h>

//kernels coefficients for sigma =1  and  kernel_size =5

float32x4_t coef_1 = { 0.054488685f , 0.122100675f , 0.20130998f , 0.122100675f };
float32x4_t coef_2 = { 0.122100675f , 0.20130998f  , 0.122100675f, 0.054488685f };

float32x4_t load_1 ,
            load_2 ,
            load_3 ,
            help_1 ,
            help_2 ,
            help_3 ;

float32x2_t a     ,
            dst_1 ,
            dst_2 ;

inline void GaussianBlur5x5_row_NEON_128bit_2_itr( const float __restrict_arr *in, float __restrict_arr *out, const unsigned int cols)
{
    //Left columns
    out[0] = (in[0]+in[2])*0.054488685f + (in[0]+in[1])*0.24420135f + in[0]*0.40261996f;
    out[1] = (in[0]+in[3])*0.054488685f + (in[0]+in[2])*0.24420135f + in[1]*0.40261996f;

    //Middle columns
    for (unsigned int j=2; j<cols-2; j+=2)
    {

     load_1 = vld1q_f32( &in[j-2] );
     load_2 = vld1q_f32( &in[j-1] );
     load_3 = vld1q_f32( &in[j  ] );

     help_1 = vmulq_f32( load_1 , coef_1 );
     help_2 = vmulq_f32( load_2 , coef_2 );
     help_3 = vaddq_f32( help_1 , help_2 );
     a      = vadd_f32 ( vget_high_f32( help_3 ) ,  vget_low_f32( help_3 ) );
     dst_1  = vpadd_f32( a , a );

     help_1 = vmulq_f32( load_2 , coef_1 );
     help_2 = vmulq_f32( load_3 , coef_2 );
     help_3 = vaddq_f32( help_1 , help_2 );
     a      = vadd_f32 ( vget_high_f32( help_3 ) ,  vget_low_f32( help_3 ) );
     dst_2  = vpadd_f32( a , a );

     out[j  ] = dst_1[0];
     out[j+1] = dst_2[0];
    }
    //Right columns
    out[cols-2] = (in[cols-4]+in[cols-1])*0.054488685f + (in[cols-3]+in[cols-1])*0.24420135f + in[cols-2]*0.40261996f;
    out[cols-1] = (in[cols-3]+in[cols-1])*0.054488685f + (in[cols-2]+in[cols-1])*0.24420135f + in[cols-1]*0.40261996f;
}

Answer 1

float32x4_t load_1 ,
            load_2 ,
            load_3 ,
            help_1 ,
            help_2 ,
            help_3 ;

float32x2_t a     ,
            dst_1 ,
            dst_2 ;

出于软件工程和性能方面的原因，这些应该是局部变量。

当它们是全局变量时，生成的汇编代码中充斥着冗余的vstr 指令。