cudaMallocManged buffer offsets supported?

In OptiX 7.0.0 SDK:
is it generally supported to use buffer offsets returned by cudaMallocManged (instead of cudaMalloc)
for all buffers (especially on inputs of optixAccelBuild and similar) ?

For what benefit?

That is not supported on all system configurations and, as far as I know, accessing the device memory from the host without migration is only supported under Linux, and that would be the main reason to use managed memory.
On my Windows 10 dual Quadro RTX6000 SLI system the cudaGetDeviceProperties() field directManagedMemAccessFromHost returns false.

1 Like