使用ARM NEON执行所需时间比C代码长

我转换了一个Brisk函数，将图像从SSE intrinsics调整为ARM NEON intrinsics，以便在ARM架构上执行它。Brisk使用SSE函数（如果支持），否则使用opencv函数。上证综指当然更快。 我在ARM neon中一步一步地转换了SSE函数，但是当我测量执行时间时，与openCV resize函数的执行时间相比，结果是我的函数慢了0.2ms，而不是0.4ms。代码如下：

苏格兰和南方能源公司：

}

手臂霓虹灯：

void halfsample(const cv::Mat& srcimg, cv::Mat& dstimg){
const unsigned short leftoverCols = ((srcimg.cols%16)/2);// take care with border...
const bool noleftover = (srcimg.cols%16)==0; // note: leftoverCols can be zero but this still false...

// make sure the destination image is of the right size:
//assert(srcimg.cols/2==dstimg.cols);
//assert(srcimg.rows/2==dstimg.rows);
//int32x4_t zero = vdupq_n_s8(0);

// mask needed later:
//register __m128i mask = _mm_set_epi32 (0x00FF00FF, 0x00FF00FF, 0x00FF00FF, 0x00FF00FF);
int32x4_t mask = vdupq_n_s32(0x00FF00FF);
// to be added in order to make successive averaging correct:
int32x4_t ones = vdupq_n_s32(0x11111111);
    print128_numhex(mask);
// data pointers:
int32_t* p1=(int32_t*)srcimg.data;
int32_t* p2=(int32_t*)(srcimg.data+srcimg.cols);
int32_t* p_dest=(int32_t*)dstimg.data;
unsigned char* p_dest_char;//=(unsigned char*)p_dest;
int k=0;
// size:
const unsigned int size = (srcimg.cols*srcimg.rows)/16;
const unsigned int hsize = srcimg.cols/16;
int32_t* p_end=p1+size*4;
unsigned int row=0;
const unsigned int end=hsize/2;
bool half_end;
if(hsize%2==0)
    half_end=false;
else
    half_end=true;
while(p2<p_end){
    k++;
    for(unsigned int i=0; i<end;i++){
        // load the two blocks of memory:
        int32x4_t upper;
        int32x4_t lower;
        if(noleftover){
            upper=vld1q_s32(p1);
            lower=vld1q_s32(p2);
        }
        else{
            upper=vld1q_s32(p1);
            lower=vld1q_s32(p2);
        }

        int32x4_t result1=vaddq_s32(upper, ones);
        result1=vrhaddq_u8(upper, lower);

        // increment the pointers:
        p1=p1+4;
        p2=p2+4;

        // load the two blocks of memory:
        upper=vld1q_s32(p1);
        lower=vld1q_s32(p2);
        int32x4_t result2=vaddq_s32(upper, ones);
        result2=vrhaddq_u8(upper, lower);
        // calculate the shifted versions:
        int32x4_t result1_shifted = vextq_u8(result1,vmovq_n_u8(0),1);
        int32x4_t result2_shifted = vextq_u8(result2,vmovq_n_u8(0),1);
        // pack:
        int32x4_t result= vcombine_u8(vqmovn_u16(vandq_u32(result1, mask)),
                vqmovn_u16(vandq_u32 (result2, mask)));

        int32x4_t result_shifted =  vcombine_u8(vqmovn_u16(vandq_u32 (result1_shifted, mask)),
                vqmovn_u16(vandq_u32(result2_shifted, mask)));
        // average for the second time:
        result=vrhaddq_u8(result,result_shifted);

        // store to memory
        vst1q_s32(p_dest, result);

        // increment the pointers:
        p1=p1+4;
        p2=p2+4;
        p_dest=p_dest+4;
        //p_dest_char=(unsigned char*)p_dest;
    }
    // if we are not at the end of the row, do the rest:
    if(half_end){
        std::cout<<"entra in half_end" << std::endl;
        // load the two blocks of memory:
        int32x4_t upper;
        int32x4_t lower;
        if(noleftover){
            upper=vld1q_s32(p1);
            lower=vld1q_s32(p2);
        }
        else{
            upper=vld1q_s32(p1);
            lower=vld1q_s32(p2);
        }

        int32x4_t result1=vqaddq_s32(upper, ones);
        result1=vrhaddq_u8(upper, lower);

