Parallel Programming Assignment VI
===

## Q1: Explain your implementation. How do you optimize the performance of convolution?

**每個Kernel計算一個Pixel的Convolution計算**，所以`Global Work Size = imageWidth * imageHeight`，Kernel內(如下)使用Loacl Variable: `imageLTIdx, imageIdx, filterLTIdx, filterIdx`來更新陣列的Index，而不是每次都使用乘法運算重算Index。

__kernel void convolution(__global float *outputImage, __global float *inputImage, __constant float *filter, int imageWidth, int imageHeight, int filterWidth)
{
    int halfFilterWidth = filterWidth / 2;
    float sum = 0;
    int x = get_global_id(0);
    int y = get_global_id(1);

    int imageLTIdx = (y - halfFilterWidth) * imageWidth + x - halfFilterWidth;
    int filterLTIdx = 0;
    for (int r = -halfFilterWidth; r <= halfFilterWidth; r++)
    {
        if (y + r >= 0 && y + r < imageHeight)
        {
            int imageIdx = imageLTIdx;
            int filterIdx = filterLTIdx;
            for (int c = -halfFilterWidth; c <= halfFilterWidth; c++, imageIdx++, filterIdx++)
            {
                if (x + c >= 0 && x + c < imageWidth)
                {
                    float factor = filter[filterIdx];
                    if (factor != 0.0)
                    {
                        sum += (inputImage[imageIdx] * factor);
                    }
                }
            }
        }
        imageLTIdx += imageWidth;
        filterLTIdx += filterWidth;
    }
    outputImage[y * imageWidth + x] = sum;
}

## Q2: Rewrite the program using CUDA.
### (1) Explain your CUDA implementation
CUDA與OpenCL一樣(資料傳輸方式也一樣，資料搬移->計算->資料搬移)，**每個Kernel計算一個Pixel的Convolution計算**，共會產生影像高度數量的Block，每個Block會產生影像寬度數量的Thread。

### (2) plot a chart to show the performance difference between using OpenCL and CUDA

### (3) explain the result.
由上圖結果可知，**CUDA比OpenCL慢一些**，在Kernel Size增加時速度的差距就越大。