c# - 在 Cudafy 中使用算法在 CPU 上工作时不起作用？

Question

我有一个工作算法可以对图像执行 2D 透视变换。

算法如下:

private Bitmap RescaleImage(double TopLX, double TopLY, double TopRX, double TopRY, double LowLX, double LowLY, double LowRX, double LowRY, int width, int height)
{
    byte[] src_bmp = bmp.ToByteArray();
    byte[] dst_bmp = new byte[src_bmp.Length];


    for (int x = 0; x < width; x++)
    {
        for (int y = 0; y < height; y++)
        {
            /*
             * relative position
             */
            double rx = (double)x / width;
            double ry = (double)y / height;

            /*
             * get top and bottom position
             */
            double topX = TopLX + rx * (TopRX - TopLX);
            double topY = TopLY + rx * (TopRY - TopLY);
            double bottomX = LowLX + rx * (LowRX - LowLX);
            double bottomY = LowLY + rx * (LowRY - LowLY);

            /*
             * select center between top and bottom point
             */
            double centerX = topX + ry * (bottomX - topX);
            double centerY = topY + ry * (bottomY - topY);

            /*
             * store result
             */
            // get fractions
            double xf = centerX - (int)centerX;
            double yf = centerY - (int)centerY;

            // 4 colors - we're flipping sides so we can use the distance instead of inverting it later

            byte cTL0, cTL1, cTL2, cTL3, cTR0, cTR1, cTR2, cTR3, cLL0, cLL1, cLL2, cLL3, cLR0, cLR1, cLR2, cLR3;

            cTL0 = src_bmp[(((int)centerY + 1) * (width * 4)) + (((int)centerX + 1) * 4)];
            cTL1 = src_bmp[((((int)centerY + 1) * (width * 4)) + (((int)centerX + 1) * 4)) + 1];
            cTL2 = src_bmp[((((int)centerY + 1) * (width * 4)) + (((int)centerX + 1) * 4)) + 2];

            cTR0 = src_bmp[(((int)centerY + 0) * (width * 4)) + (((int)centerX + 1) * 4)];
            cTR1 = src_bmp[((((int)centerY + 0) * (width * 4)) + (((int)centerX + 1) * 4)) + 1];
            cTR2 = src_bmp[((((int)centerY + 0) * (width * 4)) + (((int)centerX + 1) * 4)) + 2];

            cLL0 = src_bmp[(((int)centerY + 1) * (width * 4)) + (((int)centerX + 0) * 4)];
            cLL1 = src_bmp[((((int)centerY + 1) * (width * 4)) + (((int)centerX + 0) * 4)) + 1];
            cLL2 = src_bmp[((((int)centerY + 1) * (width * 4)) + (((int)centerX + 0) * 4)) + 2];

            cLR0 = src_bmp[(((int)centerY + 0) * (width * 4)) + (((int)centerX + 0) * 4)];
            cLR1 = src_bmp[((((int)centerY + 0) * (width * 4)) + (((int)centerX + 0) * 4)) + 1];
            cLR2 = src_bmp[((((int)centerY + 0) * (width * 4)) + (((int)centerX + 0) * 4)) + 2];

            // 4 distances
            double dTL = Math.Sqrt(xf * xf + yf * yf);
            double dTR = Math.Sqrt((1 - xf) * (1 - xf) + yf * yf);
            double dLL = Math.Sqrt(xf * xf + (1 - yf) * (1 - yf));
            double dLR = Math.Sqrt((1 - xf) * (1 - xf) + (1 - yf) * (1 - yf));

            // 4 parts
            double factor = 1.0 / (dTL + dTR + dLL + dLR);
            dTL *= factor;
            dTR *= factor;
            dLL *= factor;
            dLR *= factor;

            // accumulate parts
            double r = dTL * (double)cTL0 + dTR * (double)cTR0 + dLL * (double)cLL0 + dLR * (double)cLR0;
            double g = dTL * (double)cTL1 + dTR * (double)cTR1 + dLL * (double)cLL1 + dLR * (double)cLR1;
            double b = dTL * (double)cTL2 + dTR * (double)cTR2 + dLL * (double)cLL2 + dLR * (double)cLR2;

            byte c0 = (byte)(r + 0.5);
            byte c1 = (byte)(g + 0.5);
            byte c2 = (byte)(b + 0.5);

            dst_bmp[(y * (width * 4)) + (x * 4)] = c0;
            dst_bmp[((y * (width * 4)) + (x * 4)) + 1] = c1;
            dst_bmp[((y * (width * 4)) + (x * 4)) + 2] = c2;

        }
    }

    Bitmap bmpOut = dst_bmp.ToBitmap(width, height);
    return bmpOut;
}

这工作正常，输出正是我想要的。但是我做了一个非常微妙的改变，让它使用 Cudafy 在 GPU 上运行:

public void PerformPerspectiveCorrection(PointF TL, PointF TR, PointF LL, PointF LR)
{
    CheckIsSet();
    _gpu.Launch(Width, Height).PerspectiveCorrectionSingleOperation(_gdata.SourceImage, _gdata.ResultImage, TL.X, TL.Y, TR.X, TR.Y, LL.X, LL.Y, LR.X, LR.Y, Width, Height);
}

