OpenCL 的部分封装导致分段错误，包括代码示例

这完全让我感到困惑。两组代码在逻辑上应该是相同的，一组仅在 GPU 上崩溃，而两者都在 CPU 上运行良好。以下是测试代码：

#include <iostream>
#include <CL/cl.hpp>
class Device
{
public:
    cl::Platform platform_;
    cl::Device device_;
    cl::Context context_;
    cl::CommandQueue queue_;
    Device( void ) : platform_()
                   , device_()
                   , context_()
                   , queue_() {}
    Device(int32_t platform, int32_t device) : platform_()
                                             , device_()
                                             , context_()
                                             , queue_()
    {
        std::vector<cl::Platform> platforms;
        cl::Platform::get(&platforms);
        platform_ = platforms[platform];
        std::vector<cl::Device> devices;
        platform_.getDevices(CL_DEVICE_TYPE_GPU, &devices);
        device_ = devices[device];
        cl_context_properties properties[3] = {
            CL_CONTEXT_PLATFORM,
            (cl_context_properties)(platform_)(),
            0
        };
        cl_int clErr = CL_SUCCESS;
        context_ = cl::Context(device_, properties, NULL, NULL, &clErr);
        queue_ = cl::CommandQueue(context_,device_,0,&clErr);
    }
};
int main()
{
    Device device(0,0);
    cl::Program::Sources source;
    std::string src =
    "__kernel void Pointless(uint total, __global uint *data)"
    "{"
    "   uint perStream=total/get_global_size(0);"
    "   __global uint *dest=data+get_global_id(0)*perStream;"
    "   for(uint i=0;i<perStream;i++)"
    "       dest[i] = 1;"
    "}";
    source.push_back({src.c_str(),src.length()});
    cl_int clErr = CL_SUCCESS;
    cl::Program program = cl::Program(device.context_,source,&clErr);
    if (clErr != CL_SUCCESS)
    {
        std::cerr << "Failed to create program: " << clErr << std::endl;
        return 1;
    }
    clErr = program.build({device.device_});
    if(clErr != CL_SUCCESS)
    {
        std::cerr << "Failed to build program: " << clErr << std::endl;
        std::cerr << program.getBuildInfo<CL_PROGRAM_BUILD_LOG>(device.device_) << std::endl;
        return 1;
    }
    uint32_t samples = 16*256;
    cl::make_kernel<cl_uint,cl::Buffer> Pointless(cl::Kernel(program,"Pointless"));
    cl::Buffer device_samples(device.context_,CL_MEM_READ_WRITE,sizeof(cl_uint)*samples);
    Pointless(cl::EnqueueArgs(device.queue_, cl::NDRange(16)), samples, device_samples).wait();
    std::vector<cl_uint> host_samples(samples);
    device.queue_.enqueueReadBuffer(device_samples,CL_TRUE,0,sizeof(cl_uint)*samples,host_samples.data());
    for (auto x: host_samples)
        std::cout << x;
    std::cout << std::endl;
    return 0;
}

以上似乎失败了：我在enqueueReadBuffer上遇到了分段错误。更有趣的是，它仅在 GPU（英特尔 P4000）上失败。CPU（i3 3xxx）运行它没有问题（将CL_DEVICE_TYPE_GPU更改为CL_DEVICE_TYPE_CPU以在CPU上进行测试）。

现在，下面的代码适用于两种设备类型。

#include <iostream>
#include <CL/cl.hpp>
int main()
{
    std::vector<cl::Platform> platforms;
    cl::Platform::get(&platforms);
    cl::Platform platform = platforms[0];
    std::vector<cl::Device> devices;
    platform.getDevices(CL_DEVICE_TYPE_GPU, &devices);
    cl::Device device = devices[0];
    cl_context_properties properties[3] = {
        CL_CONTEXT_PLATFORM,
        (cl_context_properties)(platform)(),
        0
    };
    cl_int clErr = CL_SUCCESS;
    cl::Context context(device, properties, NULL, NULL, &clErr);
    cl::CommandQueue queue(context,device,0,&clErr);
    cl::Program::Sources source;
    std::string src =
    "__kernel void Pointless(uint total, __global uint *data)"
    "{"
    "   uint perStream=total/get_global_size(0);"
    "   __global uint *dest=data+get_global_id(0)*perStream;"
    "   for(uint i=0;i<perStream;i++)"
    "       dest[i] = 1;"
    "}";
    source.push_back({src.c_str(),src.length()});
    cl::Program program = cl::Program(context,source,&clErr);
    clErr = program.build({device});
    if(clErr != CL_SUCCESS)
    {
        std::cerr << program.getBuildInfo<CL_PROGRAM_BUILD_LOG>(device) << std::endl;
    }
    uint32_t samples = 16*256;
    cl::make_kernel<cl_uint,cl::Buffer> Pointless(cl::Kernel(program,"Pointless"));
    cl::Buffer device_samples(context,CL_MEM_READ_WRITE,sizeof(cl_uint)*samples);
    Pointless(cl::EnqueueArgs(queue, cl::NDRange(16)), samples, device_samples).wait();
    std::vector<cl_uint> host_samples(samples);
    queue.enqueueReadBuffer(device_samples,CL_TRUE,0,sizeof(cl_uint)*samples,host_samples.data());
    for (auto x: host_samples)
        std::cout << x;
    std::cout << std::endl;
    return 0;
}

显然，我在这里缺少一些非常基本的东西。 他们都使用英特尔 ICD（我在这个系统上没有 AMD 设备）。