GPU_BUF双线性上采样坐标计算

最新推荐文章于 2023-04-17 21:45:00 发布

YoungHong1992

最新推荐文章于 2023-04-17 21:45:00 发布

阅读量258

点赞数

分类专栏：开发笔记文章标签：双线性插值 resize upsample ocl kernel

本文链接：https://blog.csdn.net/YoungHong1992/article/details/116134792

版权

开发笔记专栏收录该内容

39 篇文章 1 订阅

订阅专栏

// 双线性插值上采样
__kernel void upsample_2x(省略参数...) {
    int dx = get_global_id(0);
    int dy = get_global_id(1);

    float scale = 0.5f; // 输入/输出
    float sx = (dx + 0.5f) * (float)scale - 0.5f;
    float sy = (dy + 0.5f) * (float)scale - 0.5f;

    int x = floor(sx);
    int y = floor(sy);

    float u = sx - x;
    float v = sy - y;

    if (x < 0) {
        x = 0, u = 0;
    }
    if (x >= width) {
        x = width - 1, u = 0;
    }
    if (y < 0) {
        y = 0, v = 0;
    }
    if (y >= height) {
        y = height - 1, v = 0;
    }

    float u1 = 1.0f - u;
    float v1 = 1.0f - v;

    int x_ = min(x + 1, width - 1);
    int y_ = min(y + 1, height - 1);

    for (int cn = 0; cn < channel; cn++) {
        //读取+计算
        float data00 = convert_float(gaussianMat[mad24(y, width, x) * channel + cn]);
        float data10 = convert_float(gaussianMat[mad24(y, width, x_) * channel + cn]);
        float data01 = convert_float(gaussianMat[mad24(y_, width, x) * channel + cn]);
        float data11 = convert_float(gaussianMat[mad24(y_, width, x_) * channel + cn]);

        half value_reup = convert_half(v1 * (u1 * data00 + u * data10) + v * (u1 * data01 + u * data11));
    }
}

// 双线性插值resize
__kernel void resize(省略参数...)
{
    int dx = get_global_id(0);
    int dy = get_global_id(1);

    __global const half* pSrc = src + mad24(offsetY, src_step, offsetX);

    float sx = (dx + 0.5f) * (float)ifx - 0.5f;
    float sy = (dy + 0.5f) * (float)ify - 0.5f;

    int x = floor(sx);
    int y = floor(sy);

    float u = sx - x;
    float v = sy - y;

    if(x < 0){x = 0, u = 0;}
    if(x >= src_w) {x = src_w - 1, u = 0;}
    if(y < 0) {y = 0, v = 0;}
    if(y >= src_h) {y = src_h - 1, v = 0;}

    float u1 = 1.0f - u;
    float v1 = 1.0f - v;


    int x_ = min(x + 1, src_w - 1);
    int y_ = min(y + 1, src_h - 1);

    float data00 = convert_float(pSrc[mad24(y, src_step,  x )]);
    float data10 = convert_float(pSrc[mad24(y, src_step,  x_)]);
    float data01 = convert_float(pSrc[mad24(y_, src_step, x )]);
    float data11 = convert_float(pSrc[mad24(y_, src_step, x_)]);

    float val = v1 * (u1 * data00 + u * data10) + v * (u1 * data01 + u * data11);
    return;
}