        // increment the pointers:
        p1=p1+4;
        p2=p2+4;

        // compute horizontal pairwise average and store
        p_dest_char=(unsigned char*)p_dest;
        const UCHAR_ALIAS* result=(UCHAR_ALIAS*)&result1;
        for(unsigned int j=0; j<8; j++){
            *(p_dest_char++)=(*(result+2*j)+*(result+2*j+1))/2;
        }
        //p_dest_char=(unsigned char*)p_dest;
    }
    else{
        p_dest_char=(unsigned char*)p_dest;
    }

    if(noleftover){
        row++;
        p_dest=(int32_t*)(dstimg.data+row*dstimg.cols);
        p1=(int32_t*)(srcimg.data+2*row*srcimg.cols);
        //p2=(__m128i*)(srcimg.data+(2*row+1)*srcimg.cols);
        //p1+=hsize;
        p2=p1+hsize*4;
    }
    else{
        const unsigned char* p1_src_char=(unsigned char*)(p1);
        const unsigned char* p2_src_char=(unsigned char*)(p2);
        for(unsigned int k=0; k<leftoverCols; k++){
            unsigned short tmp = p1_src_char[k]+p1_src_char[k+1]+
                    p2_src_char[k]+p2_src_char[k+1];
            *(p_dest_char++)=(unsigned char)(tmp/4);
        }
        // done with the two rows:
        row++;
        p_dest=(int32_t*)(dstimg.data+row*dstimg.cols);
        p1=(int32_t*)(srcimg.data+2*row*srcimg.cols);
        p2=(int32_t*)(srcimg.data+(2*row+1)*srcimg.cols);
    }
}

}

---

ARM和SSE功能的输出完全相同。问题在于执行时间。

您应该知道，无论是内部函数还是内联汇编代码，都不能像本机汇编中手工编写的代码那样完美无瑕

更糟糕的是，有时编译器——特别是像GCC这样的开源编译器——会放入一些不必要的指令，导致管道暂停，这需要花费远远超过十个周期的时间。当这种情况发生在最内部的循环中时，对性能来说是致命的

---

为什么不发布代码的反汇编？有内在问题的人应该先看一看。并尽快停止使用intrinsic

您应该从改进代码开始。第一个noleftover部分与if/else相同。您可以两次分配result1，而不使用第一个。不要使用短裤，因为手臂上有32位单词。这其中的大部分可能是由编译器优化的，但无论如何，您应该清除它们，然后使用像DS-5 Streamline这样的探查器-这将使您的任务更容易。社区/免费版本可用。@auselen Yes在剩余部分中，我复制了2次代码。我对它进行了编辑。关于result1，我的操作与原始代码完全相同，但我仍然不清楚为什么要执行该操作，我只是按原样对其进行了转换。我会像您所说的那样尝试使用探查器。您的代码无法编译。所以我想你被一些ifdef愚弄了。@auselen我肯定它是编译的。我使用这些参数-mfloat abi=softfp-mfpu=neon-flash向量转换。使用DS-5 Streamline prifiling的问题是，我使用的是没有图形界面的beagleboard，我不知道是否可以在上面使用DS-5。你可以在beaglebone上捕获数据并在PC主机上进行检查，很简单。这不是答案。我会在OP发布他的反汇编时给出答案：我如何获得反汇编？这是我第一次使用NEON？

void halfsample(const cv::Mat& srcimg, cv::Mat& dstimg){
const unsigned short leftoverCols = ((srcimg.cols%16)/2);// take care with border...
const bool noleftover = (srcimg.cols%16)==0; // note: leftoverCols can be zero but this still false...

// make sure the destination image is of the right size:
//assert(srcimg.cols/2==dstimg.cols);
//assert(srcimg.rows/2==dstimg.rows);
//int32x4_t zero = vdupq_n_s8(0);

