使用驱动程序 API 创建纹理对象时 JCuda 访问冲突

Question

我有一个 JCuda 项目在尝试使用驱动程序 API 创建纹理对象时遇到访问冲突。 Java HotSpot 声称错误来自 nvcuda.dll。

创建纹理的底层 CUarray 似乎已正确填充；将其内容复制回主机端浮点数组会生成一个与初始主机端数据相同的数组。这意味着错误本身必须是纹理声明中的内容，对吧？

使用 cuda-memcheck 运行代码没有发现任何错误。

这是遇到错误的代码：

import jcuda.Pointer;
import jcuda.Sizeof;
import jcuda.driver.*;

public class Main {

    public static void main(String[] args) {
        init();

        float[] hostArray = new float[]{0, 1, 2, 3, 4, 5, 6, 7};
        int[] dims = new int[]{2,2,2};

        CUdeviceptr deviceArray = new CUdeviceptr();
        JCudaDriver.cuMemAlloc(deviceArray, hostArray.length * Sizeof.FLOAT);
        JCudaDriver.cuMemcpyHtoD(deviceArray, Pointer.to(hostArray), hostArray.length * Sizeof.FLOAT);

        // initialize the opaque array object to represent the texture's data
        CUarray cuArray = makeCudaArray(dims);

        // populate the opaque array object
        copyDataIntoCudaArray(deviceArray, cuArray, dims);

        JCudaDriver.cuMemFree(deviceArray);

        // create the various descriptors
        CUDA_RESOURCE_DESC resourceDescriptor = makeResourceDescriptor(cuArray);
        CUDA_TEXTURE_DESC textureDescriptor = makeTextureDescriptor();
        CUDA_RESOURCE_VIEW_DESC resourceViewDescriptor = makeResourceViewDescriptor(dims);

        CUtexObject texture = new CUtexObject();

        System.out.println("About to hit an access violation:");
        JCudaDriver.cuTexObjectCreate(texture, resourceDescriptor, textureDescriptor, resourceViewDescriptor);
    }

    static void init() {
        JCudaDriver.setExceptionsEnabled(true);
        JCudaDriver.cuInit(0);

        int[] deviceCount = new int[1];
        JCudaDriver.cuDeviceGetCount(deviceCount);

        CUdevice currentDevice = new CUdevice();
        JCudaDriver.cuDeviceGet(currentDevice, 0);

        CUcontext currentContext = new CUcontext();
        JCudaDriver.cuCtxCreate(currentContext, 0, currentDevice);
    }

    static CUarray makeCudaArray(int[] dims) {
        CUarray array = new CUarray();
        CUDA_ARRAY3D_DESCRIPTOR arrayDescriptor = new CUDA_ARRAY3D_DESCRIPTOR();

        arrayDescriptor.Width = dims[0];
        arrayDescriptor.Height = dims[1];
        arrayDescriptor.Depth = dims[2];
        arrayDescriptor.Format = CUarray_format.CU_AD_FORMAT_FLOAT;
        arrayDescriptor.NumChannels = 1;
        arrayDescriptor.Flags = 0;

        JCudaDriver.cuArray3DCreate(array, arrayDescriptor);
        return array;
    }

    static void copyDataIntoCudaArray(CUdeviceptr deviceArray, CUarray array, int[] dims) {
        CUDA_MEMCPY3D copyParams = new CUDA_MEMCPY3D();
        copyParams.srcMemoryType = CUmemorytype.CU_MEMORYTYPE_DEVICE;
        copyParams.srcDevice = deviceArray;
        copyParams.srcXInBytes = 0;
        copyParams.srcY = 0;
        copyParams.srcZ = 0;
        copyParams.srcPitch = (long) dims[0] * Sizeof.FLOAT;
        copyParams.srcHeight = dims[1];
        copyParams.srcLOD = 0;

        copyParams.dstMemoryType = CUmemorytype.CU_MEMORYTYPE_ARRAY;
        copyParams.dstArray = array;
        copyParams.dstXInBytes = 0;
        copyParams.dstY = 0;
        copyParams.dstZ = 0;
        copyParams.dstLOD = 0;

        copyParams.WidthInBytes = (long) dims[0] * Sizeof.FLOAT;
        copyParams.Height = dims[1];
        copyParams.Depth = dims[2];

        JCudaDriver.cuMemcpy3D(copyParams);
    }

    static CUDA_RESOURCE_DESC makeResourceDescriptor(CUarray cuArray) {
        CUDA_RESOURCE_DESC resourceDescriptor = new CUDA_RESOURCE_DESC();
        resourceDescriptor.resType = CUresourcetype.CU_RESOURCE_TYPE_ARRAY;
        resourceDescriptor.array_hArray = cuArray;
        resourceDescriptor.flags = 0;
        return resourceDescriptor;
    }

    static CUDA_TEXTURE_DESC makeTextureDescriptor() {
        CUDA_TEXTURE_DESC textureDescriptor = new CUDA_TEXTURE_DESC();
        textureDescriptor.addressMode = new int[]{
                CUaddress_mode.CU_TR_ADDRESS_MODE_CLAMP,
                CUaddress_mode.CU_TR_ADDRESS_MODE_CLAMP,
                CUaddress_mode.CU_TR_ADDRESS_MODE_CLAMP };
        textureDescriptor.filterMode = CUfilter_mode.CU_TR_FILTER_MODE_LINEAR;
        textureDescriptor.flags = 0;
        textureDescriptor.maxAnisotropy = 1;
        textureDescriptor.mipmapFilterMode = CUfilter_mode.CU_TR_FILTER_MODE_POINT;
        textureDescriptor.mipmapLevelBias = 0;
        textureDescriptor.minMipmapLevelClamp = 0;
        textureDescriptor.maxMipmapLevelClamp = 0;
        return textureDescriptor;
    }

    static CUDA_RESOURCE_VIEW_DESC makeResourceViewDescriptor(int[] dims) {
        CUDA_RESOURCE_VIEW_DESC resourceViewDescriptor = new CUDA_RESOURCE_VIEW_DESC();
        resourceViewDescriptor.format = CUresourceViewFormat.CU_RES_VIEW_FORMAT_FLOAT_1X32;
        resourceViewDescriptor.width = dims[0];
        resourceViewDescriptor.height = dims[1];
        resourceViewDescriptor.depth = dims[2];
        resourceViewDescriptor.firstMipmapLevel = 0;
        resourceViewDescriptor.lastMipmapLevel = 0;
        resourceViewDescriptor.firstLayer = 0;
        resourceViewDescriptor.lastLayer = 0;
        return resourceViewDescriptor;
    }
}

我在这里做错了什么？

Answer 1

这种访问冲突的原因是 JCuda 0.9.0 中的一个错误。

纹理句柄被错误地作为NULL 指针传递给本机函数。这已在 this commit 中修复，该修复将包含在下一个版本中。

已添加test case based on the code in the question。

更新：此问题已在 JCuda 0.9.0d 中修复。