How to bind buffer to global memory

Gaming and Visualization Technologies Visualization OptiX
1

The following works well.

void* position_buffer_data;
    ...
    RT_CHECK_ERROR2( rtBufferMap( input_energy_buffer_obj, &position_buffer_data));
    for(int i=0;i<width;i++)
       ((float*)position_buffer_data)[i]= 1.6*(i+1);
    RT_CHECK_ERROR2( rtBufferUnmap( input_energy_buffer_obj));

But I want to bind the buffer to a global memory, then I can avoid the cost of cudaMempyDeviceToHost and then bind a host pointer to the buffer as above during iterative kernel launches.

I tried

void* position_buffer_data;
    ...
    RT_CHECK_ERROR2( rtBufferMap( input_energy_buffer_obj, &position_buffer_data));
    (float*)position_buffer_data = testm; 
    /*testm is pointer to a global memory that have been cudaMalloc. */
    RT_CHECK_ERROR2( rtBufferUnmap( input_energy_buffer_obj));

but got error “error: lvalue required as left operand of assignment”

Any ideas?
Thanks!

2

Well, optix programming guide say buffer is used to communicate between hot and device. Then can I use a argument in the ray_generation program?

3

Yes, you can do this. Look in the OptiX Programming Guide for Chapter 7, Interoperability with CUDA. Search for the functions rtBufferGetDevicePtr and rtBufferCreateForCUDA.

The collision sample uses CUDA interop.

4

What is “DevicePtr”? Isn’t it pointer to data on device?

But I have error for

RT_CHECK_ERROR2( rtBufferSetDevicePointer( input_test_buffer_obj, 0, testm));
//testm is pointer to a float array allocated by cudaMalloc()

invalid conversion from ‘float*’ to ‘CUdeviceptr {aka long long unsigned int}’ [-fpermissive]
error: initializing argument 3 of ‘RTresult rtBufferSetDevicePointer(RTbuffer, unsigned int, CUdeviceptr)’ [-fpermissive]

Edit: just found CUdeviceptr is used for CUDA driver API. How can I use it with runtime API as I usually do with CUDA?