OpenCL.clSetKernelArg 返回 -51答案

【问题标题】：OpenCL.clSetKernelArg returns -51OpenCL.clSetKernelArg 返回 -51
【发布时间】：2021-04-03 08:19:30
【问题描述】：

我尝试在 openCL 中创建并行 bfs，但我没有足够的 c++ 经验。所以这可能是内存错误，但我真的不知道如何修复它。我也找不到错误值 -51 的含义。结果，我在下一行得到“我的 project.exe 中 0x00007FFCFB06A549 (amdocl64.dll) 处的未处理异常：0xC0000005：访问冲突读取位置 0xFFFFFFFFFFFFFFFF”。

主要

    Graph G(AdjacencyList, Directed);
    int startVertex;
    vector<int> distance;
    vector<bool> visited;
    distance = vector<int>(G.numVertices);
    visited = vector<bool>(G.numVertices);
    bool done = false;
    const bool true_value = true;

    int level = 0;

    // Allocation on device
    const int size = G.numVertices * sizeof(int);
    const int adjacencySize = G.adjacencyList.size() * sizeof(int);


    //OpenCL
    cl_int status;
    cl_int ret;

    cl_platform_id platform_id;
    clGetPlatformIDs(1, &platform_id, NULL);

    cl_device_id device_id;
    ret = clGetDeviceIDs(platform_id, CL_DEVICE_TYPE_GPU, 1, &device_id, NULL);

    cl_context context = clCreateContext(NULL, 1, &device_id, NULL, NULL, &status);

    cl_command_queue command_queue = clCreateCommandQueueWithProperties(context, device_id, NULL, &status);
    cl_mem d_adjacencyList = clCreateBuffer(context, CL_MEM_READ_WRITE, adjacencySize, NULL, &status);
    cl_mem d_edgesOffset = clCreateBuffer(context, CL_MEM_READ_WRITE, size, NULL, &status);
    cl_mem d_edgesSize = clCreateBuffer(context, CL_MEM_READ_WRITE, size, NULL, &status);
    cl_mem d_distance = clCreateBuffer(context, CL_MEM_READ_WRITE, size, NULL, &status);
    cl_mem d_done = clCreateBuffer(context, CL_MEM_READ_WRITE, sizeof(bool), NULL, &status);
    status = clEnqueueWriteBuffer(command_queue, d_adjacencyList, CL_TRUE, 0, adjacencySize, &G.adjacencyList[0], 0, NULL, NULL);
    status = clEnqueueWriteBuffer(command_queue, d_edgesOffset, CL_TRUE, 0, size, &G.edgesOffset[0], 0, NULL, NULL);
    status = clEnqueueWriteBuffer(command_queue, d_edgesSize, CL_TRUE, 0, size, &G.edgesSize[0], 0, NULL, NULL);
    
    distance = vector<int>(G.numVertices, INT_MAX);
    distance[start] = 0;
    status = clEnqueueWriteBuffer(command_queue, d_distance, CL_TRUE, 0, size, distance.data(), 0, NULL, NULL);
    char* source_str = NULL;
    size_t source_size;
    FILE* fp;


    fp = fopen("bfs.cl", "r");
    if (!fp)
    {
        cout << "Failed to load Kernel\n";
        exit(1);
    }
    source_str = (char*)malloc(MAX_SOURCE_SIZE);
    source_size = fread(source_str, 1, MAX_SOURCE_SIZE, fp);
    cl_program program = clCreateProgramWithSource(context, 1, (const char**)&source_str, (const size_t*)&source_size, &status);
    status = clBuildProgram(program, 1, &device_id, NULL, NULL, NULL);

    cl_kernel kernel = clCreateKernel(program, "bfs", &status);
    
    status = clSetKernelArg(kernel, 0, sizeof(int), (void*)&G.numVertices);
    status = clSetKernelArg(kernel, 1, sizeof(cl_mem), (void*)&d_adjacencyList);
    status = clSetKernelArg(kernel, 2, sizeof(cl_mem), (void*)&d_edgesOffset);
    status = clSetKernelArg(kernel, 3, sizeof(cl_mem), (void*)&d_edgesOffset);
    status = clSetKernelArg(kernel, 4, sizeof(cl_mem), (void*)&d_edgesSize);
    status = clSetKernelArg(kernel, 5, sizeof(cl_mem), (void*)&d_distance); //here retirns -51
    status = clSetKernelArg(kernel, 6, sizeof(cl_mem), (void*)&level); 
    status = clSetKernelArg(kernel, 7, sizeof(cl_mem), (void*)&d_done);

内核

__kernel void bfs(int n, __global int *adjacencyList,__global int *edgesOffset,__global int *edgesSize,__global int *distance, int level,__global bool *done) {
    int tid = get_global_id(0); 
    if (tid < n) {
        if (distance[tid] == level) {
            for (int i = edgesOffset[tid]; i < edgesOffset[tid] + edgesSize[tid]; ++i) {
                int v = adjacencyList[i];
                if (distance[v] == INT_MAX) {
                    *done = false;
                    distance[v] = level + 1;
                }
            }
        }
    }
}

【问题讨论】：

标签： c++ opencl

【解决方案1】：

嗨@Parrison，欢迎来到 StackOverflow！

所有OpenCL 错误代码都在cl.h 中定义。在最新的（第 3 版）cl.h 中，您会发现在第 194 行和第 270 行之间定义的错误代码，您会在第 241 行找到：

#define CL_INVALID_ARG_SIZE -51

因此，OpenCL ICD 认为您为 distance 传递了错误的变量大小。

但是，在此之前我可以看到许多其他错误。例如，您需要根据 OpenCL 变量的大小而不是本机变量的大小来设置 OpenCL 缓冲区的大小，例如：

cl_int 而不是 int
cl_float 而不是 float
尤其是cl_bool 而不是bool。

无法保证 OpenCL cl_int 的大小与主机 int 的大小相同，并且 OpenCL cl_bool 被定义为 unsigned int，它极不可能与 bool 的大小相同！

确保您的 OpenCL 内核的所有参数均已正确定义，并且您正在main 程序中为它们创建正确的buffers 和variables。

【讨论】：