// mask needed later:
//register __m128i mask = _mm_set_epi32 (0x00FF00FF, 0x00FF00FF, 0x00FF00FF, 0x00FF00FF);
int32x4_t mask = vdupq_n_s32(0x00FF00FF);
// to be added in order to make successive averaging correct:
int32x4_t ones = vdupq_n_s32(0x11111111);
    print128_numhex(mask);
// data pointers:
int32_t* p1=(int32_t*)srcimg.data;
int32_t* p2=(int32_t*)(srcimg.data+srcimg.cols);
int32_t* p_dest=(int32_t*)dstimg.data;
unsigned char* p_dest_char;//=(unsigned char*)p_dest;
int k=0;
// size:
const unsigned int size = (srcimg.cols*srcimg.rows)/16;
const unsigned int hsize = srcimg.cols/16;
int32_t* p_end=p1+size*4;
unsigned int row=0;
const unsigned int end=hsize/2;
bool half_end;
if(hsize%2==0)
    half_end=false;
else
    half_end=true;
while(p2<p_end){
    k++;
    for(unsigned int i=0; i<end;i++){
        // load the two blocks of memory:
        int32x4_t upper;
        int32x4_t lower;
        if(noleftover){
            upper=vld1q_s32(p1);
            lower=vld1q_s32(p2);
        }
        else{
            upper=vld1q_s32(p1);
            lower=vld1q_s32(p2);
        }

        int32x4_t result1=vaddq_s32(upper, ones);
        result1=vrhaddq_u8(upper, lower);

        // increment the pointers:
        p1=p1+4;
        p2=p2+4;

        // load the two blocks of memory:
        upper=vld1q_s32(p1);
        lower=vld1q_s32(p2);
        int32x4_t result2=vaddq_s32(upper, ones);
        result2=vrhaddq_u8(upper, lower);
        // calculate the shifted versions:
        int32x4_t result1_shifted = vextq_u8(result1,vmovq_n_u8(0),1);
        int32x4_t result2_shifted = vextq_u8(result2,vmovq_n_u8(0),1);
        // pack:
        int32x4_t result= vcombine_u8(vqmovn_u16(vandq_u32(result1, mask)),
                vqmovn_u16(vandq_u32 (result2, mask)));

        int32x4_t result_shifted =  vcombine_u8(vqmovn_u16(vandq_u32 (result1_shifted, mask)),
                vqmovn_u16(vandq_u32(result2_shifted, mask)));
        // average for the second time:
        result=vrhaddq_u8(result,result_shifted);

        // store to memory
        vst1q_s32(p_dest, result);

        // increment the pointers:
        p1=p1+4;
        p2=p2+4;
        p_dest=p_dest+4;
        //p_dest_char=(unsigned char*)p_dest;
    }
    // if we are not at the end of the row, do the rest:
    if(half_end){
        std::cout<<"entra in half_end" << std::endl;
        // load the two blocks of memory:
        int32x4_t upper;
        int32x4_t lower;
        if(noleftover){
            upper=vld1q_s32(p1);
            lower=vld1q_s32(p2);
        }
        else{
            upper=vld1q_s32(p1);
            lower=vld1q_s32(p2);
        }

        int32x4_t result1=vqaddq_s32(upper, ones);
        result1=vrhaddq_u8(upper, lower);

        // increment the pointers:
        p1=p1+4;
        p2=p2+4;

        // compute horizontal pairwise average and store
        p_dest_char=(unsigned char*)p_dest;
        const UCHAR_ALIAS* result=(UCHAR_ALIAS*)&result1;
        for(unsigned int j=0; j<8; j++){
            *(p_dest_char++)=(*(result+2*j)+*(result+2*j+1))/2;
        }
        //p_dest_char=(unsigned char*)p_dest;
    }
    else{
        p_dest_char=(unsigned char*)p_dest;
    }

    if(noleftover){
        row++;
        p_dest=(int32_t*)(dstimg.data+row*dstimg.cols);
        p1=(int32_t*)(srcimg.data+2*row*srcimg.cols);
        //p2=(__m128i*)(srcimg.data+(2*row+1)*srcimg.cols);
        //p1+=hsize;
        p2=p1+hsize*4;
    }
    else{
        const unsigned char* p1_src_char=(unsigned char*)(p1);
        const unsigned char* p2_src_char=(unsigned char*)(p2);
        for(unsigned int k=0; k<leftoverCols; k++){
            unsigned short tmp = p1_src_char[k]+p1_src_char[k+1]+
                    p2_src_char[k]+p2_src_char[k+1];
            *(p_dest_char++)=(unsigned char)(tmp/4);
        }
        // done with the two rows:
        row++;
        p_dest=(int32_t*)(dstimg.data+row*dstimg.cols);
        p1=(int32_t*)(srcimg.data+2*row*srcimg.cols);
        p2=(int32_t*)(srcimg.data+(2*row+1)*srcimg.cols);
    }
}