JZ Li

play_w_madness
Gas OpenCL DOP nodes have "Flush Attributes" toggle - "After writing to attributes, the new values are left on the GPU until another solver requests the geometry attributes. This lets the attributes stay there and provides the most efficiency. Turning on flush attributes forces them to be copied back from the GPU into geometry memory explicitly. This should not be required."
It's off by default, meaning the data isn't flushed into the host memory until necessary.

As for your other question, HDK provides CE_Grid struct ( https://www.sidefx.com/docs/hdk/class_c_e___grid.html [www.sidefx.com] ) which stores cl::Buffer, that you can possibly use for interop with CUDA.
Basically what you need is CE_... headers from the HDK - CE_Context, CE_VDBGrid etc.

Additionally, I recommend taking a look at
https://www.sidefx.com/docs/hdk/_g_a_s___open_c_l_8h_source.html [www.sidefx.com]
https://www.sidefx.com/docs/hdk/_s_i_m___object_8h_source.html [www.sidefx.com]

void GAS_OpenCL_Lissajous::accessOpenCLBuffer(const GU_Detail* gdp, float time) {

    CE_Context *ceContext = CE_Context::getContext();

    cl::Context clContext = ceContext->getCLContext();
    cl::CommandQueue clQueue = ceContext->getQueue();

    cl::Buffer gpuBuffer = ...; // how to get my GPU buffer??

    thrust::device_vector<Particle> deviceBuffer(500000);  // 500,000 particles
    Particle* tempHostBuffer = new Particle[500000];
    ceContext->readBuffer(gpuBuffer, sizeof(Particle) * 500000, tempHostBuffer);

    ...
    ...
}

animatrix_
Hi,

You can use it inside Houdini using C++ Wrangle:

This guy integrated the Eigen library also:
https://github.com/lecopivo/cpp-wrangle [github.com]

So you can try that version to access the Eigen library directly. This is the closest you can get albeit not inside VEX or OpenCL.