context.hpp File Reference

Contains a proxy class for CUDA execution contexts. More...

#include "current_context.hpp"
#include "versions.hpp"
#include "error.hpp"
#include "constants.hpp"
#include "types.hpp"
#include <string>
#include <utility>

Include dependency graph for context.hpp:

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

Go to the source code of this file.

Classes
struct	cuda::context::stream_priority_range_t
	A range of priorities supported by a CUDA context; ranges from the higher numeric value to the lower. More...
class	cuda::context_t
	Wrapper class for a CUDA context. More...
class	cuda::context_t::global_memory_type
	A class to create a faux member in a context_t, in lieu of an in-class namespace (which C++ does not support); whenever you see a function my_context.memory::foo(), think of it as a my_dev::memory::foo(). More...

Namespaces
	cuda
	Definitions and functionality wrapping CUDA APIs.
	cuda::link
	Definitions related to CUDA linking-processes, captured by the link_t wrapper class.

Typedefs
using	cuda::context::limit_t = CUlimit
	Features of contexts which can be configured individually during a context's lifetime.
using	cuda::context::limit_value_t = size_t
	Type for the actual values for context (see limit_t for the possible kinds of limits whose value can be set)
using	cuda::context::shared_memory_bank_size_t = CUsharedconfig
	Choice of the number of bytes in each bank of the shared memory.

Functions
context_t	cuda::context::wrap (device::id_t device_id, context::handle_t context_id, bool take_ownership=false) noexcept
	Obtain a wrapper for an already-existing CUDA context. More...
void	cuda::synchronize (const context_t &context)
	Waits for all previously-scheduled tasks on all streams (= queues) in a CUDA context to conclude, before returning. More...
context_t	cuda::context::create (const device_t &device, host_thread_sync_scheduling_policy_t sync_scheduling_policy=heuristic, bool keep_larger_local_mem_after_resize=false)
	creates a new context on a given device More...
context_t	cuda::context::create_and_push (const device_t &device, host_thread_sync_scheduling_policy_t sync_scheduling_policy=heuristic, bool keep_larger_local_mem_after_resize=false)
	Creates a new CUDA context on a given device, as would create() - and pushes it onto the top of the context stack. More...
context_t	cuda::context::current::get ()
	Obtain the current CUDA context, if one exists. More...
void	cuda::context::current::set (const context_t &context)
	Set the context at the top of the stack to a specified context. More...
bool	cuda::context::current::push_if_not_on_top (const context_t &context)
	Push a (reference to a) context onto the top of the context stack - unless that context is already at the top of the stack, in which case do nothing.
void	cuda::context::current::push (const context_t &context)
	Push a (reference to a) context onto the top of the context stack. More...
context_t	cuda::context::current::pop ()
	Pop the top off of the context stack. More...
bool	cuda::context::is_primary (const context_t &context)
bool	cuda::operator== (const context_t &lhs, const context_t &rhs) noexcept
bool	cuda::operator!= (const context_t &lhs, const context_t &rhs) noexcept

Detailed Description

Contains a proxy class for CUDA execution contexts.

Function Documentation

create()

context_t cuda::context::create ( const device_t &  device, host_thread_sync_scheduling_policy_t  sync_scheduling_policy = heuristic, bool  keep_larger_local_mem_after_resize = false  )

inline

creates a new context on a given device

Parameters

device	The device which the new context will regard
sync_scheduling_policy	Choice of how host threads are to perform synchronization with pending actions in streams within this context. See host_thread_sync_scheduling_policy_t for a description of these choices.
keep_larger_local_mem_after_resize	If true, larger allocations of global device memory, used by kernels requiring a larger amount of local memory, will be kept (so that future kernels with such requirements will not trigger a re-allocation).

Note

Until CUDA 11, there used to also be a flag for enabling/disabling the ability of mapping pinned host memory to device addresses. However, it was being ignored since CUDA 3.2 already, with the minimum CUDA version supported by these wrappers being later than that, so - no sense in keeping it.

create_and_push()

context_t cuda::context::create_and_push ( const device_t &  device, host_thread_sync_scheduling_policy_t  sync_scheduling_policy = heuristic, bool  keep_larger_local_mem_after_resize = false  )

inline

Creates a new CUDA context on a given device, as would create() - and pushes it onto the top of the context stack.

Note

The effect of this function could, of course, be achieved by creating a context without a push, then pushing it onto the top of the stack - both functions offered in this same namespace. However, the CUDA Driver API actually offers this functionality directly, i.e. just creating a context means creating-and-pushing, then popping, which is a waste.

get()

context_t cuda::context::current::get ( )

inline

Obtain the current CUDA context, if one exists.

Exceptions

::std::runtime_error

in case there is no current context

is_primary()

bool cuda::context::is_primary ( const context_t &  context )

inline

Returns

true if the context is the primary context of its device

pop()

context_t cuda::context::current::pop ( )

inline

Pop the top off of the context stack.

Returns

a non-owning context wrapper for the context (previously) at the top of the stack

push()

void cuda::context::current::push ( const context_t &  context )

inline

Push a (reference to a) context onto the top of the context stack.

Note

if the specified context is already at the top of the stack - it is still pushed, i.e. there will now be one additional reference to the context on the stack.

set()

void cuda::context::current::set ( const context_t &  context )

inline

Set the context at the top of the stack to a specified context.

Note

If the context stack is empty, the context (reference) is pushed as the single element; otherwise, the top item on the stack is replaced with the specified context, with the number of elements on the stack not changing.

wrap()

context_t cuda::context::wrap ( device::id_t  device_id, context::handle_t  context_id, bool  take_ownership = false  )

inlinenoexcept

Obtain a wrapper for an already-existing CUDA context.

Wrap the fundamental information regarding a CUDA context into an instance of the context_t wrapper class.

Note

This is a named constructor idiom instead of direct access to the ctor of the same signature, to emphasize what this construction means - a new context is not created.

Parameters

device_id	Device with which the context is associated
context_id	id of the context to wrap with a proxy
take_ownership	when true, the wrapper will have the CUDA driver destroy the context when the wrapper itself destruct; otherwise, it is assumed that the context is "owned" elsewhere in the code, and that location or entity is responsible for destroying it when relevant (possibly after this wrapper ceases to exist)

Returns

a context wrapper associated with the specified context