Peak memory clock and bus width of an OpenMP target device, read from
the vendor runtime.
OpenMP offers no query for either, so this asks the CUDA driver on
NVIDIA hardware and HIP on AMD hardware, the same attributes the CUDA
backend reads. The vendor library is opened at run time with dlopen
instead of being linked: no compiler's offload runtime links it for
us, the backend builds with NVHPC, Flang and Cray alike, and a machine
without the library reports the figures as unavailable rather than
failing to link or start.
The OpenMP device number is passed as the vendor device ordinal. Both
enumerate the devices visible to the process in the same order.
Nodes of different colours represent the following:
Solid arrows point from a submodule to the (sub)module which it is
descended from. Dashed arrows point from a module or program unit to
modules which it uses.
Where possible, edges connecting nodes are
given different colours to make them easier to distinguish in
large graphs.
Nodes of different colours represent the following:
Solid arrows point from a submodule to the (sub)module which it is
descended from. Dashed arrows point from a module or program unit to
modules which it uses.
Where possible, edges connecting nodes are
given different colours to make them easier to distinguish in
large graphs.
Subroutines
public subroutine query_device_memory(device, mem_clock_khz, bus_width_bits, available)
Memory clock (kHz) and bus width (bits) of OpenMP device device.
The vendor this build targets is tried first, then the other one, as
the build only guesses the vendor for some compilers.