在JCUDA中执行cuCtxSync()时"CUDA_ERROR_ILLEGAL_ADDRESS"

Question

我正在学习JCuda并使用JCuda样本学习。

当我使用 JCuda 研究 KMeans 算法代码时，我在执行行 cuCtxSync(( 时得到了一个"CUDA_ERROR_ILLEGAL_ADDRESS";

这让我很困惑。我该如何解决？

这是 KMeansKernel.cu

extern "C"
__global__ void add(int n, float *a, float *b, float *sum)
{
    int i = blockIdx.x * blockDim.x + threadIdx.x;
    if (i<n)
    {
    sum[i] = a[i] + b[i];
    }
}

主要方法(我的类名为"CUDA"(：

public static void main(String[] args){
    // omit some code which input kinds of parameters
    try {
        // Open image file
        BufferedImage bi = ImageIO.read(picFiles);      
        if (bi == null) {
            System.out.println("ERROR: File input error.");
            return;
        }
        // Read image data
        int length = bi.getWidth() * bi.getHeight();
        int[] imageProperty = new int[length*5];
        int[] pixel;
        int count = 0;
        for (int y = 0; y < bi.getHeight(); y++) {
            for (int x = 0; x < bi.getWidth(); x++) {
                pixel = bi.getRaster().getPixel(x, y, new int[4]);
                imageProperty[count*5  ] = pixel[0];
                imageProperty[count*5+1] = pixel[1];
                imageProperty[count*5+2] = pixel[2];
                imageProperty[count*5+3] = x;
                imageProperty[count*5+4] = y;
                count++;
            }
        }
        //setup
        JCudaDriver.setExceptionsEnabled(true);
        // Create the PTX file
        String ptxFileName;
        try
        {
            ptxFileName = preparePtxFile("KmeansKernel.cu");
        }
        catch (IOException e)
        {
            System.out.println("Warning...");
            System.out.println(e.getMessage());
            System.out.println("Exiting...");
            return;
        }
        cuInit(0);
        CUdevice device = new CUdevice();
        cuDeviceGet(device, 0);
        CUcontext context = new CUcontext();
        cuCtxCreate(context, 0, device);
        CUmodule module = new CUmodule();
        cuModuleLoad(module, ptxFileName);
        CUfunction kmeansFunction = new CUfunction();
        System.out.println("x");
        cuModuleGetFunction(kmeansFunction, module, "add");
        //copy host input to device
        CUdeviceptr imageDevice = new CUdeviceptr();
        cuMemAlloc(imageDevice, imageProperty.length * Sizeof.INT);
        cuMemcpyHtoD(imageDevice, Pointer.to(imageProperty), imageProperty.length * Sizeof.INT);
        int blockSizeX = 256;
        int gridSizeX = (int) Math.ceil((double)(imageProperty.length / 5) / blockSizeX);
        long et = System.currentTimeMillis();
        System.out.println(((double)(et-st)/1000.0) + "s");
        for (int k = startClusters; k <= endClusters; k++) {
            long startTime = System.currentTimeMillis();
            int[] clusters = new int[length];
            int[] c = new int[k*5];
            int h = 0;
            for(int i = 0; i < k; i++) {
                c[i*5] = imageProperty[h*5];
                c[i*5+1] = imageProperty[h*5+1];
                c[i*5+2] = imageProperty[h*5+2];
                c[i*5+3] = imageProperty[h*5+3];
                c[i*5+4] = imageProperty[h*5+4];
                h += length / k;
            }
            double tolerance = 1e-4;
            **//got warning in following line
            CUDA.KmeansKernel(kmeansFunction, imageDevice, imageProperty, clusters, c, k, tolerance, distanceWeight, colorWeight, blockSizeX, gridSizeX);** 
            int[] output = calculateAveragePixels(imageProperty, clusters);
            BufferedImage outputImage = new BufferedImage(bi.getWidth(), bi.getHeight(), BufferedImage.TYPE_INT_RGB);
            for (int i = 0; i < length; i++) {
                int rgb = output[i*5];
                rgb = (rgb * 256) + output[i*5+1];
                rgb = (rgb * 256) + output[i*5+2];
                outputImage.setRGB(i%bi.getWidth(), i/bi.getWidth(), rgb);
            }
            String fileName = (picFiles.getName()) + ".bmp";
            File outputFile = new File("output/" + fileName);
            ImageIO.write(outputImage, "BMP", outputFile);
            long runTime = System.currentTimeMillis() - startTime;
            System.out.println("Completed iteration k=" + k + " in " + ((double)runTime/1000.0) + "s");
        }
        System.out.println("Files saved to " + outputDirectory.getAbsolutePath() + "\");
        cuMemFree(imageDevice);
    } catch (IOException e) {
        e.printStackTrace();
    }
}

