cuda::experimental::stf::exec_place

Defined in include/cuda/experimental/__stf/places/places.cuh

class exec_place

Indicates where a computation takes place (CPU, dev0, dev1, …)

Currently data and computation are together (devid == int(data_place)).

Subclassed by cuda::experimental::stf::exec_place_cuda_stream, cuda::experimental::stf::exec_place_grid, cuda::experimental::stf::exec_place_host

Public Functions

exec_place() = default

inline exec_place(const data_place &affine)

inline bool operator==(const exec_place &rhs) const

inline bool operator!=(const exec_place &rhs) const

inline iterator begin()

inline iterator end()

inline ::std::string to_string() const

Returns a string representation of the execution place object.

Returns: std::string

inline const data_place &affine_data_place() const: Returns the data_place naturally associated with this execution place.

inline void set_affine_data_place(data_place place)

inline stream_pool &get_stream_pool(async_resources_handle &async_resources, bool for_computation) const

inline decorated_stream getStream(async_resources_handle &async_resources, bool for_computation) const

inline const ::std::shared_ptr<impl> &get_impl() const

inline exec_place activate(backend_ctx_untyped &state) const

Set computation to run on this place.

Returns: exec_place The previous execution place. See deactivate below.

inline void deactivate(backend_ctx_untyped &state, const exec_place &p) const

Undoes the effect of activate.

Call with the previous exec_place object retured by activate.

Warning

Undefined behavior if you don’t pass the result of activate.

inline bool is_device() const

inline bool is_grid() const

inline size_t size() const

inline exec_place_grid as_grid() const

inline size_t grid_dim(int axid_is) const

inline dim4 grid_dims() const

template<typename Fun> inline auto operator->*(Fun &&fun) const

Execute lambda on this place.

This method accepts a functor, saves the current CUDA device, changes it to the current execution place, invokes the lambda, and finally sets the current device back to the previous one. The last step is taken even if the lambda throws an exception.

Template Parameters: Fun – A callable entity type
Parameters: fun – Input functor that will be forwarded and executed
Returns: auto the result of the executed functor.

inline exec_place(::std::shared_ptr<impl> pimpl)

Public Static Functions

static inline exec_place device(int devid)

static inline exec_place_cuda_stream cuda_stream(cudaStream_t stream)

static inline exec_place_cuda_stream cuda_stream(const decorated_stream &dstream)

static inline exec_place current_device()

Returns the currently active device.

Returns: exec_place

static inline exec_place_grid all_devices()

static inline exec_place_grid n_devices(size_t n, dim4 dims)

static inline exec_place_grid n_devices(size_t n)

static inline exec_place_grid repeat(const exec_place &e, size_t cnt)

Public Static Attributes

static const exec_place_host host = {}

static const exec_place device_auto = {data_place::device_auto}

class impl

Subclassed by cuda::experimental::stf::exec_place_cuda_stream::impl, cuda::experimental::stf::exec_place_grid::impl, cuda::experimental::stf::exec_place_host::impl

Public Functions

impl() = default

impl(const impl&) = delete

impl &operator=(const impl&) = delete

virtual ~impl() = default

inline explicit impl(data_place place)

inline virtual exec_place activate(backend_ctx_untyped&) const

inline virtual void deactivate(backend_ctx_untyped&, const exec_place &prev) const

inline virtual const data_place &affine_data_place() const

inline ::std::string to_string() const

inline virtual bool is_device() const

inline virtual bool is_grid() const

inline virtual size_t size() const

inline virtual void set_affine_data_place(data_place place)

inline virtual bool operator==(const impl &rhs) const

inline virtual stream_pool &get_stream_pool(async_resources_handle &async_resources, bool for_computation) const

inline decorated_stream getStream(async_resources_handle &async_resources, bool for_computation) const

Protected Functions

inline explicit impl(int devid)

Protected Attributes

data_place affine = data_place::invalid 

Friends

friend class exec_place

class iterator

an iterator class which goes over all subplaces in an exec place.

This is a trivial singleton unless we have a grid of places.

Public Functions

inline iterator(::std::shared_ptr<impl> impl, size_t index)

inline exec_place operator*(): Implementation deferred because we need the definition of exec_place_grid.

inline iterator &operator++()

inline bool operator==(const iterator &other) const

inline bool operator!=(const iterator &other) const