在创建 OpenCL 上下文之前设置线程关联会导致段错误

Question

I'm setting the affinity of the main thread to core 0 but the OpenCl context creation ends with a segfault. (I joined a code snippet to reproduce the problem, I removed vector size check to keep it concise as possible).

If I define more than one core in my cpuset everything works fine.

I have gut feeling about what could be wrong. But I would prefer a concrete explanation.

And then an error could be nicer than a raw segfault.

Open CL info:

• Device Version Intel OpenCL 1.2 (Build 475)
• Driver Version 1.2.0.475

#include <vector>

#include <CL/cl.hpp>


int main(int argc, char** argv)
{
    // Set affinity to core 0
    cpu_set_t cpuset;
    CPU_ZERO(&cpuset);
    CPU_SET(0, &cpuset);
    pthread_setaffinity_np(pthread_self(), sizeof(cpu_set_t), &cpuset);

    // Retrieve first platform
    std::vector<cl::Platform> platforms;
    cl::Platform::get(&platforms);
    cl::Platform platform = platforms[0];

    // Retrieve first CPU device
    std::vector<cl::Device> devices;
    platform.getDevices(CL_DEVICE_TYPE_CPU, &devices);
    cl::Device device = devices[0];

    // Create context
    cl::Context ctx(device);

    return 0;
}

Edit: add OpenCL version

Answer 1

I just tried your example with two Intel OpenCL SDKs, but didn't have your exact version at hand. Both ran without errors, so: cannot reproduce .

This one seems a bit older than yours (by the build number):

CL_DEVICE_VERSION:          OpenCL 1.2 (Build 43)
CL_DRIVER_VERSION:          1.2.0.43
CL_DEVICE_OPENCL_C_VERSION: OpenCL C 1.2 

The second one is, I think, still the most recent Intel CPU driver:

CL_DEVICE_VERSION:          OpenCL 2.1 (Build 0)
CL_DRIVER_VERSION:          18.1.0.0920
CL_DEVICE_OPENCL_C_VERSION: OpenCL C 2.0 

I used GCC 7.4.0 on Ubuntu 18.4.

To my experience it is not a problem to set affinity with the Intel OpenCL Runtime for CPUs .

I used it together with MPI on supercomputers, where you would have multiple, eg 4, MPI processes per compute node and then have the runtime set each process' CPU mask to a disjunct subset of the available cores. The Intel OpenCL Runtime (using Intel Threading Building Blocks under the hood) inside each process would then spawn one thread per assigned core, ie adhere to the host process' CPU mask as expected. Maybe the result is different if you set the first thread's affinity, and also depends on where in the program you set (relative to where the OpenCL runtime spawns it's threads).

As you seem to already have run your code in a debugger and narrowed things down to clcreatecontext . I'd suggest you try using a newer OpenCL driver.