[Cudafy]
private static void PerspectiveCorrectionSingleOperation(GThread thread, byte[] src_bmp, byte[] dst_bmp, double TopLX, double TopLY, double TopRX, double TopRY, double LowLX, double LowLY, double LowRX, double LowRY, int width, int height)
{

    int x = thread.blockIdx.x;
    int y = thread.threadIdx.x;

    /*
     * relative position
     */
    double rx = (double)x / width;
    double ry = (double)y / height;

    /*
     * get top and bottom position
     */
    double topX = TopLX + rx * (TopRX - TopLX);
    double topY = TopLY + rx * (TopRY - TopLY);
    double bottomX = LowLX + rx * (LowRX - LowLX);
    double bottomY = LowLY + rx * (LowRY - LowLY);

    /*
     * select center between top and bottom point
     */
    double centerX = topX + ry * (bottomX - topX);
    double centerY = topY + ry * (bottomY - topY);

    /*
     * store result
     */
    // get fractions
    double xf = centerX - (int)centerX;
    double yf = centerY - (int)centerY;

    // 4 colors - we're flipping sides so we can use the distance instead of inverting it later

    byte cTL0, cTL1, cTL2, cTR0, cTR1, cTR2, cLL0, cLL1, cLL2, cLR0, cLR1, cLR2;

    cTL0 = src_bmp[(((int)centerY + 1) * (width * 4)) + (((int)centerX + 1) * 4)];
    cTL1 = src_bmp[((((int)centerY + 1) * (width * 4)) + (((int)centerX + 1) * 4)) + 1];
    cTL2 = src_bmp[((((int)centerY + 1) * (width * 4)) + (((int)centerX + 1) * 4)) + 2];

    cTR0 = src_bmp[(((int)centerY + 0) * (width * 4)) + (((int)centerX + 1) * 4)];
    cTR1 = src_bmp[((((int)centerY + 0) * (width * 4)) + (((int)centerX + 1) * 4)) + 1];
    cTR2 = src_bmp[((((int)centerY + 0) * (width * 4)) + (((int)centerX + 1) * 4)) + 2];

    cLL0 = src_bmp[(((int)centerY + 1) * (width * 4)) + (((int)centerX + 0) * 4)];
    cLL1 = src_bmp[((((int)centerY + 1) * (width * 4)) + (((int)centerX + 0) * 4)) + 1];
    cLL2 = src_bmp[((((int)centerY + 1) * (width * 4)) + (((int)centerX + 0) * 4)) + 2];

    cLR0 = src_bmp[(((int)centerY + 0) * (width * 4)) + (((int)centerX + 0) * 4)];
    cLR1 = src_bmp[((((int)centerY + 0) * (width * 4)) + (((int)centerX + 0) * 4)) + 1];
    cLR2 = src_bmp[((((int)centerY + 0) * (width * 4)) + (((int)centerX + 0) * 4)) + 2];

    // 4 distances
    double dTL = Math.Sqrt(xf * xf + yf * yf);
    double dTR = Math.Sqrt((1 - xf) * (1 - xf) + yf * yf);
    double dLL = Math.Sqrt(xf * xf + (1 - yf) * (1 - yf));
    double dLR = Math.Sqrt((1 - xf) * (1 - xf) + (1 - yf) * (1 - yf));

    // 4 parts
    double factor = 1.0 / (dTL + dTR + dLL + dLR);
    dTL *= factor;
    dTR *= factor;
    dLL *= factor;
    dLR *= factor;

    // accumulate parts
    double r = dTL * (double)cTL0 + dTR * (double)cTR0 + dLL * (double)cLL0 + dLR * (double)cLR0;
    double g = dTL * (double)cTL1 + dTR * (double)cTR1 + dLL * (double)cLL1 + dLR * (double)cLR1;
    double b = dTL * (double)cTL2 + dTR * (double)cTR2 + dLL * (double)cLL2 + dLR * (double)cLR2;

    byte c0 = (byte)(r + 0.5);
    byte c1 = (byte)(g + 0.5);
    byte c2 = (byte)(b + 0.5);

    dst_bmp[(y * (width * 4)) + (x * 4)] = c0;
    dst_bmp[((y * (width * 4)) + (x * 4)) + 1] = c1;
    dst_bmp[((y * (width * 4)) + (x * 4)) + 2] = c2;
}

我返回的byte[]全为0。我也尝试直接将值 (255) 应用于 dst_bmp 中的所有字节，它似乎只对一行像素执行操作(1280 字节，因为第一行是 320px，每个像素有 4 个字节)。

有什么想法吗？这真令人气愤!

Answer 1

找到答案，与我的算法完全无关!我正在将浮点值传递给 Cudafy 方法的 double 。这不会发生，因为 GPU 正在“盲目”运行 CUDA，因此在为值分配内存类型之前将访问 float ，就好像它们是 double 而不转换它们。

将 double 参数更改为 float ，效果非常好。