Thread safety of embree device?

[saedrna]

I have a embree device and a scene created and commited in the main thread, and I want to do rtcIntersect using the same device from different threads (using openmp).

Sometimes the program just hang forever or crashes randomly.

I would like to ask that if Embree device is thread safe or not? This will help me to isolate the problem.

Thanks,

Han

[svenwoop]

API calls to the same object are not thread safe in general, thus you cannot have two threads modify some geometry. However, the rtcIntersect/rtcOccluded calls can be invoked from multiple threads, thus what you do is correct.

Where is the crash appearing?

[saedrna]

Thanks, I think I can do some unit test first to check my code

[saedrna]

Here is my code.
On the program start, this is set:

    _MM_SET_FLUSH_ZERO_MODE(_MM_FLUSH_ZERO_ON);
    _MM_SET_DENORMALS_ZERO_MODE(_MM_DENORMALS_ZERO_ON);

The function below is used parallely outside.

std::vector<bool> MeshEmbreeBVH::ray_intersected(const Vector3d &origin, const std::vector<Vector3d> &direction,
                                                 float dnear /*= 0.0f*/, float dfar /*= FLT_MAX*/,
                                                 int mask /*= 0xFFFFFFFF*/) {

    std::vector<bool> hits(direction.size());
    std::vector<RTCRay> rays(direction.size());
    for (int i = 0; i < direction.size(); ++i) {
        RTCRay ray;
        ray.org_x = origin(0);
        ray.org_y = origin(1);
        ray.org_z = origin(2);

        ray.dir_x = direction[i](0);
        ray.dir_y = direction[i](1);
        ray.dir_z = direction[i](2);

        ray.tnear = dnear;
        ray.tfar = dfar;
        ray.mask = -1;
        ray.time = 0.0f;
        ray.flags = 0;
        ray.id = i;

        rays[i] = ray;
    }

    RTCIntersectContext context;
    rtcInitIntersectContext(&context);
    context.flags = RTC_INTERSECT_CONTEXT_FLAG_NONE;
    rtcOccluded1M(scene_, &context, rays.data(), rays.size(), sizeof(RTCRay));
    CHECK_EMBREE_ERROR(device_);

    std::vector<bool> results(rays.size(), false);
    for (int i = 0; i < rays.size(); ++i) {
        if (rays[i].tfar == -std::numeric_limits<float>::infinity()) {
            results[i] = true;
        } else {
            results[i] = false;
        }
    }

    return results;
}

And the macro is

#define CHECK_EMBREE_ERROR(device)                                                                                     \
    {                                                                                                                  \
        auto error = rtcGetDeviceError(device);                                                                        \
        CHECK(error == RTC_ERROR_NONE) << "Error code " << error;                                                      \
    }

So the only thing that I am uncertain is rtcOccluded1M and rtcGetDeviceError, if all of them are thread safe, then probably it's other problems in my code.

[svenwoop]

Does the code work if you do not parallelize this function?

[saedrna]

@svenwoop yes, it works if I comment out openmp

[svenwoop]

Could you please give a stack trace where it crashes exactly?

[saedrna]

I think I have solved the problem. It is not related to embree. Thanks!