访存加速-Speed-up of Memory Access Intensive Program

【参考网址】
关于采用PLD指令进行内存预取减少内存操作指令的等待时间：
https://www.jianshu.com/p/7b3bfc3aed12
关于L1，L2，L3的缓存介绍：
https://www.cnblogs.com/arnoldlu/p/7883663.html

实际中配合NEON的v_load指令似乎并没有起到任何作用，不仅如此反而有些变慢了。U_U

代码如下：

static inline void prefetch_range(uint8_t *addr, size_t len) {
    uint8_t *cp;
    uint8_t *end = addr + len;
    for (cp = addr; cp < end; cp += 1) __builtin_prefetch(cp, 0, 3);
}

static inline void prestore_range(uint8_t *addr, size_t len){
    uint8_t *cp;
    uint8_t *end = addr + len;
    for (cp = addr; cp < end; cp += 1) __builtin_prefetch(cp, 1, 3);
}

void GetPixelOrder_CPU(cv::Mat & order, const cv::Mat & grey){
    uint8_t * data_order = order.data;
    uint8_t * data_grey = grey.data;
    if(order.type() != CV_8UC1 ||
       grey.type() != CV_8UC1 ||
       order.size() != grey.size()){
        CAP_LOGE("[GetPixelOrder] invalid inputs");
        exit(-1);
    }
    int h = grey.size().height;
    int w = grey.size().width;
    int offset;
    int idx;
    uint8_t buf[8]; // window is a 4 x 2 ( width x height ) rectangle
    uint8_t center; // center is in fact the left corner pixel in window
    for(int i=0; i<h-1; ++i){
        offset = i * w;
        idx = 6;
        buf[0] = data_grey[offset];
        buf[1] = data_grey[offset + w];

        buf[2] = data_grey[offset + 1];
        buf[3] = data_grey[offset + w + 1];

        buf[4] = data_grey[offset + 2];
        buf[5] = data_grey[offset + w + 2];
        for(int j=0; j<w-3; ++j){
            // update the outdated pixels with new ones coming in by the right side
            buf[idx] = data_grey[offset + 3];
            buf[idx + 1] = data_grey[offset + w + 3];
            // update window position
            ++offset;
            // update buffer state: position of outdating pixels
            idx = (idx + 2) % 8;
            // the center of window is always the up-left one
            center = buf[idx];
            // calculate the order for center point within current window
            uint8_t counter = 0;
            for(int k=1; k<8; ++k) counter += (buf[(idx + k)%8] > center);
            data_order[offset] = counter << 5;
        }
    }
}


void GetPixelOrder_NEON(cv::Mat & order, const cv::Mat & grey){
    uint8_t * data_order = order.data;
    uint8_t * data_grey = grey.data;
    if(order.type() != CV_8UC1 ||
       grey.type() != CV_8UC1 ||
       order.size() != grey.size()){
        CAP_LOGE("[GetPixelOrder] invalid inputs");
        exit(-1);
    }
    int h = grey.size().height;
    int w = grey.size().width;
    int offset;

    cv::v_uint8x16 v_b_0, v_b_1, v_b_2, v_b_3, v_b_4, v_b_5, v_b_6, v_b_7; // 4(w) x 2(h) window
    cv::v_uint8x16 v_counter;
    cv::v_uint8x16 v_flag;

    int y_end = h - 1;
    int x_end = w - 15;

    for(int i=0; i<y_end; ++i){
        offset = i * w;
        for(int j=0; j<x_end; j+=16){
            // memory hint instructions: PLD Commands to enable data prefetching
            //prefetch_range(data_grey + offset, 24);
            //prefetch_range(data_grey + offset + w, 24);
            //prestore_range(data_order + offset, 16);

            // update window
            v_b_0 = cv::v_load(data_grey + offset);
            v_b_2 = cv::v_load(data_grey + offset + 1);
            v_b_4 = cv::v_load(data_grey + offset + 2);
            v_b_6 = cv::v_load(data_grey + offset + 3);

            v_b_1 = cv::v_load(data_grey + offset + w);
            v_b_3 = cv::v_load(data_grey + offset + w + 1);
            v_b_5 = cv::v_load(data_grey + offset + w + 2);
            v_b_7 = cv::v_load(data_grey + offset + w + 3);

            // calculate the order for center point within current window
            v_counter = cv::v_setall_u8(0);
            v_flag = v_b_0 < v_b_1;
            v_flag = v_flag >> 7;
            v_counter = v_counter + v_flag;
            v_flag = v_b_0 < v_b_2;
            v_flag = v_flag >> 7;
            v_counter = v_counter + v_flag;
            v_flag = v_b_0 < v_b_3;
            v_flag = v_flag >> 7;
            v_counter = v_counter + v_flag;
            v_flag = v_b_0 < v_b_4;
            v_flag = v_flag >> 7;
            v_counter = v_counter + v_flag;
            v_flag = v_b_0 < v_b_5;
            v_flag = v_flag >> 7;
            v_counter = v_counter + v_flag;
            v_flag = v_b_0 < v_b_6;
            v_flag = v_flag >> 7;
            v_counter = v_counter + v_flag;
            v_flag = v_b_0 < v_b_7;
            v_flag = v_flag >> 7;
            v_counter = v_counter + v_flag;
            v_counter = v_counter << 5;
            cv::v_store(data_order + offset, v_counter);
            // update window position
            offset += 16;
        }
    }
}

CPU版本：17-26ms（手机运行不稳定）
NEON版本：1-2ms（绝大部分时间是1ms)
NEON+PLD/PST(预取/存）：3-7ms(不稳定）
将预取prefetch_range以及prestore_range加上后稍微稳定些了，但是速度变慢了。
因此还需要认真思考预取的使用策略，使其生效。