Device.hh File Reference

#include <cstddef>
#include <iosfwd>
#include <map>
#include <memory>
#include <string>
#include "corecel/Assert.hh"
#include "ThreadId.hh"

This graph shows which files directly or indirectly include this file:

Classes
class	celeritas::Device
	Manage attributes of the GPU. More...

Functions
Device const &	celeritas::device ()
	Get the shared default device.
void	celeritas::activate_device (Device &&device)
	Activate the global celeritas device. More...
void	celeritas::activate_device ()
	Initialize the first device if available, when not using MPI.
void	celeritas::activate_device (MpiCommunicator const &comm)
	Initialize device in a round-robin fashion from a communicator.
void	celeritas::activate_device_local ()
	Call cudaSetDevice using the existing device, for thread-local safety. More...
std::ostream &	celeritas::operator<< (std::ostream &os, Device const &d)
	Print device info.
void	celeritas::set_cuda_stack_size (int limit)
	Increase CUDA stack size to enable complex geometries with VecGeom. More...
void	celeritas::set_cuda_heap_size (int limit)
	Increase CUDA heap size to enable complex geometries with VecGeom. More...

Function Documentation

activate_device()

void celeritas::activate_device

(

Device &&

device

)

Activate the global celeritas device.

The given device must be set (true result) unless no device has yet been enabled – this allows make_device to create "null" devices when CUDA is disabled.

This function may be called once only, because the global device propagates into local states (e.g. where memory is allocated) all over Celeritas.

activate_device_local()

void celeritas::activate_device_local

(

)

Call cudaSetDevice using the existing device, for thread-local safety.

See https://developer.nvidia.com/blog/cuda-pro-tip-always-set-current-device-avoid-multithreading-bugs

Precondition

activate_device was called or no device is intended to be used

set_cuda_heap_size()

void celeritas::set_cuda_heap_size

(

int

limit

)

Increase CUDA heap size to enable complex geometries with VecGeom.

For the cms-hllhc.gdml detector geometry, the 8MB default heap size is too small, and a new size as high as 33554432 (=32MB) has run successfully. This should be increased as necessary, but avoid setting it too high.

set_cuda_stack_size()

void celeritas::set_cuda_stack_size

(

int

limit

)

Increase CUDA stack size to enable complex geometries with VecGeom.

For the cms2018.gdml detector geometry, the default stack size is too small, and a limit of 8K is recommended with debugging disabled (and up to 32K if debugging is enabled).