amrex::SArena Class Reference

A STREAM-ordered memory arena. More...

#include <AMReX_SArena.H>

Inheritance diagram for amrex::SArena:

Public Member Functions
void *	alloc (std::size_t sz) final
	Allocates dynamic memory from the arena of size sz. Returns a pointer to this memory. The memory can be used immediately by both the CPU and GPU with any stream. This function may synchronize all GPU streams to free up memory.
void	free (void *pt) final
	Deletes the arena pointed to by pt. After this function is called, the memory can still be used by already launched GPU kernels on the currently active stream. This function may synchronize the currently active GPU stream to free up memory.
bool	isDeviceAccessible () const final
bool	isHostAccessible () const final
bool	isManaged () const final
	Check whether it is managed GPU memory.
bool	isDevice () const final
	Check whether it is non-managed GPU device memory.
bool	isPinned () const final
	Check whether it is pinned host memory.
bool	isStreamOrderedArena () const final
	Is this a GPU stream ordered memory allocator?
Public Member Functions inherited from amrex::Arena
virtual	~Arena ()=default
	Arena () noexcept=default
	Arena (const Arena &rhs)=delete
	Arena (Arena &&rhs)=delete
Arena &	operator= (const Arena &rhs)=delete
Arena &	operator= (Arena &&rhs)=delete
virtual std::pair< void *, std::size_t >	alloc_in_place (void *, std::size_t, std::size_t szmax)
	Allocate memory in-place if possible.
virtual void *	shrink_in_place (void *, std::size_t sz)
	Shrink allocation size in-place.
virtual std::size_t	freeUnused ()
	Free unused memory back to the system. Return value is the amount memory freed.
virtual bool	hasFreeDeviceMemory (std::size_t sz)
	Does the device have enough free memory for allocating this much memory? For CPU builds, this always return true. This is not a const function because it may attempt to release memory back to the system.
void	registerForProfiling (const std::string &memory_name)
	Add this Arena to the list of Arenas that are profiled by TinyProfiler.
void	deregisterFromProfiling ()
	Remove this Arena from the list of Arenas that are profiled by TinyProfiler. This is equivalent to destructing and re-constructing the Arena.
const ArenaInfo &	arenaInfo () const
	Return the ArenaInfo object for querying.
virtual void	ResetMaxUsageCounter ()

Additional Inherited Members
Static Public Member Functions inherited from amrex::Arena
static std::size_t	align (std::size_t sz)
	Given a minimum required arena size of sz bytes, this returns the next largest arena size that will align to align_size bytes.
static void	Initialize (bool minimal)
	Used internally by amrex.
static void	PrintUsage (bool print_max_usage=false)
	Print memory usage information of all arenas.
static void	PrintUsageToStream (std::ostream &os, std::string const &space)
	Print memory usage information of all arenas to a given output stream.
static void	PrintUsageToFiles (std::string const &filename, std::string const &message)
	Print memory usage information of all arenas to given file.
static void	Finalize ()
	Used internally by amrex.
Static Public Attributes inherited from amrex::Arena
static const std::size_t	align_size = 16
	The alignment of allocated memory.
Protected Member Functions inherited from amrex::Arena
virtual std::size_t	freeUnused_protected ()
void *	allocate_system (std::size_t nbytes)
void	deallocate_system (void *p, std::size_t nbytes)
Static Protected Member Functions inherited from amrex::Arena
static void	out_of_memory_abort (std::string const &memory_type, std::size_t nbytes, std::string const &error_msg)
Protected Attributes inherited from amrex::Arena
ArenaInfo	arena_info
struct amrex::Arena::ArenaProfiler	m_profiler

Detailed Description

A STREAM-ordered memory arena.

This Arena is implemented as a wrapper around amrex::The_Arena() where the free() function calls amrex::Gpu::freeAsync(). This allows memory to be used by GPU kernels that were launched before free() was called but may execute afterward, without calling amrex::Gpu::streamSynchronize(). These kernels need to use the same GPU stream that is active when free() is called, otherwise explicit synchronization is needed.

This is achieved by holding freed memory in a temporary buffer until the next time the GPU stream is synchronized. In case too much memory is held up in this buffer or if The_Arena needs more memory, the stream is automatically synchronized to clear the buffer.

Member Function Documentation

alloc()

void * amrex::SArena::alloc ( std::size_t  sz )

finalvirtual

Allocates dynamic memory from the arena of size sz. Returns a pointer to this memory. The memory can be used immediately by both the CPU and GPU with any stream. This function may synchronize all GPU streams to free up memory.

Implements amrex::Arena.

free()

void amrex::SArena::free ( void *  pt )

finalvirtual

Deletes the arena pointed to by pt. After this function is called, the memory can still be used by already launched GPU kernels on the currently active stream. This function may synchronize the currently active GPU stream to free up memory.

Implements amrex::Arena.

isDevice()

bool amrex::SArena::isDevice ( ) const

finalvirtual

Check whether it is non-managed GPU device memory.

Note that isManaged, isDevice and isPinned are mutually exclusive. For memory allocated by cudaMalloc* etc. in GPU builds, one of them returns true. For CPU builds, they are always false.

Reimplemented from amrex::Arena.

isDeviceAccessible()

bool amrex::SArena::isDeviceAccessible ( ) const

finalvirtual

Check whether the memory is accessible on the device. Note that isDeviceAccessible and isHostAccessible can both be true.

Reimplemented from amrex::Arena.

isHostAccessible()

bool amrex::SArena::isHostAccessible ( ) const

finalvirtual

Check whether the memory is accessible on the host. Note that isDeviceAccessible and isHostAccessible can both be true.

Reimplemented from amrex::Arena.

isManaged()

bool amrex::SArena::isManaged ( ) const

finalvirtual

Check whether it is managed GPU memory.

Note that isManaged, isDevice and isPinned are mutually exclusive. For memory allocated by cudaMalloc* etc. in GPU builds, one of them returns true. For CPU builds, they are always false.

Reimplemented from amrex::Arena.

isPinned()

bool amrex::SArena::isPinned ( ) const

finalvirtual

Check whether it is pinned host memory.

Note that isManaged, isDevice and isPinned are mutually exclusive. For memory allocated by cudaMalloc* etc. in GPU builds, one of them returns true. For CPU builds, they are always false.

Reimplemented from amrex::Arena.

isStreamOrderedArena()

bool amrex::SArena::isStreamOrderedArena ( ) const

inlinefinalvirtual

Is this a GPU stream ordered memory allocator?

Reimplemented from amrex::Arena.

---

The documentation for this class was generated from the following files:

• /home/runner/work/amrex/amrex/Src/Base/AMReX_SArena.H
• /home/runner/work/amrex/amrex/Src/Base/AMReX_SArena.cpp