方法Kmeans内核：

private static void KmeansKernel(CUfunction kmeansFunction, CUdeviceptr imageDevice, int[] imageProperty, int[] clusters, int[] c,
                                        int k, double tolerance, double distanceWeight, double colorWeight,
                                        int blockSizeX, int gridSizeX) {

    CUdeviceptr clustersDevice = new CUdeviceptr();
    cuMemAlloc(clustersDevice, clusters.length * Sizeof.INT);
    // Alloc device output
    CUdeviceptr centroidPixels = new CUdeviceptr();
    cuMemAlloc(centroidPixels, k * 5 * Sizeof.INT);
    CUdeviceptr errorDevice = new CUdeviceptr();
    cuMemAlloc(errorDevice, Sizeof.DOUBLE * clusters.length);
    int[] c1 = new int[k*5];
    cuMemcpyHtoD(centroidPixels, Pointer.to(c), Sizeof.INT * 5 * k);
    // begin algorithm
    int[] counts = new int[k];
    double old_error, error = Double.MAX_VALUE;
    int l = 0;
    do {
        l++;
        old_error = error;
        error = 0;
        Arrays.fill(counts, 0);
        Arrays.fill(c1, 0);
        cuMemcpyHtoD(centroidPixels, Pointer.to(c), k * 5 * Sizeof.INT);
        Pointer kernelParameters = Pointer.to(
                Pointer.to(new int[] {clusters.length}),
                Pointer.to(new int[] {k}),
                Pointer.to(new double[] {colorWeight}),
                Pointer.to(new double[] {distanceWeight}),
                Pointer.to(errorDevice),
                Pointer.to(imageDevice),
                Pointer.to(centroidPixels),
                Pointer.to(clustersDevice)
        );
        cuLaunchKernel(kmeansFunction,
                gridSizeX, 1, 1,
                blockSizeX, 1, 1,
                0, null,
                kernelParameters, null
        );
        **cuCtxSynchronize(); //got warning here.why?**
        cuMemcpyDtoH(Pointer.to(clusters), clustersDevice, Sizeof.INT*clusters.length);
        for (int i = 0; i < clusters.length; i++) {
            int cluster = clusters[i];
            counts[cluster]++;
            c1[cluster*5] += imageProperty[i*5];
            c1[cluster*5+1] += imageProperty[i*5+1];
            c1[cluster*5+2] += imageProperty[i*5+2];
            c1[cluster*5+3] += imageProperty[i*5+3];
            c1[cluster*5+4] += imageProperty[i*5+4];
        }
        for (int i = 0; i < k; i++) {
            if (counts[i] > 0) {
                c[i*5] = c1[i*5] / counts[i];
                c[i*5+1] = c1[i*5+1] / counts[i];
                c[i*5+2] = c1[i*5+2] / counts[i];
                c[i*5+3] = c1[i*5+3] / counts[i];
                c[i*5+4] = c1[i*5+4] / counts[i];
            } else {
                c[i*5] = c1[i*5];
                c[i*5+1] = c1[i*5+1];
                c[i*5+2] = c1[i*5+2];
                c[i*5+3] = c1[i*5+3];
                c[i*5+4] = c1[i*5+4];
            }
        }

        double[] errors = new double[clusters.length];
        cuMemcpyDtoH(Pointer.to(errors), errorDevice, Sizeof.DOUBLE*clusters.length);
        error = sumArray(errors);
        System.out.println("" + l + " iterations");
    } while (Math.abs(old_error - error) > tolerance);
    cuMemcpyDtoH(Pointer.to(clusters), clustersDevice, clusters.length * Sizeof.INT);
    cuMemFree(errorDevice);
    cuMemFree(centroidPixels);
    cuMemFree(clustersDevice);
}

堆栈跟踪：

Exception in thread "main" jcuda.CudaException: CUDA_ERROR_ILLEGAL_ADDRESS
    at jcuda.driver.JCudaDriver.checkResult(JCudaDriver.java:330)
    at jcuda.driver.JCudaDriver.cuCtxSynchronize(JCudaDriver.java:1938)
    at com.test.CUDA.KmeansKernel(CUDA.java:269)
    at com.test.CUDA.main(CUDA.java:184)

Answer 1

正如@talonmies所提到的，传递给cuLaunchKernel方法kernelParameters与内核函数签名不一致add。

您在cuCtxSynchronize时收到错误，因为 CUDA 执行模型是异步的：cuLaunchKernel立即返回，并且设备上内核的实际执行是异步的。 cuCtxSync文档如下：

请注意，此函数还可能从以前的异步启动返回错误代码。

第二个kernelParameters条目是一个int k，其中add方法的第二个参数是一个pointer to float，因此很可能是非法访问错误。