rmm Namespace Reference

The top-level namespace containing all RMM functionality. More...

Namespaces
namespace	mr
	Memory Resource classes and adaptors.

Classes
class	bad_alloc
	Exception thrown when an RMM allocation fails. More...
struct	cuda_device_id
	Strong type for a CUDA device identifier. More...
struct	cuda_error
	Exception thrown when a CUDA error is encountered. More...
struct	cuda_set_device_raii
	RAII class that sets the current CUDA device to the specified device on construction and restores the previous device on destruction. More...
class	cuda_stream
	Owning wrapper for a CUDA stream. More...
class	cuda_stream_pool
	A pool of CUDA streams. More...
class	device_buffer
	RAII construct for device memory allocation. More...
class	device_scalar
	Container for a single object of type T in device memory. More...
class	device_uvector
	An uninitialized vector of elements in device memory. More...
class	exec_policy
	Helper class usable as a Thrust CUDA execution policy that uses RMM for temporary memory allocation on the specified stream. More...
class	exec_policy_nosync
	Helper class usable as a Thrust CUDA execution policy that uses RMM for temporary memory allocation on the specified stream and which allows the Thrust backend to skip stream synchronizations that are not required for correctness. More...
class	invalid_argument
	Exception thrown when an argument to a function is invalid. More...
struct	logic_error
	Exception thrown when logical precondition is violated. More...
class	out_of_memory
	Exception thrown when RMM runs out of memory. More...
class	out_of_range
	Exception thrown when attempting to access outside of a defined range. More...

Typedefs
template<typename T>
using	device_vector = thrust::device_vector<T, rmm::mr::thrust_allocator<T>>
	Alias for a thrust::device_vector that uses RMM for memory allocation.
using	thrust_exec_policy_t
	Synchronous execution policy for allocations using Thrust.
using	thrust_exec_policy_nosync_t
	Asynchronous execution policy for allocations using Thrust.
using	device_resource_ref = cuda::mr::synchronous_resource_ref<cuda::mr::device_accessible>
	Alias for a cuda::mr::synchronous_resource_ref with the property cuda::mr::device_accessible.
using	device_async_resource_ref = cuda::mr::resource_ref<cuda::mr::device_accessible>
	Alias for a cuda::mr::resource_ref with the property cuda::mr::device_accessible.
using	host_resource_ref = cuda::mr::synchronous_resource_ref<cuda::mr::host_accessible>
	Alias for a cuda::mr::synchronous_resource_ref with the property cuda::mr::host_accessible.
using	host_async_resource_ref = cuda::mr::resource_ref<cuda::mr::host_accessible>
	Alias for a cuda::mr::resource_ref with the property cuda::mr::host_accessible.
using	host_device_resource_ref
	Alias for a cuda::mr::synchronous_resource_ref with the properties cuda::mr::host_accessible and cuda::mr::device_accessible.
using	host_device_async_resource_ref
	Alias for a cuda::mr::resource_ref with the properties cuda::mr::host_accessible and cuda::mr::device_accessible.

Functions
bool	is_pow2 (std::size_t value) noexcept
	Returns whether or not value is a power of 2.
bool	is_supported_alignment (std::size_t alignment) noexcept
	Returns whether or not alignment is a valid memory alignment.
bool	is_supported_base_resource_alignment (std::size_t alignment) noexcept
	Returns whether or not alignment is supported by base CUDA memory resources.
std::size_t	align_up (std::size_t value, std::size_t alignment) noexcept
	Align up to nearest multiple of specified power of 2.
std::size_t	align_down (std::size_t value, std::size_t alignment) noexcept
	Align down to the nearest multiple of specified power of 2.
bool	is_aligned (std::size_t value, std::size_t alignment) noexcept
	Checks whether a value is aligned to a multiple of a specified power of 2.
bool	is_pointer_aligned (void *ptr, std::size_t alignment=rmm::CUDA_ALLOCATION_ALIGNMENT) noexcept
	Checks whether the provided pointer is aligned to a specified alignment.
cuda_device_id	get_current_cuda_device ()
	Returns a cuda_device_id for the current device.
int	get_num_cuda_devices ()
	Returns the number of CUDA devices in the system.
std::pair< std::size_t, std::size_t >	available_device_memory ()
	Returns the available and total device memory in bytes for the current device.
std::size_t	percent_of_free_device_memory (int percent)
	Returns the approximate specified percent of available device memory on the current CUDA device, aligned (down) to the nearest CUDA allocation size.
rapids_logger::sink_ptr	default_sink ()
	Returns the default sink for the global logger.
std::string	default_pattern ()
	Returns the default log pattern for the global logger.
rapids_logger::logger &	default_logger ()
	Get the default logger.
void	prefetch (void const *ptr, std::size_t size, rmm::cuda_device_id device, cuda::stream_ref stream)
	Prefetch memory to the specified device on the specified stream.
template<typename T>
void	prefetch (cuda::std::span< T const > data, rmm::cuda_device_id device, cuda::stream_ref stream)
	Prefetch a span of memory to the specified device on the specified stream.
bool	process_is_exiting () noexcept
	Returns true if the process has entered exit() / atexit handler execution.
template<class Resource>
device_async_resource_ref	to_device_async_resource_ref_checked (Resource *res)
	Convert pointer to memory resource into device_async_resource_ref, checking for nullptr.

Variables
static constexpr std::size_t	CUDA_ALLOCATION_ALIGNMENT {256}
	Default alignment used for CUDA memory allocation.

Detailed Description

The top-level namespace containing all RMM functionality.

The rmm namespace provides a comprehensive set of memory management utilities for CUDA applications, including memory resources, CUDA stream management, device-side data containers, and memory allocation utilities.

Function Documentation

default_logger()

rapids_logger::logger & rmm::default_logger

(

)

Get the default logger.

Returns

logger& The default logger

default_pattern()

std::string rmm::default_pattern

(

)

Returns the default log pattern for the global logger.

Returns

std::string The default log pattern.

default_sink()

rapids_logger::sink_ptr rmm::default_sink

(

)

Returns the default sink for the global logger.

If the environment variable RMM_DEBUG_LOG_FILE is defined, the default sink is a sink to that file. Otherwise, the default is to dump to stderr.

Returns

sink_ptr The sink to use